The AI for E-Commerce Newsletter

Make Every Mistake a Rule

Ritu Java — Tue, 04 Aug 2026 15:07:00 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Make Every Mistake a Rule💥
Marie Kondo Your Claude Chats 🧹
Voice Finally Grew Up 🎙️
4 AI Hacks From the Seven-Figure Seller Summit 🎤
News and Events Worth Following

Make Every Mistake a Rule

Ever corrected your AI, then watched it make the exact same mistake a week later?

Then you already know the problem. Today I am going to show you how to make a fix permanent, so you never have to give the same correction twice.

This one changed how I work. It is not a tool. It is a habit, plus two upgrades that make the habit stick even when you switch between AI tools.

The Leak Nobody Plugs

Here is the part most people miss.

The expensive thing is not the mistake. It is the re-explaining. You tell your AI how you like your subject lines written. You fix the tone. You correct a formatting choice. It gets it right. Then the session ends, and all of that context evaporates.

Next session, you start from zero. Same corrections, same tuition, paid again.

Think about that for a second. Every correction you give is a small piece of training. And most of us throw it in the trash the moment we close the window.

The fix is to treat every correction as an asset you keep, not a note you lose. There are three ways to do it, from the ten second version to the one that carries across every tool you use.

Level One: Write the Rule

This is the one you can do today, and it costs nothing.

Most AI tools read a plain instructions file when they start up. Claude Code reads a file called CLAUDE.md. Many other tools, including Codex, Cursor, and Windsurf, read one called AGENTS.md. Whatever the name, it is a file that loads before your AI does any work, and it treats what is inside as standing instructions.

So when your AI gets something wrong, do not just fix the output. Add one line to that file.

🔶 It keeps calling your brand "PPCNinja" instead of "PPC Ninja"? Add the rule.
🔶 It writes in long paragraphs when you want short ones? Add the rule.
🔶 It keeps slipping an em dash into your copy when your style forbids them? Add the rule.

That last one is mine. I have a hard no dashes rule, and every time it broke, I wrote the correction down instead of re-explaining it. Now it is one line in a file, and every new session already knows.

Pro Tip: Spend the extra ten seconds. Fixing the output solves today. Writing the rule solves every day after.

Level Two: Enforce It With a Hook

Here is where it gets a little more powerful.

Some rules the model just keeps breaking. You wrote it down. It read it. And it still slips up now and then, because a written instruction is a suggestion the model is supposed to remember, not a wall it cannot cross.

That is what a hook is for.

A hook is a small check that runs automatically at a specific moment, on its own, whether or not the model remembers the rule. In Claude Code, you can set one to run right before the AI writes or edits a file. It can inspect what is about to happen and block it if it breaks your rule.

The difference is simple:

🔶 A rule is a note. The model reads it and tries to comply.
🔶 A hook is a guardrail. It does not care whether the model remembered.

Take my no dashes rule again. I can write it in my instructions all day, and once in a while a dash still sneaks through. A hook closes that gap for good: a tiny check that reads anything about to be saved and refuses it if it finds a dash. The model can forget. The hook cannot.

You would not do this for every rule. It is more setup than a one line note, so save it for the rules that actually matter, or the ones the model keeps breaking no matter how many times you write them down. For those, it is the difference between "usually right" and "never wrong."

I am keeping the setup out of this piece on purpose, because the idea matters more than the syntax. If you want my starter hook, there is a reply link at the end.

Level Three: Hand It Off

Now the part I have been most excited about, especially if you use more than one AI tool.

I run both Claude and Codex. They are good at different things, and I move between them depending on the task. For a long time, switching meant re-briefing the second tool from scratch. Everything the first one had learned about the job stayed locked in the first one's session.

Two things fixed that.

First, a shared instructions file. AGENTS.md is becoming the common standard that many tools read, and it is stewarded by the Linux Foundation, so this is not one vendor's walled garden. Codex, Cursor, Windsurf, and others read it natively. Claude Code reads its own CLAUDE.md, but you can point it at the same AGENTS.md with a single import line, so both tools work from one source of truth. Fix a rule once, and every tool you use inherits it.

Second, a handoff document. This is a short note you have your AI generate at the end of a working session. Not the whole transcript. Just the state:

🔶 What we did
🔶 What we decided, and why
🔶 What is still open
🔶 Where the important files live

Then when you open the other tool, you hand it that document, and it starts where the last one stopped instead of asking you to explain the whole project again. It is the same move as writing a rule, just aimed at carrying context across a tool switch instead of across a session.

Here is one I actually did.

A few weeks ago I had a big multi step build to hand off. Instead of trying to hold the whole thing in one long chat, I wrote a single handoff document first. It opened with one line that tells you exactly why this works: this document is self contained, you do not need the originating conversation.

Inside it was everything the next tool needed. The business context. Which existing pieces to reuse instead of rebuilding from scratch. The exact phased plan. And a list of known gotchas that would have tripped up anyone starting cold.

Then the actual build happened in Codex, on its own branch. I never re-explained the project once. The document was the briefing, and Codex just picked it up and ran with it. That is the whole trick: the handoff doc is the thing that remembers, so you do not have to.

Why This Is the Quiet Advantage

Step back and look at what these three levels really are.

Every seller has access to the same models now. Same Claude, same Codex, same everything. The model is not the advantage anymore. It is table stakes.

The advantage is everything you have taught your AI that the person next to you has not. Your rules. Your guardrails. Your handoffs. The accumulated corrections that make your AI a little sharper every single week, while everyone else re-explains the basics every time they open a window.

This is the amnesia fix I keep coming back to. Your AI does not remember anything on its own. But you can give it a memory, and the sellers who do are quietly building a tool that fits their business better every day.

The best part is how small the first step is. You do not need to set any of this up in an afternoon. You just need to stop throwing your corrections away.

Next time you correct your AI, ask one question: which rung does this belong on? Write it in your instructions file, enforce it with a hook, or add it to a handoff doc. Then it is fixed for good.

Know anyone who wants to work with an agency that has AI and tech in their DNA and are pros at managing advertising and creatives for E-commerce? Forward them this newsletter and let them connect with us at support@ppcninja.com.

NERD BYTES

Marie Kondo Your Claude Chats 🧹

Ever open Claude, stare at a wall of half remembered sessions, and give up on finding the one you actually needed? I’ve got to say, Search on Claude sucks and is not coming to your rescue.

Good news though! Claude Code desktop app just redesigned its sidebar, and now a little tidying is finally within reach. Started looking a bit like Codex honestly but it is really making my life easier.

🔶 Group the sidebar by project, so related sessions cluster under one heading instead of one endless list.

🔶 Pinned Chats help you move your most important chats to the top in a Pinned section
🔶 Open in a New Window allows you to open multiple browser like windows that you can detach away from Claude main and keep an eye on simultaneously running tasks.

🔶 Filter by status, project, or environment to shrink the list to just what you want right now. Honestly this is straight from the Claude documentation and I am still looking for this filter. If anyone figured this out hit reply!
🔶 Rename a session from the toolbar, so "the one about the scraper" is actually called that.

A clean workspace is not vanity. When you can find and resume the right session in two seconds, you actually reuse your past work instead of starting from scratch. That is the quiet advantage hiding in a tidy sidebar.

COOL TOOLS

Voice Finally Grew Up 🎙️

Voice is the highest bandwidth input we have. You can say about five times more than you can type in the same stretch of time. For a long while the tools did not deserve that bandwidth. In the last couple of weeks that changed, and the one you can actually use today is Claude.

Claude Voice: From Worst to Genuinely Good

Claude voice used to be the punchline. The reason: it only ran on the smallest, fastest model, so it got words wrong and felt shallow.

That is fixed. In its late July upgrade, Claude voice now runs on the full lineup, Opus, Sonnet, and Haiku, so you can talk to the smartest model, not just the quickest one. It reaches your connected tools, handles many more languages, and lets you switch models mid conversation. It is in beta across mobile, desktop, and web.

Talking a hard decision through with Opus out loud is a different thing than the old Haiku only version. This is the one to open first.

ChatGPT GPT-Live: Just Shipped, Still Rolling Out

OpenAI also rebuilt its interactive voice on a new model called GPT-Live, and on paper it is the more ambitious one. It parks a small voice window beside your work on the desktop app, can capture your screen, and can drive the Work and Codex surfaces by voice.

I say "on paper" because I could not get it on my account yet. That is expected. GPT-Live is going to consumer plans first (Free, Plus, Pro) and is not on Business or Team workspaces at launch, so if you live in a work workspace like I do, you will hit the same wall. I will report back once it reaches me.

One thing worth knowing now: the new GPT-Live model does not do live video or screen sharing. For that you drop back to the older Advanced voice mode.

Want to see it in action before it reaches you? The hands-on walkthrough that put me onto this is worth a few minutes.

The Catch Both Share

Voice gives you the bandwidth. Neither tool remembers your rules between sessions, so it starts fresh every time. The fix is this issue's main feature: put your rules in the file the AI reads, and voice starts every conversation already knowing them.

The other downside: These voice apps are a bit TOO chatty for the kinds of work I do. So I am going to stick to plain dictation for now.

VOICES FROM THOUGHT LEADERS

4 AI Hacks From the Seven-Figure Seller Summit 🎤

I joined the AI hackathon panel at the Seven-Figure Seller Summit, in a room full of eight-figure sellers and tool builders, each one demoing a live hack. Here are the best one-liners, and who shared them.

🔶 Moshe Shiner (eight-figure seller): Bulk-swap seasonal product images across your whole catalog through a tool wired to Amazon's API, using Perplexity Comet to build the prompt, then let it run while you work.

🔶 Andrew Erickson (seven-figure seller): Drop a free MCP into any AI you use (Claude, ChatGPT, or Codex) to pull top keywords, top competitors, and seasonal trends for any ASIN.

🔶 Samuel (BeCool, AI PPC agency): Turn one webinar or YouTube video into a LinkedIn post, an X thread, Instagram captions, an email, and a blog post with a single Gemini prompt.

🔶 Yours truly (Ritu Java): Use the =AI() formula in Google Sheets to run one prompt down hundreds of rows for listing rewrites, review mining, and negative keyword triage. It is a pipeline, not a chat.

Want to watch the full session? The Seven-Figure Seller Summit AI hackathon panel is worth an hour.

NEWS AND EVENTS WORTH FOLLOWING

If you have spent more than five minutes in the Amazon world, you know Kevin King. He is one of the sharpest, most generous operators in the space, and his events have a reputation for being where serious sellers go to actually level up.

His next one is Market Masters 4, August 20 to 24 in Austin, Texas. And it is not a sit-in-a-ballroom conference.

It is a 3.5-day mastermind at an 18,000 square foot mansion on Lake Travis. The format is built around "hot seats": you put your real business in front of a room of experts and walk out with a plan, not a tote bag of pamphlets.

🔶 Built for e-commerce operators at every stage, from pre-revenue founders to teams running $200M-plus.
🔶 Seven 90-minute workshops on AI, content, PR, email, and more.
🔶 Nine hot seats and thirty audience seats, all meals included. It is intimate on purpose.

I am flagging it because it is genuinely good, and because I have a little something for you. Use code EX1000 for $1,000 off your seat.

If you have been meaning to get out of your own head about your business, this is the room to do it in. You can see the full lineup on the Market Masters page.

Other News 📡

🔶 Claude Opus 5 is here. Anthropic's new top tier model landed on July 24, with the biggest gains for long running agents and coding work.

🔶 ChatGPT's cheapest model just got 5x cheaper. OpenAI cut the price of GPT-5.6 Luna by 80 percent on July 30, down to about 20 cents per million input tokens, which makes running AI work at scale far cheaper.

🔶 The Anthropic Economic Index is now a Claude connector. Turn it on in the connectors directory and ask what AI is actually used for in your role, straight from real usage data, with nothing to install.

🔶 Meta launched a Seller app for Facebook Marketplace. It brings AI listing creation, a unified inbox, and inventory tools, and anything you make posts straight to Marketplace.

🔶 Shoppable's one click checkout arrived inside ChatGPT. Its Universal Checkout plugin opens a catalog of 500 million products for multi retailer checkout without leaving the chat.

🔶 Sellers are deploying fleets of AI agents to protect margins. Forbes profiled one operator running dozens of agents for about 1,000 dollars a month to watch listings, ads, and market trends.

🔶 A cheaper, faster Nano Banana is rolling out. Nano Banana 2 Lite is Google's most cost efficient image model yet, now reaching the Gemini app and AI Mode for lower cost product shots.

We hope you liked this edition of the AI for E-Commerce Newsletter! Hit reply and let us know what you think! Thank you for being a subscriber! Know anyone who might be interested to receive this newsletter? Share it with them and they will thank you for it! 😃 Ritu

"Pest Control" for Your Gmail Inbox

Ritu Java — Tue, 21 Jul 2026 15:07:00 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
“Pest Control” for Your Gmail Inbox💥
The Shiny Object Syndrome 🤖
News and Events Worth Following

“Pest Control” for Your Gmail Inbox

Who here dreads opening their inbox to a wall of "just 15 minutes?" pitches?

Same. So I built something that makes them vanish before I ever see them. And here is the twist: it does not run on Claude.

Let me back up.

The pain that finally broke me

Every morning, same story. Between the emails that actually matter (clients, my team, invoices) sat a pile of cold pitches from complete strangers.

"Worth a look?" "Can I grab 15 minutes?" Backlink spam. Lead-gen offers. Someone selling me ad inventory for a sports site I have nothing to do with.

None of it was relevant. All of it wanted a reply. And every one I even glanced at was time I did not get back.

The worst part? Replying, even to say "no thanks," just tells them you are a real, live human. That is the kick they want. So the pitches multiply.

What I actually wanted

Not a spam filter. Spam filters catch the obvious junk. This stuff is sneakier: personalized, polite, and technically "legitimate" sales outreach.

I wanted a little agent that would do what I would do if I had the time. Read each new email in my Primary tab. Decide, "is this a cold pitch from a stranger, or something real?" And quietly move the pitches into Promotions where I never have to think about them.

Move them. Not delete them. Big difference. (More on that in a second.)

So I built exactly that. I call it Pest Control.

The twist: I reached for Gemini, not Claude

If you read this newsletter, you know I basically live in Claude Code. But this job had a catch. It needs to run on its own, every day, without me sitting there watching.

For a scheduled, unattended job, my Claude setup would not stay logged in in the background. Google's Gemini command line did. So the piece that reads each email and makes the call runs quietly on Gemini, on a timer.

Lesson I keep relearning: use the tool that fits the job, not the tool you are loyal to.

Fast, but with guardrails

Here is the part I care about most. An agent that moves your email around is terrifying if it gets it wrong. Bury one real client email and the whole thing is dead to you.

So I gave it three guardrails before it is allowed to touch anything:

🔶 It protects everyone I have ever emailed. On setup, it scans my own sent mail from the past year and builds an allowlist. If I have written to you, you are safe. Period.

🔶 It skips any conversation I am part of. If I have already replied in a thread, it leaves that thread alone. Real relationships never get swept.

🔶 When in doubt, it does nothing. Only a clear cold-sales pitch gets moved. Anything ambiguous (a podcast invite, someone who might actually know me) gets listed in a daily summary for me to eyeball. It never guesses.

That last one matters. The agent would rather leave a borderline email in my inbox than risk hiding something I needed.

The unsubscribe trick most people get wrong

When there was a safe way to unsubscribe, I wanted it to do that too. But carefully.

There are two kinds of unsubscribe. One is a one-click web link (safe, no human reads it on the other end). The other asks you to send an email to unsubscribe. That second one is a trap: replying confirms you are real, and usually gets you more mail, not less.

So Pest Control only ever uses the one-click web version. It never sends a reply. No kick, no confirmation, no "great, this inbox is active."

What happened when I ran it

I pointed it at my last seven days as a test. Here is the real tally:

🔶 It moved 9 cold pitches from 7 different senders out of my Primary tab and into Promotions.

🔶 It safely unsubscribed from the 3 that had a one-click link. The rest it just filed away, silently.

🔶 It correctly left my accountant alone. (A dumber, keyword-only version would have flagged her. The guardrails caught it.)

And then one surprise that taught me something.

In some cases, I find emails valuable, just not at that moment. I did not want to unsubscribe and burn the relationship. I just wanted the clutter gone from my primary inbox.

So I added one more rule: move to Promotions, but never unsubscribe. Declutter without cutting anyone off. That little nuance is the difference between a blunt tool and one you actually trust.

The safety net: it is all reversible

Every move gets logged. If it ever files something wrong, one command puts that day's mail right back in Primary. Nothing is ever deleted. Nothing is lost.

That is the "fast with guardrails" idea I keep coming back to. Move quickly, automate aggressively, but always leave yourself an undo button.

Now it runs on a timer every evening and sends me a short digest: here is what I filed today, here is what I was unsure about. I skim it in ten seconds and move on with my morning.

You do not need to build the whole thing

Want a taste without any code? Gmail already does a lightweight version.

Find a repeat offender, open Gmail's filters, and set "from this sender" to skip the inbox and land in Promotions. Same idea, just manual and one sender at a time. Good enough to feel the relief.

The full agent (the part that reads and decides for you) is the upgrade. And because the brain runs on Gemini, you do not need a Claude subscription to use it.

Want it to build the whole thing for you? You do not need to know how to code. Paste this starting prompt into your AI of choice (Claude, Gemini, ChatGPT, whatever you have open) and go from there:

Build me a small script that checks my Gmail Primary inbox once a day, uses an AI model to spot unsolicited cold-sales pitches from strangers, and moves only those into Promotions without ever deleting anything. Protect any sender I have emailed or replied to before, and only unsubscribe through safe one-click web links, never by replying. Then email me a daily summary of what it filed and what it was unsure about.

Start there, tell it which email account you use, and let it walk you through the setup.

The quiet advantage

Here is what I keep noticing. The tools that change your day are not always the flashy ones.

Nobody is going to be impressed that my inbox is quieter. But I get a little slice of my morning back, every single day, and I never feel that tiny jolt of "who is this person and what do they want from me."

That is the quiet advantage of building small agents for your own annoyances. Not one big dramatic win. A hundred tiny ones that add up while everyone else is still hitting archive.

So here is my question for you: what is the small, daily annoyance you would hand to an agent first?

Hit reply with "pest control" and I will send you the exact setup guide (the no-code version and the full Gemini build, both included). No Claude plan required.

COOL TOOLS

The Shiny Object Syndrome 🤖

Be honest. Every time a new model drops (GPT-5.6, Gemma, DeepSeek), you feel the itch to try it. But testing each one usually means a new account, a new API key, and a new little bill.

The fix has a name: a model aggregator (or AI gateway). One account that carries a whole shelf of models, so you reach any of them through a single door instead of a dozen.

The new option on my radar is Amazon Bedrock.

Claude, Llama, Gemma, OpenAI's GPT, Amazon's Nova, and dozens more, all pay-as-you-go. Want to line up three models on the same task? Pennies to find out which one wins.

Here is the honest breakdown:

🔶 The good: pay-as-you-go, zero subscription. Some models are almost free to run. Amazon's Nova Micro is around $0.03 per million input tokens. That is not a typo. For quick, high-volume jobs (classifying, tagging, sorting emails) it is close to free.

🔶 Also good: on-demand prices basically match what each provider charges directly, and batch jobs run at half price. So you are not paying a "convenience tax" to get everything under one roof.

🔶 The catch: it is an AWS product, so it feels like a developer tool, not a chat app. You make an account, request access to the models you want, and it is happiest when a script or an agent is calling it, not when you are typing into a box.

🔶 The surprise: the shelf is deeper than you would guess. When I looked, there were 51 models on it, from the full Claude line (Opus 4.8, Sonnet 5) to OpenAI's brand-new GPT-5.6, to a row of names you may not know yet (GLM 5, MiniMax, NVIDIA Nemotron). Yes, even OpenAI's models now sit right next to everyone else's, which was not true a few months ago.

About that rumor. You might have heard Claude is free on Bedrock is free for six months. I heard it too, could not find anything to back it up, and I have a call in to AWS to confirm this week. So do not count on free. The real story is better anyway: it is so cheap per use that for testing models, "free" barely moves the needle. (One honest note on the prices: Bedrock labels its own numbers "illustrative" and ties them to a single AWS region, so treat any figure you see as a ballpark, not a quote.)

Who it is for: anyone building little tools and agents (like my Pest Control sweeper from up top) who wants to try models without stacking five monthly subscriptions. If you only ever chat in a browser, keep your ChatGPT or Claude tab. But the moment you start automating, Bedrock is the cheap sandbox.

How to peek: create an AWS account, search "Bedrock" in the console, and request access to a model or two. That is the whole on-ramp.

My take? The win here is optionality. The "best" model changes every few weeks now, so not being locked into one is quietly worth more than picking the right one.

I have not run it yet, but it is high on my list. Hit reply with "bedrock" if you want me to test it and send back the setup.

VOICES FROM THOUGHT LEADERS

I recently went live with Augustas from Orange Klik and we talked about how Alexa for Shopping is changing the conversational e-Commerce landscape. Have a listen to this information packed episode!

NEWS AND EVENTS WORTH FOLLOWING

His next one is Market Masters 4, August 20 to 24 in Austin, Texas. And it is not a sit-in-a-ballroom conference.

I am flagging it because it is genuinely good, and because I have a little something for you. Use code EX1000 for $1,000 off your seat.

If you have been meaning to get out of your own head about your business, this is the room to do it in. You can see the full lineup on the Market Masters page.

Other News 📡

🔶 Catch me at the 7 Figure Seller Summit. On July 28 and 29 I am walking through 5 Claude skills you can use to run your PPC. It is 100% online and free to attend.

🔶 Salesforce's commerce agents just went live. Its Shopper, Buyer, and Merchant agents are now generally available and plug straight into ChatGPT, another sign that "AI does the shopping" is leaving the slide decks and actually shipping.

🔶 There is now a scorecard for AI shopping readiness. Digital Commerce 360 and ReFiBuy launched a quarterly ranking of how prepared the top 1,000 retailers are for agent-driven discovery, so "AI ready" just became something people measure out loud.

🔶 Amazon's AI shopping prompts are now a paid ad slot. Sponsored Products and Brands prompts exited beta, they bill under your existing CPC, and every campaign was auto-enrolled. Worth checking your Prompts tab if you have not looked.

🔶 AI already steers a fifth of online sales. AI influenced roughly 20% of global online sales last holiday season (about $262 billion), and retailers running their own shopping agents grew 59% faster than the ones sitting it out.

Fable Built Me a Wispr Flow Clone

Ritu Java — Tue, 14 Jul 2026 15:07:00 +0000

Fable Built Me a Wispr Flow Clone

Anyone else talk to their computer more than they type into it these days?

I do. Emails, messages, half of my Claude Code prompts. And for a while I was paying for the privilege.

I'd been using WisprFlow, the voice-to-text app that lets you dictate into any app on your Mac. It's a genuinely good product. But two things kept nagging at me every time I hit record. My voice was leaving my Mac, getting processed on someone else's servers, and coming back as text. And I was paying $144 to $180 a year for that round trip, with a weekly word cap waiting for me the moment I dropped to the free tier.

So this month I sat down with Claude Code and started building my own version. No subscription. No cloud. Just my Mac doing the whole job by itself.

And obviously, I had to try handing this problem off to Fable.

Fable, Anthropic’s most capable model

On June 9th, Anthropic released Claude Fable 5. Their most capable publicly available model yet.

Three days later, the US government slapped export controls on it. Anthropic had no way to verify who was on the other end of a login in real time, so they did the only thing they could. They pulled it. For everyone.

For eighteen days, the most capable model Anthropic had ever shipped publicly just... wasn't there.

Then on July 1st, it came back. Available again on Claude.ai, Claude Code, Claude Cowork, all of it.

What Fable Actually Is

Skip this if you already know. Otherwise, here's the short version.

Fable 5 is built for the projects you'd normally hand to a person, not a prompt. Anthropic points to software engineering, knowledge work, and vision as its strong suits. Run it inside an agent harness like Claude Code, and it can keep working for days at a stretch instead of stopping after one exchange.

It also comes with hard limits baked in. Ask it something that touches cybersecurity, biology, chemistry, or a handful of other high risk categories, and it declines and falls back to Opus 4.8 instead. Anthropic built the guardrail into the model, not just the policy.

That combination (long runway, real capability, a fence around the dangerous stuff) is why the export control story mattered in the first place. This wasn't a toy getting pulled.

The Build

My FlowDictate app is a local Wispr Flow clone for macOS. The functionality is simple. Hold down a key, speak, release: text appears at the cursor in any app. All processing on-device.

The problem I was trying to solve is that many apps do not come with their own voice dictation software. I wanted every surface on my Mac to be voice-dictatable.

Here's the part I find genuinely cool. Every piece of this runs locally, for free, using tools that already exist.

🔶 WhisperKit, running OpenAI's open source Whisper model, handles the actual transcription, on my machine, with no internet required. (NOTE: WhisperKit, different from WisprFlow, is an open-source Swift package maintained by Argmax, MIT licensed that runs OpenAI's Whisper speech-to-text model)
🔶 Apple's on-device Intelligence model (built into every current Mac) cleans up the transcript afterward, stripping filler words and fixing punctuation
🔶 A simple hotkey (hold Ctrl, talk, let go) triggers the whole thing and drops the finished text wherever my cursor is sitting (later I switched the hotkey to Alt to prevent triggering each time a common CTRL+key combo was used)

No API bills. No per-word pricing. No model to babysit.

Fable’s Implementation Architecture: Steal this!

The Part That Actually Made Me Do This

Cost was the easy argument. Wispr Flow runs $144 to $180 a year per person, and that is before you even think about rolling it out to a team. My version costs $0 a month, forever, with no cap on how much I dictate.

But privacy is the part that tipped me over. Wispr Flow processes every recording through AWS and a third-party AI model before it comes back to you as text. Even their optional "Privacy Mode" still sends the audio out first, it just claims to delete it after. And earlier this year, Wispr had a publicized data exposure through a third-party compliance vendor, which is exactly the kind of thing that happens when your voice has to leave your machine to get turned into words.

Mine never leaves. Nothing to intercept, nothing to leak, because there is no server in the loop at all.

The Quiet Advantage

The part I did not expect to love: I can tune the vocabulary myself. I dropped in a list of terms my transcript kept mangling, ACOS, ASIN, ROAS, and now it spells them correctly every time, because I own the file, not a support ticket queue.

I'm still in the build and test phase, so I won't hand you exact latency numbers yet. But the architecture is done, and it already tells me something bigger. A lot of the tools we pay for monthly are really just a thin wrapper around models and APIs that are sitting right there for anyone willing to build the wrapper themselves.

(PS: I also replaced 2 heavy duty accounting apps (DEXT and Synder) last week with Fable, saving me $2500/year in subscription costs)

What's a tool you're paying for every month that might just be a weekend build away?

NERD BYTES (FOR NERDS)

Codex, Claude and a battle of models

GPT-5.6 (Sol, Terra, Luna) launched July 9th. And Anthropic's "Fable 5 is free" promo, originally set to end July 7th, has been extended twice since. First to July 12th. Then to July 19th.

That's not a coincidence. That's moving the goalposts before anyone notices the free window closing.

🔶 Fable 5, Anthropic's flagship, launched June 9th, briefly pulled by export controls, back since July 1st and included at no extra cost for up to 50% of weekly usage limits on Pro, Max, Team, and select Enterprise plans, for now
🔶 Sol reportedly beats it on general agentic benchmarks. Fable still leads on real coding work, where it counts
🔶 Grok 4.5 is the wild card, trained alongside Cursor, undercutting both on price

The quiet advantage: you're building anything real this month in the middle of a price and access war nobody announced outright. The free window won't last. While it's open, it's the cheapest shot you'll get at testing Anthropic's best model on a real project.

COOL TOOLS

This week Google turned the plain search box into a multimodal AI tool. Click the plus icon next to the search bar and you'll find Upload image, Upload file, a Create Images option, and a straight up Gemini 3 model picker (Fast or Pro), sitting right on the homepage. Not tucked inside a separate AI Mode tab. The front door, not a side door.

Google's own framing for the change: "the biggest upgrade to our Search box in over 25 years," part of the I/O 2026 rollout.

🔶 Upload gets analyzed, not just reverse searched. Drop in a competitor's product photo and ask what's actually working in it
🔶 Files work too. A PDF or spec sheet, no copy pasting text out first
🔶 Create Images runs Nano Banana without leaving Search
🔶 Chrome tabs count as input, per Google's own announcement, alongside text, images, files, and video
🔶 It's free, same as Search always was

It also allows you to search within recent tabs. If you click on that + icon you will now be able to search within that tab.

Like here, I asked it a question about one of my open tabs:

The catches: it can't see behind a login, so Seller Central and anything else authenticated is off limits. It's not built for long multi turn conversations the way a real chat session is. And rollout is still uneven depending on region and account.

VOICES FROM THOUGHT LEADERS

I recently sat down with Andrew Bell, host of RetailPlaybook, for a conversation on where AI shoppers and agentic commerce are taking retail next. Podcast rolls out on the 15th and I am honored to be Andrew’s first guest!

If the name is new to you, it shouldn't stay that way for long. Andrew is VP of Research at ReFiBuy and one of the leading independent researchers tracking Amazon's Alexa for Shopping. He published the original Alexa for Shopping Blueprint, writes the LinkedIn newsletter Amazon Science Made Simple, and his frameworks for how shopping agents interpret and recommend products have been picked up by Forbes and referenced across the seller and retail tech world.

In the meantime, here is a sneak peak into a section of our conversation: How Alexa for Shopping is Transforming Product Discovery

NEWS WORTH FOLLOWING

🔶 Amazon's new Canvas gives sellers an AI dashboard built on Bedrock and Claude. Free, live now for US and UK sellers, and Amazon says sellers accept its recommendations nearly 90% of the time.

🔶 AI-referred shoppers converted 40% better than search, email, or social during Prime Day 2026. First time that's ever been true, per Adobe's data. The friction that used to sink AI-driven traffic is gone.

🔶 Walmart's Sparky agent is now live inside ChatGPT and Gemini. Baskets built with it run about 35% higher in average order value. Not an Amazon story, but every seller should be watching where else shopping agents are showing up.

🔶 Amazon raised $25 billion in bonds to fund its AI infrastructure buildout. The scale of the bet is a good sign of how permanent this shift is meant to be.

🔶 Guideline can now track verified ad spend on ChatGPT and Perplexity. Roughly $200 billion in annual media investment, now measurable. The AI platforms are becoming real ad channels, not just chat interfaces.

AI Found a Gap Between Our Two Best Reports

Ritu Java — Tue, 30 Jun 2026 19:53:45 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Claude Found the Gap Between Our Two Best Reports💥
How We Made Sure We Did Not Go Out of Budget on Prime Day
ChatGPT’s New Schedules Feature!
News Worth Reading

Claude Found the Gap Between Our Two Best Reports

Every week, new keywords show up in our clients’ Search Query Performance reports.

Real customers, typing real searches, and Amazon is already putting our products in front of them. For terms we had never added to our keyword rank tracker.

And we used to miss some of them.

Not because the data was hidden. It was right there in SQP. We just were not catching the new ones fast enough to start tracking them and bidding on them. Why? Because it is a manual process, and manual processes are always flawed.

So we built a Claude skill to fix it. Then we turned it into a Routine so it does the catching for us.

Why SQP Is the List That Matters

If you have Brand Registry, you have Search Query Performance inside Brand Analytics.

It is not a keyword tool guessing at search volume. It is Amazon telling you the actual queries customers used, and whether your product showed up for them.

That is the part most people skip past. These are not keywords you brainstormed. These are keywords Amazon already decided your product is relevant for. The demand is proven. The only question is whether you are tracking it and ranking for it.

New ones appear all the time. Buyer language shifts. Seasons change. A term nobody was using last month starts driving impressions to your listing this month.

If your rank tracker only watches the keywords you already knew about, you never see the new ones coming.

What We Built

We built a Claude skill that compares the SQP keywords against the keywords already in our rank tracker, and flags anything new.

No exporting. No copy paste. No sitting there with two tabs open trying to spot what changed.

The reason it can do that is simple: both reports already live in our data warehouse (Big Query). The SQP data and the rank tracking data are sitting in the same place. So the skill just reads both, lines them up, and tells us what is new. Nothing to download, because nothing left the warehouse to begin with.

The output from Claude is simple. It tells the keywords showing up in SQP reports that you we not tracking yet.

SQP —> Rank Tracker

Then We Made It a Routine

The skill worked. But someone still had to remember to run it.

So we turned it into a Routine. Now it runs on its own and tells us what is new. Our standard approach is to send these auto-generated emails to the right team members for their accounts.

Why This Matters Beyond SQP

Here is the thing that stuck with me.

The skill is not just about SQP. It is about the gap between any two reports that should be compared and almost never are because of the manual effort required.

You have plenty of these. Your search term report and your negative keyword list. Your prices and a competitor's. Your reviews and your own bullet points. Each one is fine on its own. The signal is in lining two of them up, and that is exactly the work that keeps getting pushed to next week.

Claude is good at that. Reading two sources and telling you where they do not match. That is the quiet advantage.

NERD BYTES (FOR NERDS)

How We Made Sure We Did Not Go Out of Budget on Prime Day

If you run super successful, money-printing ads on Amazon, you know the worst mistake you can make is running OOB (Out of budget) on them.

Your campaign hits its cap, the ads stop, and you go dark right when shoppers are still buying. On a normal day it stings. On a big traffic day it is money walking out the door.

We did not want to find out hours later that we had gone dark. So we built a Claude Routine to watch for us.

It runs every three hours through the event and does one job: tell each person on my team exactly which of their campaigns have gone out of budget, so they can top them up fast.

But a raw list of every OOB campaign is just noise (also, we could have made it hourly but we felt that was too noisy for the team). So we made it smart about what actually counts:

🔶 It only flags campaigns that are genuinely out of budget and still turned on
🔶 It skips campaigns whose whole portfolio is already capped. No point chasing those, the limit is somewhere else.
🔶 It ignores campaigns that have not been spending anyway. Dead weight does not need an alert.
🔶 It groups what is left by account, so each person gets only their own campaigns to deal with.

Here is the part I am proudest of. It only emails you when your list has changed since the last check. If nothing moved in the last three hours, you get nothing. No alert fatigue, no inbox full of identical snapshots. Just a tap on the shoulder when something actually needs you.

Think about how different that is from babysitting a dashboard. Nobody had to remember to look. The Routine watched, filtered out the noise, and pinged the right person at the right moment.

Now, could we have used an Amazon Budget rule for this? Sure, but that would be WAY limiting in triangulating our specific set of requirements of noise. Plus, it would miss the side bonus of team members noticing which campaigns are actually running out multiple times in a day with our alert counter.

The lesson holds beyond one event. Most people manage budget by checking whether it already ran out. The better move is to let something watch the pace for you and speak up only when it matters.

COOL TOOLS

ChatGPT’s New Schedules Feature!

Here is a small feature that quietly took a task off my plate.

It is called ChatGPT Schedules. You give ChatGPT a job, you tell it when to run, and it emails you the result. That is it.

I use it to keep tabs on a product I follow closely. I told it to search every week for any real news: can people test it, demo it, join a waitlist, preorder, or buy, and to email me a summary. Now every Monday it just shows up in my inbox.

I could have built this with Zapier. But honestly, why bother. This was easier, and I am already paying for ChatGPT.

How To Set One Up

You will find Scheduled Tasks inside ChatGPT (you need a paid plan).

You write the instruction in plain English, pick how often it runs, and you are done. It emails the results to you.

A Few Ways Sellers Could Use It

🔶 Watch a competitor. "Every Monday, search for any new product launches, price changes, or review trends for [brand], and send me a summary."

🔶 Take the pulse of your niche. "Every Friday, search for news in [category]: new entrants, waitlists, preorders, big price moves, and send me a short summary."

🔶 Keep an eye on your own brand. "Every week, search for mentions of [your brand] across news, forums, and social, and flag anything I should see."

One Honest Limit

It searches the public web. It is not looking inside Amazon, so it will not catch BSR shifts or ad changes. For keeping up with public news on a competitor or a category, it is great. For your own account data, you still go to your reports (or, as above, a skill that reads them).

Have you tried Schedules yet? What would you point it at? Hit reply and let me know!

NEWS WORTH FOLLOWING

🔶 The market is shifting from "tokenmaxxing" to efficiency. Buyers are done paying for raw power and want right-sized model spend instead. The same logic behind picking the cheapest model that gets the job done.

🔶 Alexa+ can now close a sale inside the ad itself. New agentic ads on Echo Show let a shopper buy through a spoken conversation, no click and no second screen. The funnel is shifting under our feet. (June 23)

🔶 Prime Day 2026 set a spending record, but with smaller carts. Record transactions, lower average order value than last year. More buyers, thinner baskets, tighter margins to manage. (June 27)

Claude Works Nights. You Take the Weekend.

Ritu Java — Tue, 23 Jun 2026 15:07:00 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Claude Works Nights. You Take the Weekend.💥
Stop Paying Opus Prices for Haiku Work 🤓 🤓
Midjourney V8.1 vs Nano Banana 2 🎬
News Worth Reading

Claude Works Nights. You Take the Weekend.

Who here feels like the work never actually stops?

Well, lucky you, because today I am going to show you how to hand off the parts that repeat. To something that keeps working long after you have closed the laptop.

Here is the thing most sellers do not realize. You can hand Claude a task, walk away, go live your life, and come back to a finished summary sitting in your inbox. No terminal left running. No machine humming in the corner. No babysitting.

This is called a Routine. And once you set one up, it changes how you think about repetitive work.

First, What Is a Routine?

A Routine is a Claude Code task you set up once and then let run on its own on a pre-determined frequency. You give it a prompt, point it at the data it needs, and pick a schedule. After that, it runs without you.

The magic word here is Remote. Routines run on Anthropic's own infrastructure, not on your computer. That means your laptop can be closed, asleep, or in your bag at the airport. The task still runs.

Compare that to the old way. Before, if you wanted Claude to do something on a schedule, your machine had to be on and a session had to be open. Forget to leave it running? Nothing happens.

Routines remove that fragility entirely.

The Three Ways to Schedule (And When to Use Each)

Here is where most people get confused, so let me lay it out clean. There are three ways to put Claude on a schedule, and they are not the same.

	Cloud Routine	Desktop Task	/loop
Runs on	Anthropic cloud	Your machine	Your machine
Laptop can be off?	Yes	No	No
Open session needed?	No	No	Yes
Sees your local files?	No (fresh copy)	Yes	Yes
Shortest interval	1 hour	1 minute	1 minute
Best for	Set and forget	Local file work	Quick polling

Read that table twice. It is the whole framework.

🔶 Cloud Routine: use this for anything that should run reliably whether or not you are around. Weekly summaries, nightly checks, scheduled reports. This is the one that works while you sleep.
🔶 Desktop Task: use this when the job needs your actual files on your actual machine. It runs locally, so the computer has to be on.
🔶 /loop: use this for quick checks during a session you are already sitting in. It needs the window open, so it is not a leave-it-overnight tool.

[Heads up: Cloud caps out at one run per hour no matter what, but Desktop only defaults to hourly. Ask Claude in plain English and a Desktop task will go down to one minute.]

For the "walk into Monday with it done" use case, you want a Cloud Routine. Every time.

How I Set One Up

I wanted a simple thing. Every Sunday at 7am, pull my key numbers from BigQuery, write a tight summary, and email it to me before I am even awake.

Here is the experience, start to finish.

I went into Claude Code on the web and opened the Routines area. Creating one is a simple form: the prompt to run, what data it should reach, and the cadence. I picked weekly, Sunday morning. I described the summary I wanted in plain English, the same way I would ask a teammate.

Then I closed my laptop and forgot about it.

Sunday morning, the email was there. Clean summary, the numbers I asked for, ready before coffee.

That is the part that still feels a little magic. I did nothing. The work just showed up.

Pro Tip: describe the output format in your prompt as specifically as you can. "Three bullets, top metric first, flag anything that moved more than 10 percent." The more precise the brief, the better the Sunday email reads.

The Guardrails You Should Know

I would be doing you a disservice if I only showed you the shiny part. So here are the honest limits.

🔶 There is a daily cap. Pro plans get 5 routine runs a day, Max gets 15, Team and Enterprise get 25. Plenty for most sellers, but worth knowing before you schedule ten of them.
🔶 The shortest cloud interval is one hour. Routines are built for "every night" and "every Monday," not "every five minutes." For fast polling, that is what /loop is for.
🔶 Cloud Routines run in the cloud, not on your machine. A Routine cannot reach files sitting on your local desktop. It gets what it needs through connectors (Slack, Google Drive, and the like) and whatever sources you point it at when you set it up.
🔶 They run on their own. No approval prompts mid-task. So write the prompt carefully and start with low-stakes, read-only jobs (summaries, reports) before you let one take any real action.

None of these are dealbreakers. They are just the shape of the tool. Match the job to the tool and you are fine.

Why This Matters for Sellers

Step back and look at what just happened. A task that used to need your attention, your open laptop, and your memory to kick it off now runs itself on a schedule you set once.

That is the difference between doing the work and designing the work.

The seller who sets up three Cloud Routines (a Monday summary, a midweek check, a Friday wrap) has effectively hired a junior analyst who never sleeps and never forgets. The seller who does not is still opening spreadsheets by hand.

This is the quiet advantage I keep coming back to. The tools are here. The access is here. Most people just have not set them up yet.

Your laptop deserves a weekend. So do you.

Know anyone who needs an AI-first agency to build cool, creative videos for E-commerce? Forward them this newsletter and let them connect with us at support@ppcninja.com.

NERD BYTES (FOR NERDS)

Stop Paying Opus Prices for Haiku Work 🤓

Quick question. Which Claude model are you using right now?

If your answer is "the one it picked for me" or "the same one I started with months ago," this one is for you.

Here is what most people miss. There are three Claude models on the menu, and they cost wildly different amounts. Using the most expensive one for simple work is like taking a sports car to pick up groceries. It works. You just overpaid.

Here are the current API prices, per million tokens:

🔶 Haiku 4.5: $1 in, $5 out. The cheapest. Fast and great for bulk, repetitive jobs.
🔶 Sonnet 4.6: $3 in, $15 out. The middle. The workhorse for most real tasks.
🔶 Opus 4.8: $5 in, $25 out. The most powerful. Save it for the hard stuff.

Look at that spread. Opus output costs five times what Haiku does. If you are running 500 product description rewrites a week on Opus, you are lighting money on fire for work Haiku handles fine.

So here is the simple rule of thumb:

🔶 Simple, high-volume, repetitive work: ... Reach for Haiku.
🔶 Everyday real work: ... Reach for Sonnet. It is the workhorse for the bulk of what you do. Rumor has it that Sonnet 5 is on its way (gaining some of Fable’s capabilities without triggering the powers above)
🔶 The genuinely hard stuff: ... That is where Opus 4.8 earns its keep. It scored 84 percent on a leading browser-agent benchmark, the best of any model tested.

One more lever. If you run big batches that are not time sensitive, the Batch API takes 50 percent off both numbers. Half price for work you do not need back this second.

The caveat: do not cheap out on the tasks that matter. A bad competitor analysis costs you more than the Opus tokens would have. Match the model to the job, do not just minimize the bill.

That is the whole trick. The model picker is the cheapest optimization most sellers never touch.

COOL TOOLS

Midjourney V8.1 vs Nano Banana 2 🎬

If you make product images with AI, two tools keep coming up. And people keep asking me which one to use.

The honest answer is: it depends what you are making. So let me put them side by side.

Midjourney V8.1 dropped April 30. It is fast now, around 4 to 5 times faster than the older version, with HD 2K output and a new "Raw mode" built for realistic lighting and product visualization. It also finally lives on the web. Discord is now optional.

Nano Banana 2 is the Google model most of us already use for listing photos. Cheap, quick, and accurate where it counts.

The Scoreboard

	Midjourney V8.1	Nano Banana 2
Speed	4 to 5x faster than V7	~10 to 15 seconds
Cost	From $10/month (GPU hours)	~$0.03 per image
Resolution	HD 2K	Production ready
Text on packaging	Often gibberish	Better, readable
Best at	Mood, lighting, concept	Accurate product shots
Web access	Yes (Discord optional)	Yes

Where Midjourney Wins

🔶 Atmosphere and mood. Cinematic lighting, editorial styling, scenes that feel like a magazine shoot.
🔶 Concept and ideation. Raw mode gives you direct control to explore looks before you commit.
🔶 Lifestyle backdrops. Beautiful environments to drop a product into.

Where It Falls Short

🔶 Text and logos. Packaging copy comes out as gibberish. Reflections sometimes defy physics.
🔶 Branding accuracy. Not reliable enough for the actual main image on a listing. 🔶 Consistency. Holding the same product look across many images is still hard.

So Who Is This For?

Here is my take. These are not really competitors. They are teammates.

Use Midjourney V8.1 for the dreaming: mood boards, lifestyle concepts, creative direction, the hero-shot vibe you are chasing. Then use Nano Banana 2 for the doing: the accurate, on-brand product images that actually go on your listing where the text has to be correct and the details have to be real.

Midjourney shows you the destination. Nano Banana gets the product there accurately. Reach for the right one and you stop fighting either of them.

NEWS WORTH FOLLOWING

🔶 Amazon's Ads AI image generator adds aspect ratio control. More flexibility for sellers generating creative directly inside the Ads console.

🔶 Amazon now shows AI-generated product images inside mobile search. As of June 3, US mobile shoppers see AI-rendered "inspiration" thumbnails sitting right next to real listings. Time to audit your main images at thumbnail size and finish your structured attributes.
🔶 Google retired Gemini CLI on June 18. If you had any script calling the gemini command, it broke with no grace period. The replacement is Antigravity CLI.
🔶 Antigravity 2.0 brings agent teams and a Planning mode. Google's new agent-first platform pairs a 1M context window with squads of subagents working in parallel.

Claude Chat, Cowork or Claude Code?

Ritu Java — Tue, 09 Jun 2026 17:34:39 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Claude Chat, Cowork or Claude Code?
The Magic Google Sheets formula for AI at Scale
Higgsfield Supercomputer
News Worth Reading

Claude Chat, Cowork or Claude Code? It Was Never Three Doors

Have you ever paused, cursor hovering, not sure whether to open Claude Chat, Cowork, or Claude Code?

You are not alone. And the hesitation is not your fault. No one explained it to you.

I have been living in these tools every single day, and the three way comparison of which one to use when just makes a simple decision feel complicated, and it hides the one distinction that actually matters.

The Only Line That Matters

Forget features. Ask one question about any task in front of you.

Am I talking, or do I want it done?

🔶 Talking is single turn. You ask, it answers, you think together. You do not expect Claude to wander off and run a ten step job alone. That is Chat.
🔶 Doing is multi turn. You hand over an outcome and expect Claude to plan it, take many steps, use tools, check its own work, and come back finished. That is the agent.

Chat is a conversation. The agent is a worker. The moment a task has more than one step and you would rather not babysit each one, you have crossed from talking to doing.

Cowork and Code Are One Engine

Here is the part nobody says plainly. On the doing side, Cowork and Code are the same thing wearing different clothes.

Anthropic built Cowork using Claude Code. Same engine underneath. The only real difference is control versus convenience. Code gives you the full set of controls. Cowork hides them: point at a folder, describe the outcome, walk away.

Cowork is Claude Code with training wheels. No shame in that, it is how you learn to ride. And the old excuse for needing it, that Code is a scary black terminal, is already out of date. Claude Code now runs as a friendly desktop app, in your browser, and on your phone. The terminal is one face of it, not the whole thing.

The Part Almost Nobody Has Seen

Now the real reason I am writing this.

Hardly anyone uses the doing side. Most people live in Chat, type the same request over and over, copy answers in and out, and never hand Claude a whole job. They have no idea what they are missing, because the agent is not a better Chat. It is a different league. Three things make it so.

🔶 Multi turn. It works in a loop. Plan, act, check the result, adjust, keep going until the job is actually done. You describe the finish line, not the steps.
🔶 Subagents. It splits a big messy job into many helpers that work in parallel, then pulls the results together. You manage none of it.
🔶 Dynamic Workflows. Brand new with Claude Opus 4.8 on May 28. Inside Claude Code, one session can now decide on its own how many helpers it needs, spin up hundreds at once, and synthesize everything into a single answer.

One honesty note: Dynamic Workflows is a research preview, and it landed on the Code side first for the higher plans. But look at the direction, not the fine print. The agent is being built to run a small army on your behalf.

In seller terms: "go through all 2,000 of my reviews, tag each one by complaint type, find the patterns by product, and draft me a fix list." Chat makes you feed it in batches and stitch the output yourself. The agent splits it across helpers and hands you the finished report. Same model, completely different experience.

But here is the kicker. Claude Code can do the talking too. Once you live in it, you just ask, and you stop opening Chat at all. I barely do. Chat is not wrong, it becomes the convenient front door: the tab you open on your phone for a ten second question. The capability now lives in one place.

My Prediction, and the Catch

Here is my prediction, and I will call it exactly that.

I think Cowork and Claude Code eventually fold into one product, maybe under a new name. One agent, with a simple face for convenience and deeper controls for power, no longer two brands you have to choose between.

Now the part that argues against me. Anthropic is not signaling that today. They just doubled Cowork usage limits for a month, and the new Opus 4.8 effort controls landed in Cowork too. That is a company investing in a product, not retiring it. So Cowork is not going away. Watch the trajectory, not the press release.

Know anyone who needs an AI-first agency to build cool, creative videos for E-commerce? Forward them this newsletter and let them connect with us at support@ppcninja.com.

NERD BYTES (FOR NERDS)

One Formula, Hundreds of Rows 🤓

Quick story. I entered a hack contest called 15x5 a little while back and walked away with 4th place. My whole hack was a single Google Sheets formula. =AI().

Here is the trick. In any Google Sheet, you can now type =AI("your prompt", and point it at a cell. The model runs right there, inside the cell. Then you grab the corner and drag it down, and it runs on every row. One prompt, hundreds of rows, as fast as you can pull.

It is not a chat. It is a pipeline.

A few of the demos I showed:

🔶 Listing rewrites. =AI("Rewrite this as an SEO-optimized Amazon title under 200 characters, front-load the main keyword: " & A3). Drag down. Every title rewritten.
🔶 Review mining. Three columns, three prompts: sentiment (positive, negative, neutral), the single biggest complaint, and a theme tag (quality, shipping, size, battery). A thousand reviews sorted while you watch.
🔶 Search term triage. Feed it your product and ask "relevant, or a negative candidate?" It flags "plastic water bottle" as a negative for your stainless steel one, and tells you why.
🔶 The Magic Row. One row that generates a title, bullet points, a German translation, and backend keywords all at once. Drag it across your whole catalog.
🔶 CRM touch. Pull the likely company name from an email domain, then draft a friendly two-sentence reorder reminder for whatever that customer last bought.

So why do this in Sheets instead of just opening a chat? Because it scales with a drag. The output feeds straight into your filters, IFs, and pivots, so one cell can drive the next step automatically. And it stays live and shareable. Change a cell and the row regenerates. The prompt sits in the formula bar for your whole team to see, tweak, and reuse. Chats disappear. A sheet has permanence.

The honest ceiling: there are daily rate limits, but they are not small. You can generate hundreds and hundreds of cells before you bump into them. It does need Gemini in your Google Workspace, and even a basic Business Standard account works.

Want the exact prompts from the demo? Hit reply with the word sheets and I will send them over.

This is the quiet advantage I keep talking about. The model has been one cell away the whole time. Most people just never typed the formula.

COOL TOOLS

Higgsfield Supercomputer

One chat that builds the whole brand, not just the video

You may know Higgsfield as a video tool. They just went much bigger.

Higgsfield Supercomputer is their new cloud native, self learning AI agent. One chat, more than 40 built in tools, three layers of memory, and it does not lean on a single model. It orchestrates the best of each: GPT-5.5, Claude Opus, Gemini, Seedance, Veo, Kling. You describe an outcome, it plans the steps, picks the right tool for each one, and runs the whole thing end to end. You can drive it from a browser, or from Telegram on your phone.

I will let the demo do the talking. The Amazon workflow is the one that should make every seller sit up.

The "Build Me a Brand" Workflow

In the demo, the prompt is almost lazy: suggest three unbranded Amazon products worth selling, with price, margin, and competition.

What happens next is the interesting part. The agent browses Amazon best seller lists live, runs the margin math (wholesale cost, FBA fees, profit per unit), and reads through thousands of reviews to find what shoppers keep complaining about. It lands on dog socks, and pegs them at a 19 dollar sale price, about a dollar to make, with tens of thousands of monthly searches and no real branding in the niche.

Then, from the same chat, it builds the brand. A name, a tagline, a product sheet, then a full asset kit: hero image, lifestyle shots, infographics, a size guide, a comparison card, logo, colors, typography. After that it studies the top performing Meta, TikTok, and YouTube ads in the niche (it actually watches them, audio and hook timing included), then writes and generates 15 platform native video ads and pings the finished files to Telegram when they are done.

The video also shows it making a short animated film from one line, and coding plus deploying a live website with its own free hosting. But the brand build is the one that matters for us.

Why Sellers Should Care

🔶 It collapses a five specialist process into one chat. Niche research, brand identity, product photography, listing images, ad creative. The work that usually costs thousands and takes weeks.
🔶 It orchestrates the models, so you do not have to. You are not choosing between GPT, Claude, and the video engines. It routes each subtask to the right one.
🔶 It runs while you are away. Kick off a job, get pinged on Telegram when it is finished.

But Stay Honest About It

🔶 That is a vendor demo, their best take. Smooth on camera does not mean smooth on your first try. Budget for misfires and re runs.
🔶 The numbers are AI estimates, not gospel. The margin and the search volume are a starting hypothesis. Verify every figure yourself before you order a single unit.
🔶 "Production ready" is their words. AI ad creative still needs your eye. Most clips are throwaways, and your taste is the filter that keeps the good ones.
🔶 Autonomy cuts both ways. An agent that signs up for hosting and makes creative calls on its own is powerful and a little unnerving. Watch what it touches, especially anything tied to accounts or money.

So Who Is This For?

Sellers who want to test a new product or a fresh brand fast and cheap, and who can already tell a good asset from a bad one. It does not stand in for your judgment. It removes the grunt work between your idea and a first draft of everything.

My Take

This is the clearest picture yet of where things are heading. Not a tool you operate, an agent you delegate to. The dog socks brand it spun up is genuinely impressive, and also exactly the kind of thing that floods Amazon with sameness if everyone runs the same prompt.

So the advantage is not having the agent. Soon everyone will. The advantage is the taste to point it at something real, and the eye to keep only the good ten percent. Treat it like the fastest junior team you have ever had, not a substitute for knowing your customer.

Want to see it build a brand live? The full demo is here. If you test it, hit reply and tell me what it made you, the honest result, not the highlight reel.

NEWS WORTH FOLLOWING

A few things from the past week that actually matter for sellers. (Editor's note: refresh links the morning of send.)

🔶 Amazon's AI shopping assistant tops 250 million shoppers ahead of Prime Day. AI is now a primary way people discover products on Amazon, and Prime Day 2026 runs June 23 to 26. If your listing does not read well to the assistant, you are invisible to a quarter billion shoppers.

🔶 Amazon search now renders AI-generated product images in the US app. As of June 3, descriptive searches can surface invented, prompt-built product images that are not real listings. Worth watching closely, because it changes what "showing up in search" even means.

🔶 Amazon starts selling its Rufus-based shopping AI to other retailers. Kate Spade is an early customer. The same assistant tech that reshaped Amazon search is about to show up across the rest of the web.

🔶 AWS launches the Agentic Shopping Assistant for any retailer. Built on the tech behind Amazon's own assistant, which the company credits with nearly 12 billion dollars in incremental sales last year. Agentic commerce is leaving the Amazon walls.

🔶 Sponsored Products CPC hits a record high before Prime Day. Average CPC has climbed about 35 percent over three years. The bid strategy that worked in spring may already be underwater, so audit before the Prime Day surge.

🔶 Sellers are quietly using AI to grow sales and cut costs. A good roundup of where AI is actually moving the needle for sellers right now, not in theory but on the P&L.

🗞️ Blotato + Claude MCP for Social Posting💥🚀

Ritu Java — Tue, 02 Jun 2026 15:39:55 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Blotato + Claude MCP for Social Posting💥
The $20 Plan That Stretches Like $100 🤓
Wait, Claude Can Animate?
News Worth Reading
AI Related Events worth Tracking

Blotato + Claude MCP for Social Posting💥🚀

Who here is tired of posting the same announcement to six platforms by hand?

Well, lucky you, because today I am going to show you a game-changer.

You write your post once. You hand it to Claude. Claude pushes it to LinkedIn, X, Instagram, Facebook, Pinterest, and wherever else you sell, all in one breath. No tabs. No copy paste. No logging into a separate scheduler.

The bridge that makes this work is a tool called Blotato, wired into Claude through something called an MCP. And I will be honest with you. Of all the MCPs I have tested, this is the simplest one to set up.

First get yourself a free Blotato account to play with:

Blotato Makes Cross-Posting a Breeze

You have a launch, a restock, a sale. So you paste it into Facebook. Then trim it for X. Then make a square image for Instagram and a vertical one for Pinterest. Twenty minutes gone, and you do it again tomorrow.

Most sellers pay this tax without noticing. Almost nobody has wired it into the AI they already use every day. That last group has the quiet advantage.

Blotato also comes with a bunch of AI features. It generates the visuals, spinning up branded images from templates so you are not opening a separate design tool for every platform. It repurposes a source, turning a video or a link into ready to post social copy. And it adapts your post per platform, reshaping one message into native versions for LinkedIn, X, Instagram, Pinterest, and the rest instead of cloning the same text everywhere. Wired into Claude, the AI writing, the AI image, and the AI formatting all run from one plain English request.

Setting it Up with Claude

Go to Claude Connectors and look up the directory of available connectors. Search for Blotato. When prompted provide this URL: https://mcp.blotato.com/mcp
That’s it. Your connected.

Chat with Claude and Post via Blotato

I drop a launch post into Claude and say: "Publish this to my LinkedIn, X, Instagram, and Pinterest. Adapt the wording for each and attach the product image."

Claude checks which accounts are connected, reshapes the copy per platform, attaches the visual, and posts or schedules each one. Then I can keep talking: "Push the Instagram version to tomorrow at 9am and make the LinkedIn one warmer." It just adjusts.

Twenty minutes of tab juggling became one sentence.

The content is the hard part, and you already do that. Distribution is the boring part, and that is exactly what to hand off.

Know anyone who needs an AI-first agency to build cool, creative videos for E-commerce? Forward them this newsletter and let them connect with us at support@ppcninja.com.

NERD BYTES (FOR NERDS)

The $20 Plan That Stretches Like $100 🤓

Quick question. Does the $20 Claude plan feel too small, but the $100 plan feel like way too much?

You are not alone. A lot of us jumped to the $100 plan and never actually burn the full $100 each month, and that unused chunk just evaporates when the cycle resets. There is a happy middle, and I just moved my whole team to it (myself included, and I am a heavy user).

Here is the trick.

You drop down to the $20 Pro plan, which still gives you Claude Code in your terminal. Then you open your Usage settings, turn on extra usage, and click Adjust limit to set a cap, say $40.

Now you pay the $20 base, and anything past your included limit is billed only as you use it, at standard rates, and never a penny over your cap.

🔶 You pay for what you actually use above the base, instead of pre buying $100 and wasting the rest.
🔶 The cap is a hard ceiling. Set it to $40 and it will not run away on you.
🔶 Nudge, do not leap. If $40 runs short one month, add another $20 and see how it goes. The jump straight to $100 is usually overkill.
🔶 It is per person. Standardize it across a team and the savings stack up fast.

Who is this for?

🔶 Light or occasional users: you almost certainly never needed the $100 plan. 🔶 Power users: set a comfortable cap and let overage cover your busy weeks only.
🔶 Teams: this is exactly the move I just rolled out to mine. Same capability, far less waste.

One heads up. Starting June 15, 2026, Anthropic splits plan billing into two buckets: one for hands on work (chat and Claude Code in your terminal) and a separate pool for automated and SDK driven jobs. If you run scheduled or autonomous tasks, keep an eye on your usage page so your cap still covers what you need.

That is the quiet advantage. Same Claude, same power. You just stop paying for credits that were quietly going to waste.

COOL TOOLS

Wait, Claude Can Animate?

Not stock clips. Not a template you fill in. Actual custom animations, made from scratch, that explain your exact idea. And you do not touch a design tool to get one.

Let me show you what I mean. This complex gif below was built by simply describing what I wanted to Claude.

How It Actually Works

Here is the part that surprised me. Claude does not reach for a video editor. It writes the animation as a tiny web page (the same HTML and CSS that builds any site), and then turns that page into a GIF you can drop anywhere. You never see the machinery. You just describe the thing and a finished animation pops out.

And because it is built from a simple recipe, changes are easy. "Make the arrows orange." "Slow it down." "Add a fourth panel." It just redoes it.

Bonus: it also handed me a still image version of each one. So I have the animated GIF for the places that move, and a clean static picture for the places that do not.

NEWS WORTH FOLLOWING

🔶 Amazon drops the Rufus name for "Alexa for Shopping" Amazon folded Rufus into Alexa and added Buy for Me and Scheduled Actions, so the assistant now researches, compares, and checks out for shoppers. The buy box is quietly becoming a conversation.

🔶 Amazon's Seller Assistant just went agentic The Seller Central helper can now reason, plan, and take actions for you, and a new dynamic canvas builds dashboards from a prompt. Sellers reportedly accept its suggestions around 90 percent of the time.

🔶 Amazon's "Enhance My Listing" puts AI on your detail pages New tools draft listings, A plus content, and lifestyle imagery with far less manual effort. Handy, as long as you still edit before you publish.

🔶 Claude Opus 4.8 arrives with Dynamic Workflows The new flagship can orchestrate up to 1,000 subagents at once, and its fast mode runs roughly 3x cheaper than before. Bigger automations for less money.

🔶 Anthropic launches Claude for Small Business A version of Claude aimed squarely at small teams, which is most of us. Worth a look if you have wanted Claude wired into how your business actually runs.

🔶 Amazon's Alexa rebrand hints at its advertising playbook As shopping goes agentic, the ad surface shifts with it. If you run Amazon ads, this is an early map of where spend is heading.

EVENTS

This past Sunday was a wild one. Nine speakers, all going deep on AI, and I was lucky enough to be one of them. The energy in that room was something else.

Missed it? The recordings can still be purchased from Kevin King and Norm Farrar, and they are absolutely worth it. Hit reply and I will send you the details.

🗞️ Nano Banana for Video is here💥🚀

Ritu Java — Fri, 29 May 2026 02:11:12 +0000

If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
🗞️ The Nano Banana for Video Is Here💥🚀
Would You Let AI Run Your Ad Account Live? 🛠️
Alexa now enables AutoBuy
News Worth Reading
AI Related Events worth Tracking
SellerLingo Launch

The Nano Banana for Video Is Here 💥🚀

Remember the Nano Banana moment?

Last August it was a fun toy. By November it was Pro. By February it was the default behind almost every AI product image you saw on Amazon. The whole shift happened in six months, and the sellers who started early were generating clean A+ hero shots while everyone else was still asking "but is it any good yet."

Last week, the same thing happened for video.

Google announced Gemini Omni at I/O on May 19. The first model, Gemini Omni Flash, is already rolling out. It accepts any input (image, audio, video, text) and generates video back. Watch the launch reel here: Introducing Gemini Omni.

But the announcement is not the story. The story is what it does.

Why This One Is Different

Every AI video tool before this had the same fatal flaw for product content. You'd render a clip, tweak the prompt, render again, and your product would morph. The bottle was teal, then blue, then a slightly different bottle. The hero became a stranger between edits. You could not finish a video without watching the product drift off model.

Omni edits like a conversation. Each instruction builds on the last.

Google's signature demo is a violinist. Across three rounds of edits, the face, the posture, and the way she holds the instrument all stay locked. Change the environment, swap the lighting, push in for a tighter angle. The person stays the same person.

Now apply that to a product. Your hero SKU on a wood floor. "Now put it on a kitchen counter." "Now show steam rising." "Now render the whole thing as cinematic realism, then as watercolor, then as claymation." Same product. Same shape. Same label. The iteration loop finally stops fighting you.

The deeper move under the hood: you can swap a product, prop, outfit, or reference image without rebuilding the whole video. The original camera path, the contact shadows, and the surrounding scene all stay in place. That is the unlock. You change one thing at a time and the rest holds.

What This Means for Sellers

The story of AI product images was never "they got better." It was "the iteration loop got fast enough that you could actually finish one." That is exactly what just happened to video.

A Sponsored Brands video, a short social ad, a 6-second clip for your A+ content. Until now those required either a real shoot or a brave gamble on a fresh render every time. Omni reduces all of that to a chat window where your product holds together.

There are honest caveats. The full tool runs on the paid Google AI Plus, Pro, and Ultra tiers in the Gemini app and Google Flow. The free door is YouTube Shorts and the YouTube Create app. The developer API is "coming weeks." Every output is watermarked with SynthID. And the model is days old, so verify the physics on your category (liquids pouring, fabrics draping) before you trust it for paid placements.

The one thing I would do this week: watch the launch reel and look at the violinist demo with seller eyes. If your product can hold across three edits the way her face does, your video cost just dropped to a chat window.

That is the Nano Banana moment for video. It will not feel huge today. It will feel obvious in three months.

Know anyone who needs an AI-first agency to build cool, creative videos for E-commerce? Forward them this newsletter and let them connect with us at support@ppcninja.com.

NERD BYTES (FOR NERDS)

Hot Take: Would You Let AI Run Your Ad Account Live? 🤓

Time for the controversial one. Amazon shipped the Ads MCP Server (open beta since February). Connect Claude, ChatGPT, or Gemini straight to your live ad account, and just talk to it. "Pause these. Raise bids on those. Build me a new Sponsored Products campaign." It turns your sentence into API calls and does it. On the spot. 50+ tools across SP, SB, SD, DSP, and AMC.

It is genuinely impressive. It is also exactly where I get nervous.

Here is the part nobody likes to say out loud. When you chat with the MCP, the change is live the second it acts. No file to look at first. No diff. No record of what changed or why. If it misread you, the mistake is already in your account. Put two people or two sessions on the same account and you have chaos with no paper trail.

So here is what my team does instead. We connect Claude to BigQuery through the CLI, pull the numbers, and have Claude generate a bulk file based on a chat. A human reviews the file. Then we upload it. Nothing touches the account until we say so.

Two philosophies. Pick a side.

The Ads MCP (chat and it does it)
🔶 Pro: Fast. Conversational. One sentence can build a whole campaign.
🔶 Pro: Official, 50+ tools, broad coverage across every ad type.
🔶 Con: Changes go live as you chat. There is no artifact to review first.
🔶 Con: No clean audit trail. Hard to diff, hard to roll back, easy to lose the thread across sessions.

The bulk file (data, then a file, then you upload)
🔶 Pro: Complete traceability. The file is a saved, reviewable, diffable record of every change.
🔶 Pro: Nothing goes live until a human approves it. The guardrail is built in.
🔶 Con: More steps. Slower. Needs the BigQuery and CLI plumbing in place.
🔶 Con: Not the instant magic of just talking to your account.

I am firmly in the bulk file camp. The file is the guardrail. But I know sharp operators who love the speed of the MCP and happily accept the risk.

So tell me where you land. Hit reply with "chat" if you would let the MCP run live, or "bulk" if you want the file and the paper trail. I will share the split in a future issue.

COOL TOOLS

Alexa now enables AutoBuy

A new Alexa related feature showed up. It is called Auto Buy. A shopper taps "Add to Auto Buy" on your product page, sets a max price ("$27.99 or less, the 90 day low"), a start date, and an expiry. Then they leave. When your listing hits that number, Amazon executes the purchase on their behalf. No return visit. No second decision.

Vanessa Hung was one of the people that noticed this and shared this screenshot.

Credit: Vanessa Hung

The button is rolling out inside the new Alexa for Shopping experience, right next to the "Ask Alexa" prompt chips. If you haven't seen it on your own ASINs yet, go look. Some categories already have it. Others are days away.

The reason it matters: every Auto Buy rule is essentially a conditional purchase order sitting in a queue. A loyal repurchase customer can set the rule once and stop coming back to comparison shop. That has structural consequences for how you read your own search and demand data (see Voices below for two sharp takes).

For now, the only action item for you: open one of your hero ASINs and check whether the button is on the page. Knowing it is there changes how you read the rest of this issue.

NEWS WORTH FOLLOWING

Andrew Bell on the new Alexa search stack. He took apart 659 result sets and 12,000 product cards across 48 categories and concluded that "fit beats popularity" in Position 1: feature and title overlap drives 94% of top placements, not star rating alone. He also argues that Alexa generates more searches than before, not fewer, because one human request now fans out into many machine generated queries underneath. Read part one of his blueprint.

Read together, the two pieces frame the same question from opposite sides. Demand is queueing, and searches are multiplying. Which means search volume is no longer a clean read on demand. Worth thinking about before you make confident calls on what your data is telling you this quarter.

🔶 Anthropic Released Claude Opus 4.8. Sharper judgment, more honesty about its own work, and a fast mode roughly 2.5 times quicker than 4.7. Pricing stays the same.

🔶 Dynamic Workflows Lands in Research Preview. Alongside Opus 4.8, Anthropic shipped a new agentic surface for Claude Code where one task can spawn up to 1,000 parallel subagents.

🔶 Anthropic Teases "Mythos-Class" Models. Buried in the Opus 4.8 announcement: a new "Mythos" family of models is already in development. No specs, no date, but the name is on the record now.

🔶 Amazon Licenses Its AI Shopping Engine to Other Retailers. AWS launched the Agentic Shopping Assistant with Kate Spade as first customer. Same tech that drove roughly $12B of incremental Amazon sales last year is now available to competitors.

🔶 Google Ships Antigravity 2.0 at I/O 2026. Agent-first developer platform with a desktop app, CLI, SDK, and a new $100/month AI Ultra tier. Gemini 3.5 Flash is the new default model and runs roughly 4 times faster than other frontier models.

🔶 Amazon Brings Natural Language Campaign Setup to the 2026 Upfront. Full-Funnel Campaigns let advertisers spin up cross-surface campaigns (sponsored ads, streaming TV, display, audio) from a single prompt, with AI-assisted creative and audience strategy included.

EVENTS

May 27, 2026, Seoul
My session: Getting Ready for Conversational AI: How Rufus Is Reshaping Product Discovery. Click to read the takeaways!

15×5 Hacks Webinar, 1200+ Registrations!

Wow! So honored to be voted #4 on this hack series!

According to Nick Penev this one had the best hacks so far with 1200+ peeps who signed up for it ... 17 Amazon speakers shared 17 hacks in a short 5 minute format in 90 minutes competed for eCom Hot Sauce prize!

Vincenzo Toscano - How to lower your ACOS by over 50% with AMC!
https://lnkd.in/dukwmfJy
Ben Leonard - How to Jumpstart TikTok Shop Product Launches to Attract Customers and Affiliates!
https://lnkd.in/dMmzMw_6
John Aspinall - The Codex Playbook for High-CTR Product Images!
https://lnkd.in/d8rieXRf
Ritu Java - Turn Google Sheets into a powerful LLM interface for Scale!
https://lnkd.in/d8fYyA85
Matt Kostan - Prompt Injections That Drive Sales!
https://lnkd.in/dwwAewzK

You can watch the full recording here: https://lnkd.in/dgJC9rXE

Did you miss the free webinar with Kevin and Norm on May 28th? Here’s the replay and don’t miss this event on May 31st!!

🗞️ One Action. Half the Tokens💥🚀

Ritu Java — Thu, 14 May 2026 02:44:09 +0000

By the way, if this newsletter is turning into what sounds like a programmer's manual, my recommendation to you is to soak it in anyway. Trust me, you will need this stuff in the agentic era!

In this edition (shortcuts):
🗞️ One Action. Half the Tokens💥🚀
CLI vs MCP, the 60-Second Decision 🛠️
Seedance 2.0 for Video Production
News Worth Reading
Upcoming Events

One Action. Half the Tokens💥🚀

Who here is frustrated because you frequently run out of tokens? Well, lucky you because today I am going to show you a game-changer!

If you are on Claude, open the Mode picker in Claude Code right now. (here’s a screenshot below if you have not seen it.)

A menu pops up. Five modes. Most sellers I talk to have never opened this menu. They use the default mode that came with the install and they burn tokens like the bill is free.

It is not free. And one of those five modes is sitting there waiting to save you half your spend on every complex task.

That mode is Plan Mode. And if you pair it with one extra step, you have a workflow that quietly beats the people running Claude on autopilot.

This is the workflow my podcast co-host Danny McMillan has been refining for months. He calls the second piece the push back step. We walked through the whole thing on the latest Sellers Sessions episode: Building Repeatables in Claude Skills, CLI vs MCP, and Token Discipline. Worth a listen on its own. But here is the short, hands-on version.

Quick note before we dive in. The examples below use Claude Code because that is where I run most of my work. The technique itself is universal. Cursor has it as Plan mode. Windsurf has it as Cascade plan. Any AI coding tool with a plan or agent mode plays by the same rules. Only the labels change.

The 5 Modes in 10 Seconds

Every Claude Code session runs in one mode. You cycle through them from the Mode picker (screenshot above). The exact shortcut depends on your platform and your keyboard settings, so the picker itself is the safest entry point. Here is what each one does.

🔶 Ask permissions. Claude asks before every action. Slowest. Safest. Use when you are new to a codebase or about to do something destructive.

🔶 Accept edits. Claude makes edits automatically but asks before running commands. The default for most users. Fine for routine work.

🔶 Plan mode. Claude reads, thinks, and proposes a plan. Does not edit. Does not execute. Just plans. This is the one we are here for.

🔶 Auto mode. Claude runs without permission prompts until something fails. Fast. Use only after you trust the plan.

🔶 Bypass permissions. Claude does everything without asking. Use only when you have verified the plan and you are inside a worktree or sandbox.

Most sellers I see live in Accept Edits or Bypass. They skip Plan Mode entirely. That is the leak.

Why Plan Mode Is Practically Free

Plan Mode does not edit your files. It does not run shell commands. It reads what it needs and writes a structured plan back to you.

The tokens it spends are tiny compared to a full run. Reading is cheap. Planning is cheap. The expensive parts of a Claude session are the exploration loops, the wrong edits, the rollbacks, and the regenerations that come after each mistake.

In Plan Mode, none of that happens. You get the model's best read of the task before any tokens get burned on doing it wrong.

So the question is not should I use Plan Mode. The question is what do I do with the plan once I have it.

The default answer is "approve and execute." That is fine. But it leaves the biggest savings on the table. Because the plan you got on first pass is rarely the plan you want.

The Push Back Step

Danny's contribution is the discipline of what to do between getting the plan and executing it.

Instead of clicking approve, you push back. You ask Claude to defend the plan. Question its assumptions. Find the cheap path. Show its work.

These are the four prompts I now run on every non-trivial Plan Mode output:

🔶 "Walk me through the most expensive step in this plan and tell me why it has to be that expensive."

🔶 "What assumption are you making here that, if wrong, breaks the whole plan?"

🔶 "Is there a cheaper way to get the same outcome?"

🔶 "Which step could I skip without breaking the result?"

Run them in order. Read the answers carefully. Half the time Claude will rewrite the plan on its own after the second prompt. The other half it will defend a step that you genuinely needed to know was that expensive.

Why This Actually Works

A first-pass plan optimizes for getting the task done. A pushed-back plan optimizes for getting the task done with the fewest tokens.

Those are different objectives. The model will not switch from one to the other unless you tell it to.

When you run the push back prompts, you are forcing a reframe. The model has to argue with itself. Steps that survive the argument are the ones that actually mattered. Steps that get dropped were padding the model added because nobody told it to be cheap.

That is the quiet advantage. The pros are not using a different model. They are running the model with one more conversational loop before they let it touch anything.

A Real Example

Last week I asked Claude to mine 90 days of search term reports for one of our accounts and propose a refreshed negative keyword list.

The first-pass plan had 11 steps. Read every search term file, build a per-campaign negative list, run a final dedupe pass, then format for Amazon bulk upload. That plan would have worked. It would also have spent half an hour re-exploring patterns I had already documented in the account brief.

So I pushed back.

"What assumption are you making here that, if wrong, breaks the whole plan?"

Claude flagged that it was assuming each campaign needed its own threshold for what counted as "underperforming." I told it the threshold was uniform across the account. The revised plan dropped from 11 steps to 6.

"Walk me through the most expensive step."

Claude pointed to the read loop, which was opening every search term file twice (once for filtering, once for the negative-list write). I asked for a single pass with a buffered write. Down to 5 steps.

"Is there a cheaper way to get the same outcome?"

Claude suggested running the three campaigns as parallel sub-agents instead of sequentially. Same total token spend, a third of the wall clock time.

That final plan was the one I executed. It is not the plan I would have run without the push back loop.

When to Skip Plan Mode

Honest caveat. Plan Mode is overkill for small tasks.

🔶 One-file edit. Just do it.
🔶 A quick grep or lookup. Just do it.
🔶 A task you have done five times before. You already know the plan.

But the moment a task touches three or more files, or you would struggle to write the plan yourself in two minutes, Plan Mode pays for itself. The push back loop pays for itself again on top.

The rule I use: if the task feels complicated enough that I want to think before I start, Claude probably does too.

What This Looks Like for Amazon Sellers

The technique transfers cleanly out of pure code work. Every one of these is a Plan-Mode-plus-push-back candidate.

🔶 A bulk campaign restructure. Plan first. Push back on which campaigns to touch, in what order, with what rollback if performance drops.

🔶 A multi-account performance audit. Plan first. Push back on which signals are actually predictive vs which are noise.

🔶 An ad copy refresh across 50 ASINs. Plan first. Push back on whether you really need new copy on all 50 or whether the bottom 10 by CTR carry the whole gain.

🔶 "Build me a Claude skill that does X." Plan Mode is the entire conversation. The skill is the artifact of the conversation. The push back loop is how the skill ends up tight instead of bloated.

The pattern is always the same. Plan first. Push back. Then run.

The Quiet Advantage

Plan Mode has been sitting in Claude Code the whole time. Cycle through it once and you understand why most users skip it. It looks like a slower version of normal.

It is not slower. It is the version that does not waste your tokens on the wrong path.

The sellers who learn the Mode picker and the push back loop are going to look like they are running tighter, cheaper sessions than the rest of us for the next year. They are not on a different plan. They are just on a different mode.

Have you been running Plan Mode already? What is the push back prompt that has saved you the most? Hit reply and let me know. If you want the four prompts as a copy-paste block plus the Mode picker how-to, hit reply with "plan mode" and I will send it over.

Know someone who is still running Claude Code on Auto and complaining about the bill? Forward this to them and earn Karma points.

PPC Ninja is helping brands future proof their listings for AI, helping you build RUFUS enabled, stunning images and videos with AI. Hit reply on this to chat with us. Explore how we can scale your content production across Social media, Amazon ads, Amazon Posts efficiently and affordably.

NERD BYTES (FOR NERDS)

CLI vs MCP, the 60-Second Decision 🛠️

Every week somebody asks me whether they should install a new tool as a CLI command or as an MCP server. Both work. They are for different things. Here is the rule that takes 60 seconds to apply.

🔶 Use the CLI route when the tool has a local binary you can run from a terminal. Examples: gws for Google Workspace, gh for GitHub, bq for BigQuery. Fastest setup. Lowest moving parts. Survives an MCP outage.

🔶 Use the MCP route when the tool only lives in the cloud, has no local binary, or you want one integration usable across multiple AI clients (Claude Desktop, Claude Code, Cursor, etc.). Build it once. All your AI surfaces talk to it.

🔶 Use both for the same tool when: never. Pick one and document the choice in your project notes. Installing both is the fastest way to get your AI to call the wrong one when you are not looking.

For sellers specifically:

🔶 Pulling BigQuery data, sending email, reading sheets, listing GitHub PRs → CLI route. These tools already have rock-solid command line versions.
🔶 Figma reads, Canva pulls, Notion writes, Linear tickets → MCP route. These are cloud-native services where MCP servers shine.
🔶 Anything specific to a single niche workflow that you only use in one AI client → just a CLI is fine.

The quiet advantage. Pros pick once and move on. Amateurs install both, run into the duplicate-command-path bug, and lose half a day debugging it. Pick the rule. Apply it for 60 seconds. Move on.

We unpacked this in more detail on the same Danny episode Building Repeatables in Claude Skills, CLI vs MCP, and Token Discipline. Worth a listen if you are about to install your next batch of tools.

COOL TOOLS (FOR THE COOL KIDS)

Seedance 2.0, The Video Tool We Cannot Stop Using 🎬

The PPC Ninja team has been busy creating videos for our clients. And we cannot stop using Seedance 2.0. It is a game changer.

If you want the deep walkthrough that pulled it all together for us, watch this one: Seedance 2.0 is CRAZY for AI Filmmaking, Full Course. Twenty minutes. Worth every one of them. Below is the short version with the techniques that translate cleanly to Amazon product video.

Why Seedance 2.0 Is Different

Most video models give you one or two reference images and a prompt. Seedance 2.0 takes 9 images, 3 video clips, and 3 audio clips in a single generation. That is fifteen reference assets shaping one clip.

For sellers, that is the whole game. You can stack the product hero, a lifestyle scene, brand color and lighting plates, a motion reference, and your recorded voiceover. All into one prompt. The model resolves them into a coherent clip with native stereo audio baked in.

Under the hood, ByteDance's Temporal Anchor tech is the anti-morphing layer. It keeps your product shape stable across the clip. Most AI product videos drift halfway through. This one does not.

Know anyone who needs an AI-first agency to build cool, creative videos for E-commerce? Forward them this newsletter and let them connect with us at support@ppcninja.com.

NEWS WORTH FOLLOWING

🔶 RUFUS is now Alexa for shopping!
https://www.linkedin.com/feed/update/urn:li:activity:7460329139674198016/

🔶 Anthropic New Monthly Limits: Starting June 15, paid Claude plans will transition to a "metered" agentic era, which includes specific monthly credits for programmatic usage across tools like the Claude Agent SDK and Claude Code GitHub Actions.

🔶 Codex 5.5 is Now Better Than Ever (I am using both Opus 4.7 and Codex 5.5 equally for coding projects)

UPCOMING EVENTS

May 27, 2026, Seoul
I am honored and excited to take the stage once again at the Amazon Ads Prime Day Summit in Seoul. I will be sharing the stage with Bradley Sutton, Bopyo Park, and Ha Eun-young. My session: Getting Ready for Conversational AI: How Rufus Is Reshaping Product Discovery.

🗞️ Claude Emailed my Team💥🚀

Ritu Java — Wed, 29 Apr 2026 14:23:27 +0000

In this edition (shortcuts):
The Email I Didn’t Send
Tell Claude You're Going to Sleep 🥱
GPT-5.5 vs Opus 4.7 (The April Showdown) 🥊
News Worth Reading
Upcoming Events

🗞️ The Email I Didn't Send💥🚀

I sent an email to my team this morning. Eight people, with me Cc'd.

I didn't write it. I didn't paste it into Gmail. I wasn't even at my keyboard for the actual send.

I asked Claude to do it. And Claude did it.

This is not another "Claude drafted my email" story. We've all done that. Claude has been writing my drafts for two years now. This is something different. I told Claude what I wanted, Claude wrote it, showed me a preview, and then actually sent it from my Gmail to eight of my teammates. Subject line, formatting, recipients, Cc, signature. The whole thing.

That sounds small. It's not.

What Actually Happened

I'm rolling out a new tool to my PPC Ninja team. Nothing dramatic, just a one-time setup they all need to run on their laptops. Five minutes, one paste of a command, and they're done.

Normally I would write that email myself. Sit down, draft it, format the Mac and Windows instructions separately, double check the recipient list, send, then field the followups when someone got stuck on step three.

This time I just told my AI assistant Claude:

"Send an email to my eight teammates with separate instructions for Mac and Windows, as non-technical as possible. Cc me. Say that you are my assistant Claude."

That was the whole instruction.

Claude drafted the email. Showed me a preview in my browser. I asked for two changes (give the Mac and Windows sections colored banners, and account for the fact that some teammates have the folder named slightly differently). Claude updated the preview. I said "send."

The email landed in eight inboxes within a second. With me Cc'd, so I have a clean record on my side.

[IMAGE: screenshot of the rendered email preview with a blue MAC banner and an orange WINDOWS banner, recipient block visible at top]

So What's Actually New Here

You've been able to ask AI to write emails for years. Old news.

What's new is the send part. The "AI did the thing in the world" part.

Until this week, every AI email workflow I'd used had a human-shaped gap right at the end. The AI writes the draft. You copy it into Gmail. You add the recipients. You fix the subject line. You hit send. The AI is the writer, you are the operator.

That gap closed for me today. Claude is now the writer and the operator. I'm just the editor and the "yes, ship it" button.

The thing that made it work isn't a fancy new model. It's a small, free, boring CLI tool from Google called gws (Google Workspace CLI). It runs on my Mac. Claude knows how to use it. When I say "send," Claude calls the gws send command with the right arguments, gws calls the Gmail API, and Gmail does what Gmail does.

No Zapier. No make.com. No monthly subscription stack. Just my AI, my CLI, and my Google account.

But Wait, Don't I Want a Human in the Loop?

Yes. And I had one. Me.

The thing that makes this safe (and not "AI gone rogue spamming my contacts") is the workflow. Every write action goes through three checkpoints before it actually happens:

🔶 The dry run. Before sending anything, Claude calls the API in "dry run" mode, which validates the request without firing it. If the recipients are wrong or the formatting breaks, this is where it fails, harmlessly.

🔶 The preview. Claude renders the actual email in a browser tab so I can see it the way my recipients will see it. Subject, banners, code blocks, spacing. I'm not reading raw markdown, I'm looking at the finished thing.

🔶 The "ship it" gate. Nothing sends until I literally type "send." That's the only word that matters. Until I say it, the draft just sits in a temp file.

That third one is the most important. The AI is fast and confident. I still own the decision to put something in someone else's inbox. The skill that runs Claude's workspace tools has a hard rule baked in: no writes, no sends, no calendar invites without explicit confirmation. Claude refuses to skip that step, even when I'm in a hurry.

[IMAGE: a flow diagram showing the three checkpoints, draft → preview → send with a person icon at the send gate]

That's what "fast with guardrails" looks like in practice. The AI moves at AI speed. The human stays in charge of consequences.

Why This Matters For Sellers

Forget the email for a second. Picture the next twelve months of your business.

🔶 "Email the agency, ask them about Q3 budget reallocation" 🔶 "Tell the warehouse the cutoff date moved to Friday" 🔶 "Send our top 20 ASIN list to the new hire and Cc Mike" 🔶 "Reply to the photographer and ask if she can reshoot the lifestyle set this week" 🔶 "Reply to that buyer who asked about MOQ. Polite but firm. Cc legal."

Every single one of those is something you do today by stopping whatever you were doing, opening Gmail, fishing the right thread, writing the message, and sending. Five to ten minutes each. Twenty of them in a day. That's two to three hours of your week being a Gmail operator.

Now imagine just saying the request out loud (or typing it into a chat) to your AI, and the email is sent. Not drafted. Sent. With you Cc'd or Bcc'd if you want a record. With your tone, your judgment, and your audit trail attached.

That's where this is going. And it's not a 2027 thing. It works today.

The Bigger Shift

There's a phrase I keep coming back to: AI is moving from a thing you talk to into a thing that acts on your behalf.

For two years we've been chatting with AI in browser tabs. Asking it questions, getting answers, copying and pasting back into our actual work. The AI was clever, but trapped behind a chat box.

What's happening now is the chat box is leaking. AI is starting to reach into your real tools. Your Gmail. Your Drive. Your calendar. Your bulk file uploader. Your ad account.

Email is just the simplest, oldest, most universal version of this. Every business runs on email. Every team coordinates through email. Every customer reaches you through email. If your AI can send email as you, with your judgment behind it and your audit trail attached, you've quietly upgraded from "I have an AI assistant" to "I have an AI employee."

The model didn't get smarter this month. The wrapper around the model did.

What You Can Do This Week

You don't need a Claude Code setup or a CLI tool to start moving toward this. Three things you can do today:

🔶 Pick one repetitive email task. The kind you do five times a week without thinking. Status updates, supplier check-ins, agency followups, "did you see this" forwards. That's your candidate.

🔶 Write the request like a delegation, not a prompt. "Email Anna and ask if the August invoice is being processed, polite, Cc me, sign off as Ritu" reads differently than "Write me an email about an invoice." The first one is how you'd talk to a real assistant. Get used to that voice.

🔶 Decide your "ship it" gate. Even when you have AI that can send, you still want a moment of human judgment before each send. Decide what that looks like for you. Mine is a one line preview I can scan in three seconds. Yours might be different. The point is to pick it on purpose, not by accident.

The rest is plumbing. The plumbing exists today, and it will be in front of you, dressed up nicely inside the next round of AI products that ship over the next quarter.

The sellers who are practiced at delegating to AI by then will use those tools fluently. The ones who aren't will still be copy-pasting drafts into Gmail.

Cool, right?

Want to try it? Hit reply.

If you want the actual install commands and the small Claude skill that makes this work on your own laptop (Mac or Windows), hit reply with "try it" and I'll send the setup over. Takes about five minutes from start to finish, no developer experience needed.

Have you handed off a real action to your AI yet? Not a draft. An actual send, or save, or post. Hit reply and tell me what it was. I'm collecting examples.

Know a seller who's still in the "AI writes my drafts" stage? Forward this to them.

~Ritu

NERD BYTES

Tell Claude You're Going to Sleep 🥱

Has Claude ever stopped you mid-flow with seven follow-up questions before doing the actual work?

"Should the heading be H2 or H3?" "Comma or semicolon here?" "Should I send the email now or wait until you confirm?"

That's because Claude defaults to a collaborative mode. It checks in. It asks permission. Which is great when you're sitting at the keyboard answering in real time. Less great when you handed it a 30-minute task and went to grab coffee.

Here's the trick I've started using.

The Magic Words

When you're about to step away (or just don't want to be pinged), tell Claude this:

Or, more dramatic:

Then walk away.

What Actually Happens

Claude shifts modes. It stops being a polite assistant asking permission. It becomes an orchestrator, breaking the work into chunks and dispatching subagents to handle them in parallel. The subagents do their pieces, hand back results, and Claude assembles the final output.

The results are honestly a little wild. I came back from a coffee break to find:

🔶 Three open questions answered with reasonable assumptions, each clearly labeled
🔶 Five subtasks dispatched and completed in parallel
🔶 A summary of every decision made, with the reasoning
🔶 A short list of "wasn't sure, please confirm" items at the bottom for me to scan

Way better than coming back to find Claude waited four hours on question number one.

Pro Tip

This works best for tasks with inherently independent sub-pieces. Audits across multiple ASINs. Bulk file generation across many campaigns. Research across a list of competitors. Anything you'd describe as "do the same thing, just for 10 of them."

It works less well when each step depends on the exact answer from the step before. For tightly sequential work, you really do need to stay in the loop.

This is the "delegate, don't supervise" pattern. The tools have been there for a while. Most people just don't tell Claude it's allowed to use them.

COOL TOOLS

GPT-5.5 vs Opus 4.7 (The April Showdown) 🥊

Two Heavyweight Releases In Eight Days

April was wild. Anthropic shipped Claude Opus 4.7 on the 16th. OpenAI fired back with GPT-5.5 (the model behind the new Codex) on the 23rd. Eight days, two of the strongest coding models the industry has ever seen, both raved about by developers I trust.

The "which one is better" question is genuinely interesting this time. Because the answer is: it depends on what you're doing.

Here's how I'm thinking about it, and how to access both inside the IDE you already use.

The Scoreboard

I went through the actual benchmarks instead of the hot takes. Here's the honest comparison.

	Claude Opus 4.7	GPT-5.5 (Codex)
Released	April 16, 2026	April 23, 2026
SWE-bench Verified	87.6%	competitive
SWE-bench Pro (real GitHub work)	64.3% (industry highest)	~58%
Terminal-Bench 2.0 (drive a real terminal)	69.4%	82.7% (state of the art)
Output token efficiency	baseline	72% fewer tokens for the same answer
Context window	1M (via API)	400K (IDE), 1M (API)
Pricing per 1M tokens	$5 in / $25 out	$5 in / $30 out
Vision	3.75MP (jumped from 1.15MP)	strong, no recent jump

Neither model "wins." They optimize for different things, and that's actually useful.

Where Each One Pulls Ahead

After a week of poking at both, here's my read.

🔶 Opus 4.7 is better when the task is "understand this whole thing first, then change it carefully." Big codebases. Refactors. "Why does this break under that condition." Anything where the AI has to hold a lot in its head before touching anything. The 1M context window plus the highest SWE-bench Pro score in the industry is not a coincidence.

🔶 GPT-5.5 is better when the task is "go drive the terminal and finish this for me." Set up a project. Run the test suite, fix what fails, run it again. Generate a bunch of files in parallel. Anything agentic and "go do the thing." The Terminal-Bench score (82.7%) tells you the truth here. It's the best terminal driver currently shipping.

🔶 GPT-5.5 is also significantly cheaper in practice, despite a slightly higher per-token output price. The 72% reduction in output tokens means an average task costs less, even before you count the speed gain.

🔶 Opus 4.7 has a real lead on visual reasoning. The vision resolution jumped from 1.15MP to 3.75MP this release. If your work involves screenshots, charts, or mockups (looking at you, Amazon Creative Studio reviewers), this matters more than benchmark deltas.

How To Use Both In Cursor

The good news: you don't have to pick one and stick with it. Cursor supports both, and switching takes one click.

For Opus 4.7:

Open the chat panel in Cursor
Click the model dropdown at the bottom
Select Claude Opus 4.7 (it's at the top of the list right now)
That's it. Your existing Cursor Pro plan covers usage, or you can bring your own Anthropic API key

For GPT-5.5 (via the Codex extension):

Open Cursor's Extensions panel
Search for "Codex (OpenAI)" and install
Sign in with your ChatGPT account (Plus, Pro, Business, Enterprise, or Go all work) or paste an OpenAI API key
Drag the Codex icon to your right sidebar so it sits next to your Cursor chat
Pick a mode: Chat, Agent, or Agent (Full Access)
Pick a reasoning effort: low, medium, or high

[IMAGE: Cursor sidebar showing both the native chat with Opus 4.7 selected, and the Codex extension panel side by side]

If you already pay for ChatGPT, your existing plan covers Codex. No second subscription needed. Same goes for Cursor Pro and Opus.

Pro Tip: Bind both to keyboard shortcuts so you can flip between them without breaking flow. I run Cmd+L for Cursor's native chat (Opus 4.7) and Cmd+Shift+K for Codex (GPT-5.5). Different jobs, different tools, one keypress apart.

My Workflow This Week

I've started using them in tag team. Plan and architect with Opus. Execute and orchestrate with GPT-5.5.

Concretely: when I'm starting a new internal tool, I ask Opus 4.7 to read the existing codebase, understand the patterns, and write the implementation plan. Then I hand the plan to Codex in Agent mode and let it actually build the thing while I'm in another meeting (see this issue's Nerd Byte for the "go do this while I'm asleep" pattern).

Opus is the architect. Codex is the foreman. Both feel less anxious than the previous generation. Both ask fewer permission questions before they get to the answer.

Should You Switch?

If you're already vibe coding and you're paying for both ChatGPT and Cursor (which I am), you don't have to choose. Use Opus 4.7 for thinking, GPT-5.5 for doing.

If you can only afford one, here's the honest split:

🔶 Lots of small, well-scoped tasks? GPT-5.5 wins on speed and cost.
🔶 Bigger refactors and "understand my whole codebase" work? Opus 4.7 still has the edge.

If you're not vibe coding yet at all, this isn't the news that should pull you in. Start with either one inside Cursor's free tier. The frontier models are now so close that the choice barely matters until you're shipping real volume.

My take: This is the first time in two years the gap between OpenAI and Anthropic at the top of the leaderboard genuinely feels like a rotation, not a winner. Both are excellent. Both are getting cheaper. Both are getting more agentic. Crazy times we live in.

I'll report back with real numbers after a full month of running them in parallel.

Are you Team Opus or Team Codex right now? Or running both like me? Hit reply and tell me what's working in your stack.

NEWS WORTH FOLLOWING

🔶 Google Commits Up to $40 Billion in Anthropic at a $350B Valuation

🔶 Amazon Q1 2026 Earnings: Ad Revenue Up 21% to $16.8B, $200B AI Capex Locked In

🔶 Amazon Launches AI-Powered Audio Q&A on Product Pages, Shoppers Can Now Ask Listings Questions

🔶 Amazon Brings Generative AI to Amazon Marketing Cloud (AMC) for Audience Building

🔶 Walmart Connect Bets Agentic AI Is the Next Battleground in Retail Media

🔶 Shopify's Storefront Catalog MCP Flips to Universal Commerce Protocol (May 30 Cutover)

UPCOMING EVENTS

Time to tune in to the Go with the Flow Podcast where Danny McMillan and I dive into AI topics. We will be talking about Claude Code, specifically some of the newer roll-outs and ways to improve Claude Skills. Listen here!

Join this free webinar by Helium10 today!
https://streamyard.com/watch/qUmtMXiMKvKt

🗞️ ChatGPT-Images-2.0 Thinks Before It Draws💥🚀

Ritu Java — Wed, 22 Apr 2026 14:14:10 +0000

🗞️ ChatGPT-Images-2.0 Thinks Before It Draws💥🚀

If you've been loyal to Nano Banana for the last six months, you're not alone. I have too.

But something shipped yesterday that changes the math.

On April 21, OpenAI released ChatGPT Images 2.0, powered by a new model called gpt-image-2. And for the first time, an image model does something none of the others do. It thinks before it draws.

That sounds like marketing copy. It isn't. It's a structural shift in how image generation works. And for those of us generating product images, A+ content, and ad carousels at scale, it changes the kinds of prompts we can realistically ship.

Check out the difference:

GPT-Image-2

Nano Banana 2

What Actually Shipped

Three things matter.

🔶 Native reasoning. The model plans multiple images from a single prompt, searches the web during generation (for product refs, brand styles, factual accuracy), and double-checks its own output before rendering. Thinking mode is paid-gated. Standard mode is free for all ChatGPT users.

🔶 2K resolution by default. 2,048 pixels on the long edge. Print-ready without upscaling.

🔶 Dense text rendering. This is the big one. Typography, labels, barcodes, data charts, and UI screenshots render with accuracy that Nano Banana Pro and Flux 2 can't currently match. The detail most of us have been hacking around for two years.

Available through ChatGPT, Codex, and the API as gpt-image-2.

What "Thinks Before It Draws" Actually Means

Most image models, including Nano Banana 2, generate in one shot. Prompt in, image out. Want three variations? Run it three times.

Images 2.0 plans first. If your prompt says "generate a 5-slide Instagram carousel with consistent faces," it asks itself which poses, which backgrounds, which continuity rules keep this consistent, then renders with that plan baked in.

The practical result: fewer "same prompt, different faces" disasters. Better multi-image continuity. Smarter object placement in complex scenes.

The difference between a model that follows instructions and a model that understands the brief.

The Test: Four E-Commerce Prompts, Two Models

Sellers don't need one beautiful image. We need a set that shares a visual language. Hero, lifestyle, carousel, data infographic, brand collab. All consistent, all on-spec, all on-brand.

I picked 4 prompts that cover the categories sellers actually use every week. Each one stress-tests a specific capability:

🔶 Prompt 1: Instagram carousel with character consistency (multi-image continuity)
🔶 Prompt 2: Lifestyle product shot (baseline photography realism)
🔶 Prompt 3: Data infographic (dense text plus data viz)
🔶 Prompt 4: Limited-edition brand collab mockup (dual-brand synthesis)

I'm running each on ChatGPT Images 2.0 (Thinking mode) and Nano Banana 2. Same prompt, same reference images where applicable. Then grading on brand coherence, text accuracy, multi-image consistency, Amazon-spec readiness, and honest usability.

Copy these into your own sessions and run the parallel test. Your results will teach you more than mine will.

Prompt 1: Instagram Carousel With Character Consistency

Tests multi-image continuity. The prompt that most exposes the gap between a one-shot model and a planning model.

Create a 5-slide Instagram carousel for an allergen-free bakery called
"Sweet Inclusion." Target audience: parents of kids with food allergies.

Slide 1: A young girl (age 7, light brown hair, blue eyes, wearing a
yellow sweater) looking sad at a birthday party dessert table, unable
to eat the cake.
Slide 2: Same girl, same sweater, now smiling brightly, eating one of
our allergen-free cupcakes.
Slide 3: Same girl and her mom (warm, 35ish, brunette) baking together
in a bright home kitchen.
Slide 4: Clean product shot of the cupcake range on light linen. Small
"Sweet Inclusion" wordmark.
Slide 5: Call to action slide: "Every kid deserves a yes. Order today."
Brand colors soft pink and cream.

Keep the girl and mom visually consistent across slides 1 through 3.
1:1 square. Warm natural light across all slides.

Here’s the comparison:

GPT-Images-2

Nano Banana 2

Prompt 2: Lifestyle Product Shot

Tests baseline photography realism and prompt nuance. The workhorse prompt every Amazon seller runs weekly.

Photograph: a wireless earbuds charging case (matte black, rounded
rectangle, about the size of a dental floss container) on a wooden
gym bench. Soft morning light from a window on the left, creating
gentle highlight on the case.

Supporting elements: a small hand towel folded nearby, faint
condensation droplets on the wood, a water bottle blurred in
background.

Shallow depth of field, 50mm equivalent, product sharp, background
soft bokeh. Natural color grading, slight warm temperature. No
branding text on the product.

Output: 4:5 vertical, suitable for Instagram feed.

Watch for: Light direction (does it actually come from the left?). Bokeh (natural or painted-on?). Scale (does the earbuds case look the right size or like a giant lunchbox?).

GPT-Images-2

Nano Banana 2

Prompt 3: Data Infographic

Tests text rendering plus data visualization. Most sellers still build these in Canva because image models butcher numbers and chart geometry. Images 2.0 is betting this is solvable.

Design a vertical infographic titled "Why HydroCore Beats the Bottle."
Aspect ratio: 4:5 (Instagram post).

Four data points, each with a labeled circular progress indicator:
- 24 hours: "Cold Retention"
- 98%: "Customer Satisfaction"
- 0: "BPA Content"
- 14 oz: "Total Weight"

Below: "Built for the 5 AM runner, the 9 AM commuter, and the 3 PM
break."

Modern sans-serif, bold numerals, lighter labels. Deep navy background,
warm cream text, coral accent on the progress arcs. Small HydroCore
wordmark at the bottom.

Numbers must render big, readable, and exact.

GPT-Images-2

Nano Banana 2

The Scorecard

Grade each output on five axes. 1 to 5 per axis, one note per prompt.

🔶 Brand coherence: on-brand or stock?
🔶 Text accuracy: numbers, labels, words rendered correctly?
🔶 Multi-image consistency: across Prompt 1's 5 slides, do people and scenes hold?
🔶 Amazon-spec readiness: uploadable without a second pass?
🔶 Usability: would you actually ship this in a live campaign?

The model with the higher average wins that prompt. But pay attention to where each model wins. It's rarely a blowout.

Who Should Switch, Who Should Stay

🔶 Dense text work (A+ content, comparison charts, data infographics, UI mockups, product labels): Images 2.0 Thinking mode looks like the new leader. The text rendering gap is real.
🔶 High-volume product photography (10 to 100 variations a week): Nano Banana 2 is still the workhorse. 3x faster, half the cost, 95% of the quality for standard product shots.
🔶 Multi-image campaigns (carousels, storyboards, before-and-afters): Images 2.0 Thinking mode is worth the subscription test. Native planning is built for this job.
🔶 Free plans: The non-Thinking version of Images 2.0 is free for everyone. Even without reasoning, the 2K default and improved text rendering justify adding it to your stack.

This isn't a "switch and never look back" moment. It's a "add a second tool and learn when to reach for which" moment.

We are living in exciting times that keep us on our toes!

NERD BYTES

Neat Little Hack: The WebP Drop Folder 🤓

Every image on your hard drive (png, jpeg, etc) is probably 3 to 10 times bigger than it needs to be.

That fat tax shows up everywhere. Slow Shopify pages. Ad platforms taking forever to approve creative. Supplier emails bouncing on attachment size. Blog and newsletter images that slow your own site down. Each one is small. Together they're a drag you've stopped noticing.

The fix is a file format called WebP.

What WebP Actually Is

Honestly? I didn't know what WebP was until a few months ago either.

WebP is a modern image format Google built in 2010. Roughly 70 to 80% smaller than PNG, 25 to 35% smaller than JPG, at the same visual quality. Shopify, Meta, Google Ads, and every modern browser support it natively. It's just not the default anywhere, so most of us never switched.

One heads up for Amazon sellers: Amazon Seller Central does not accept WebP uploads for product listing images. Amazon still requires JPEG, PNG, TIFF, or GIF. So WebP is for everything around your Amazon listings (Shopify stores, ad platforms, email, blog images), not the listings themselves.

What makes the conversion possible: FFMPEG

What is FFMPEG? A free, twenty-year-old command-line tool that can convert, compress, and resize any image or video. Powerful, intimidating. Nobody casually opens a terminal for it.

Until Claude became the front end.

The Drop Folder Workflow

Here's the whole setup:

Make a folder on your desktop called webp-drop.
Drop all the images you want to shrink into it.
Point Claude Code to that foldr and type: "Convert the images in /webp-drop to WebP format."

Claude runs FFMPEG, saves the smaller version back to the folder, done. About four seconds per file. Done.

What You Actually Save

🔶 1.5MB PNG screenshot → ~200KB
🔶 800KB product JPG → ~280KB
🔶 4MB creative export → ~700KB

Same visual quality. Tiny fraction of the file size.

One-Time Setup

Install FFMPEG (brew install ffmpeg on Mac, winget install ffmpeg on Windows). Make the folder. Done forever.

The quiet advantage isn't WebP, and it isn't FFMPEG. It's that boring workflows you used to ignore can now be a one-sentence request. Every seller I know has a dozen of these hiding in their week.

COOL TOOLS

Anthropic quietly shipped Claude Design on April 17 as a new Anthropic Labs product. It's built for the "I need a polished visual in ten minutes and I don't want to open a design tool" moment we all have three times a week.

What It Actually Does

You describe what you need. Claude builds a first version. You refine through conversation, inline comments, direct edits, or custom sliders (Claude generates the sliders based on what you might want to tweak). When you're happy, export to Canva, PDF, PPTX, or standalone HTML.

It also learns. Feed it your brand colors, fonts, and logo once. Every design after that uses them by default.

The Quiet Upgrade: It Runs on Opus 4.7

Claude Design is powered by Opus 4.7, which means it inherits the tripled vision resolution. Drop in a screenshot of a competitor's A+ content and it can actually read every headline, bullet, and comparison row. Then ask it to riff on the structure for your own product. That combination (competitor analysis plus fast visual iteration) is the specific use case I'm testing on a real listing this week.

The Catch

Research preview. Included free for anyone already on Claude Pro, Max, Team, or Enterprise. No extra charge, but also no standalone free tier. If you're on Claude Free, you'll need to upgrade to try it.

I'm keeping this one in my workflow rotation for a month and will share what held up (and what didn't) in a future issue.

NEWS WORTH FOLLOWING

🔶 Anthropic Passes OpenAI in Revenue for the First Time

🔶 Amazon CEO Jassy Defends $200B AI Spend: "We're Not Going to Be Conservative"

🔶 Shopify AI Toolkit Lets Claude Code Run Your Entire Store

🔶 Amazon's "Buy For Me" AI Agent Shops Other Websites Without Leaving the App

🔶 OpenAI, Anthropic, and Google Unite Against Chinese AI Model Theft

🔶 Google Gemini Now Generates Interactive Visualizations in Chat

UPCOMING EVENTS

April 28th is the last Tuesday of the month. Time to tune in to the Go with the Flow Podcast where Danny McMillan and I dive into AI topics. We will be talking about Claude Code, specifically some of the newer roll-outs and ways to improve Claude Skills. Listen here!

My good friend, Gary Huang, (who runs the 7 Figure Seller Summit) is hosting a live session next week focused on: From AI Tools to AI Systems: What Actually Works for Amazon Sellers. RSVP here: 👉 https://newsletter.7figuresellersummit.com/ocea3

HubSpot's ex-Head of Paid shares his 2026 playbook

Rex Gelb spent a decade building HubSpot's paid engine. Now he's showing founders exactly how to do it.

On April 27th, get the framework to structure, launch, and scale paid media that drives pipeline, not just traffic. 20 minutes. Live Q&A. Free.

Join the free session

🗞️ Build Your Own 'Keepa' with Claude 💥🚀

Ritu Java — Wed, 15 Apr 2026 12:07:00 +0000

In this edition (shortcuts):

1) Build your own version of Keepa for free!
2) A neat little hack to save precious tokens
3) Leveraging a FREE TikTok data resource
4) Make Claude run on a Schedule. Discover Routines
5) AI news you can’t miss

I Canceled Keepa. Here's What I Built Instead.

I've been paying for Keepa for a while now. €19 a month. About $21 USD. It's a great tool. Price history, BSR tracking, alerts, the whole deal.

But here's the thing. I wasn't using 90% of it. What I actually needed was simple: check a list of competitor ASINs every day, grab the current price and title, and log the running list to a spreadsheet so I can spot changes.

That's a $21/month problem. And I solved it in 20 minutes with Claude Code. No coding. No subscription. Just a script that runs every morning at 9am, opens each Amazon product page in a headless browser, pulls the price and product title, and drops the results straight into a Google Sheet.

$250 a year back in my pocket. But honestly, the money isn't the point. The point is I own this now. I can customize it, extend it, add BSR tracking later, share it with my team without paying per seat. It does exactly what I need and nothing I don't.

Here's exactly how I built it, step by step, so you can build yours today.

What You're Building

A daily automated task that does three things:

🔶 Opens each Amazon product page in a headless browser (so you don't even see it happen)
🔶 Extracts the product title and current price
🔶 Logs everything to a Google Sheet with today's date

It runs on autopilot. Every morning. No clicking, no copying, no pasting.

The whole setup takes about 30 minutes. You'll need three free tools and a Mac (sorry Windows friends, I'll cover you in a future issue). But the best thing is that you can simply paste my entire post into Claude Code and ask IT to do the setup for you.

The Three Tools You Need

Here's your toolkit. All free.

Tool	What It Does	Where to Get It
Claude Code	Orchestrates the entire workflow	claude.ai/code
Playwright Plugin	Headless browser that visits Amazon pages	Installed via Claude Code (see Step 1)
gws CLI	Reads and writes Google Sheets from your terminal	github.com/googleworkspace/cli

A quick note on gws. It stands for Google Workspace CLI. It's an official Google tool that lets you talk to Google Sheets, Docs, Drive, Gmail, and more from your terminal. Think of it as a remote control for your Google account. You'll also need the Google Cloud SDK (gcloud) for authentication. Grab it at cloud.google.com/sdk.

Once you've downloaded Claude Code, gws, and gcloud, run these two commands in your terminal to log in:

gcloud auth login
gws auth login

Both will open a browser window where you sign in with your Google account. Do it once and you're set.

Step 1: Enable the Playwright Plugin

Open your terminal and run this:

claude mcp add playwright npx @playwright/mcp@latest

That's it. One line. This tells Claude Code to use Playwright as its browser engine.

To verify it worked, open Claude Code and type:

Open https://www.amazon.com using Playwright and take a screenshot

You should see a screenshot of the Amazon homepage appear. If you do, you're good.

Pro Tip: You'll need Node.js 18 or higher installed for Playwright to work. Run node --version in your terminal to check. If you're on an older version, grab the latest from nodejs.org.

[IMAGE: Screenshot of Claude Code showing the Playwright verification step with Amazon homepage screenshot]

Step 2: Set Up Your Google Sheet

Create a new Google Sheet (or use an existing one). Add these headers in Row 1:

A	B	C	D
Date	ASIN	Product	Price

Now grab the spreadsheet ID from your URL. It's the long string between /d/and /edit:

https://docs.google.com/spreadsheets/d/THIS_PART_RIGHT_HERE/edit

Copy that ID. You'll need it in a minute.

To make sure gws can see your sheet, run this in your terminal (replace the ID with yours):

gws sheets spreadsheets get --params '{"spreadsheetId": "YOUR_SPREADSHEET_ID"}' --format json

If you see sheet details come back, you're connected.

[IMAGE: Screenshot of Google Sheet with headers set up and the spreadsheet ID highlighted in the URL bar]

Step 3: Create the Script

This is where it gets fun. You're going to create a small script that tells Claude Code exactly what to do each morning.

Open your terminal and run these commands one at a time.

Create a folder for your scripts:

mkdir -p ~/Scripts

Now create the script file. This is a longer command, so I'll walk you through it:

cat > ~/Scripts/amazon-price-tracker.sh << 'SCRIPT_END'
#!/bin/bash
LOGFILE="$HOME/Scripts/amazon-price-tracker.log"
echo "$(date): Starting Amazon price tracker" >> "$LOGFILE"

PROMPT='You are an automated Amazon price scraper. Do the following steps exactly:

SHEET_ID: YOUR_SPREADSHEET_ID
ASINs: ASIN1, ASIN2, ASIN3

STEP 1 - Check for duplicates:
Run: gws sheets spreadsheets values get --params "{\"spreadsheetId\": \"YOUR_SPREADSHEET_ID\", \"range\": \"Sheet1!A:B\"}" --format json
Build a list of "date|ASIN" pairs already in the sheet. Skip any ASIN that already has an entry for today.

STEP 2 - Scrape prices:
For each non-duplicate ASIN, use Playwright MCP:
- Navigate to https://www.amazon.com/dp/{ASIN}
- Extract title and price using browser_evaluate with this function:
  () => { const pw = document.querySelector(".a-price .a-price-whole"); const pf = document.querySelector(".a-price .a-price-fraction"); const t = document.querySelector("#productTitle"); return JSON.stringify({ title: t ? t.textContent.trim() : "N/A", price: pw ? "$" + pw.textContent + (pf ? pf.textContent : "") : "Not found" }); }

STEP 3 - Write to Google Sheet:
Append results using gws CLI. Use today date as a text string in YYYY-MM-DD format (wrap in single quotes so Sheets does not convert it). Keep $ in prices.
Use amazon.com only (not .in or other locales).
Do not ask questions. Execute silently and report a summary.'

claude --print --allowedTools "mcp__plugin_playwright_playwright__browser_navigate,mcp__plugin_playwright_playwright__browser_evaluate,mcp__plugin_playwright_playwright__browser_snapshot,Bash,Read" -p "$PROMPT" >> "$LOGFILE" 2>&1

echo "$(date): Finished" >> "$LOGFILE"
SCRIPT_END

Before you move on, open the script in any text editor and replace:

🔶 YOUR_SPREADSHEET_ID (it appears twice) with your actual Google Sheet ID 🔶 ASIN1, ASIN2, ASIN3with your comma-separated list of ASINs

Then make it executable:

chmod +x ~/Scripts/amazon-price-tracker.sh

Step 4: Test It

Run the script manually first:

bash ~/Scripts/amazon-price-tracker.sh

This takes 5 to 10 minutes depending on how many ASINs you're tracking. Claude Code will launch in the background, open each Amazon page with Playwright, extract the data, and write it to your Google Sheet.

When it's done, check the log:

cat ~/Scripts/amazon-price-tracker.log

Then open your Google Sheet. You should see today's prices.

[IMAGE: Screenshot of Google Sheet populated with real price data from the first test run]

If everything looks good, you're ready to automate it.

Step 5: Schedule It to Run Every Morning

This part uses a Mac feature called launchd. It's like a built-in alarm clock for scripts.

First, find your Mac username by running whoami in Terminal. Then paste this command, replacing YOUR_USERNAME with what you got:

cat > ~/Library/LaunchAgents/com.pricetracker.plist << 'PLIST_END'




    Label
    com.pricetracker
    ProgramArguments
    
        /bin/bash
        /Users/YOUR_USERNAME/Scripts/amazon-price-tracker.sh
    
    StartCalendarInterval
    
        Hour
        9
        Minute
        0
    
    StandardErrorPath
    /Users/YOUR_USERNAME/Scripts/amazon-price-tracker-error.log
    EnvironmentVariables
    
        PATH
        /usr/local/bin:/usr/bin:/bin:/Users/YOUR_USERNAME/.local/bin
    


PLIST_END

Activate the schedule:

launchctl load ~/Library/LaunchAgents/com.pricetracker.plist

Verify it's running:

launchctl list | grep pricetracker

You should see a line with com.pricetracker. That means you're live.

You're Done. Here's What Happens Now.

Every morning at 900 AM, your Mac will automatically

🔶 Launch Claude Code in headless mode (no window, no interruption)
🔶 Open each Amazon product page in an invisible browser
🔶 Extract the product title and current price
🔶 Skip any ASINs already scraped today (smart duplicate checking)
🔶 Append the results to your Google Sheet

You don't touch anything. You don't click anything. You open your laptop, check your Sheet, and see exactly what your competitors are charging. Every single day.

Quick Reference: Managing Your Tracker

What You Want	What to Paste in Terminal
Run it right now	`bash ~/Scripts/amazon-price-tracker.sh`
Check the log	`cat ~/Scripts/amazon-price-tracker.log`
Pause the schedule	`launchctl unload ~/Library/LaunchAgents/com.pricetracker.plist`
Resume the schedule	`launchctl load ~/Library/LaunchAgents/com.pricetracker.plist`
Change ASINs	Open `~/Scripts/amazon-price-tracker.sh` in any text editor and update the list

A Few Things to Know

🔶 Your Mac needs to be on at 9am. If it's asleep, the task runs when it wakes up. If it was off, you'll need to run it manually that day.

🔶 Auth can expire. If gws stops working, run gws auth login again. Takes 30 seconds.

🔶 The duplicate check is smart. If you run the script twice in one day (or a teammate runs it too), it won't create duplicate entries. It checks what's already in the sheet before writing.

🔶 This tracks amazon.com prices. If you sell on other locales (.co.uk, .de, .co.jp), change the URL in the script from amazon.comto your marketplace.

The Bigger Picture

Here's what I keep coming back to.

Keepa is a good product. I'm not trashing it. But I was paying $21 a month for a fraction of what it offers, because that fraction was all I needed. And I built that fraction in 30 minutes with free tools.

That's the real lesson here. It's not about replacing Keepa specifically. It's about recognizing when you're renting a Swiss Army knife to open one bottle. Claude Code, Playwright, and gws are the bottle opener. Free. Yours. Customizable.

This is what I mean when I talk about the "Own, Don't Rent" mindset. The tools are here. The access is here. The gap is just knowing how to wire them together.

And that gap? It's closing fast.

What competitor data are you tracking manually right now? Hit reply and tell me. I bet there's an automation hiding in your workflow.

Know someone still paying for a price tracker? Forward this to them. They'll thank you.

~Ritu

PPPC Ninja is helping brands future proof their listings for AI. We'll audit your listings for RUFUS readiness, build stunning AI images and videos, and scale your content across Amazon ads, posts, and social media. Learn more about our RUFUS Audit or hit reply to chat with us. https://www.ppcninja.com/rufus-audit.html

Want to learn how you can access INSANE amounts of free data from TikTok? You need to check out TikTok Creative Center.

Now, you don’t need to be advertising or promoting any products at all on TikTok to use this free platform. Just go here, login with a Google or TikTok account and instantly get access to all kinds of data related to Top Ads, Keywords, Top Products, Trends and a lot more.

To access trending hashtags, click on Trends in the menu bar.

TikTok categorizes trends based on their “Speeds of Culture” into three buckets:

Trend Moments: Creative prompts that quickly gain traction and buzz through high participation
Trend Signals: An emerging user behavior or interest revealed through new content patterns
Trend Forces: Enduring, large-scale behavioral transformations

If you want to view the Top Ads on TikTok go to their latest platform called TikTok One. Here you can watch all their top videos, access

To access specific trends, choose a category, such as Beauty & Personal care and pick your time window (7d, 30d, 120d). You can now see the top trending hashtags, the trend line, the top creators, and detailed analytics for each where you can analyze the actual videos themselves.

Even if you don’t advertise on TikTok, this is a great resource for learning what keywords are popular in Trend Moments (7d), Trend Signals (30d) or Trend Forces (120d) buckets.

To make things super easy, this platform comes with “Symphony Assistant”, an AI chatbot that can give you all the answers you seek quickly:

I asked: “Give me trend signals and keyword insights related to foundation”, and this is what it returned:

Since we know that Amazon is also combing the internet for signals and connecting the dots, the BIGGEST use case of this data is to include viral keywords coming from TikTok in your product’s listing or backend, as appropriate.

Given that RUFUS, Amazon’s shopping assistant is scraping your listing to get a good idea of it’s use cases for different personas, you might want to proactively insert viral keywords and use cases into your listing so that RUFUS can lead potential shoppers to your product detail page.

Claude Code Just Got Autopilot - Say Hello to Routines 🤓

You guys know I love a good automation that runs without me. Well, Anthropic dropped a quiet but big one this week. On April 14, 2026 they launched a research preview of Claude Code Routines. Think of it as scheduled tasks on steroids, except the tasks run in the cloud, not on your laptop.

If you ever typed /schedule in Claude Code before, that feature just grew up.

What's Actually New

A Routine is a saved Claude Code config: a prompt, a repo, and a set of connectors, packaged once and run automatically. Three ways to kick one off:

🔶 On a schedule. Hourly, nightly, weekly. Cron style, friendly UI
🔶 By API. Every routine gets its own endpoint and auth token. POST a message, get back a session URL
🔶 By webhook or GitHub event. React to a PR open, a push, a new issue, or any external signal

The biggest shift? Routines run on Anthropic's cloud infrastructure, not your machine. Close your laptop at 6pm. Your 9am Monday routine still runs.

There's a "local routines" option too (you may have seen the "Local routines only run while your computer is awake" note in the UI). But the headline feature is cloud, and that's where the real power lives.

The Fine Print

🔶 Pro plan: 5 routines per day
🔶 Max plan: 15 routines per day
🔶 Team/Enterprise: 25 routines per day
🔶 Available today in research preview at claude.ai/code or via /schedule in the CLI

Use Cases Worth Stealing

For sellers:
🔶 Weekly PPC health check. Every Monday 9am, Claude pulls your top campaigns, flags spikes or drops, emails you a summary
🔶 Daily competitor price log. Every morning, scrape your tracked ASINs, log to Google Sheets (see the main feature if you haven't yet)
🔶 Review drop alert. Webhook fires when a 1 star review lands, Claude reads it, drafts a response

For content creators:
🔶 Newsletter idea pipeline. Sunday night, scrape your favorite AI newsletters, de duplicate, drop 10 topic ideas in a Google Doc
🔶 LinkedIn queue from Git. Push a new blog post to a repo, a GitHub event triggers a Routine that generates 3 LinkedIn posts in your voice

For operations:
🔶 SOP from screen recording. Save a screen recording to a watched folder, a local Routine picks it up, extracts the steps, writes a Google Doc and a reusable skill file

Why This Matters More Than It Looks

Claude used to need you. You opened it, you typed, you waited. Routines flips that. Claude sits quietly in the cloud, watching for signals, acting on its own. That's the move from assistant to agent.

And Anthropic is capping it at 5 to 25 runs per day for now because they know the second this gets unlimited, people will build actual companies on top of it. Fair.

What's the one recurring task you'd hand to a Routine tomorrow? Hit reply and tell me. I'm collecting favorites for a walkthrough issue.

Should You Really Be Pasting Screenshots Into Claude? 🤓

Here's a habit most of us picked up without thinking. Something goes wrong. Something looks weird. We hit ⌘+Shift+4, drag a box, and paste the screenshot into Claude.

Quick. Easy. And way more expensive than you realize.

The Screenshot Tax

Anthropic has a published formula for how images are priced. It's this:

tokens = (width × height) / 750

That means a plain full screen grab on a standard 1920×1080 monitor costs you about 2,765 tokens. A Retina or 4K screenshot? Closer to 5,000 to 10,000 tokens. An iPhone screenshot is around 1,850 tokens.

For context, 2,765 tokens is roughly 10,000 characters of plain text. That's about 15 pages of documentation. For one screenshot of one browser window.

The Text Alternative

Here's the gap. If the thing you're pasting is mostly words (an error log, a terminal output, a product description, a spreadsheet row), copying the actual text instead of the image can cut your token cost by 10x or more.

🔶 A 200 word paragraph pasted as text: ~267 tokens 🔶 The same paragraph as a screenshot: 2,000 to 5,000 tokens

Same information. Tenth of the cost. Often a better answer too, because Claude doesn't have to OCR anything.

The Playwright Back Door

If you're using Playwright (like in the tracker above), there's a third option most people miss: browser_snapshot. It skips the image entirely and returns a structured text tree of the page. For data tasks (prices, titles, specs, reviews), it's dramatically cheaper than a screenshot and more accurate than OCR.

When Screenshots Actually Win

Screenshots aren't evil. They're just overused. Paste an image when the visual itself is the information:

🔶 A layout bug or weird spacing issue 🔶 A chart, graph, or hand drawn sketch 🔶 A design you want critiqued 🔶 Handwriting you need transcribed

Paste text (or use browser_snapshot) for everything else.

The tokens you save today are the tokens you get to spend tomorrow. Quiet advantage.

News worth following 📡 ⚡ 🌐 📰

🔶 Anthropic Passes OpenAI in Revenue for the First Time
🔶 Amazon CEO Jassy Defends $200B AI Spend: "We're Not Going to Be Conservative"
🔶 Shopify AI Toolkit Lets Claude Code Run Your Entire Store 🔶 Amazon's "Buy For Me" AI Agent Shops Other Websites Without Leaving the App
🔶 OpenAI, Anthropic, and Google Unite Against Chinese AI Model Theft
🔶 Google Gemini Now Generates Interactive Visualizations in Chat

🗞️ AI Without Plumbing Is Just a Party Trick💥🚀

Ritu Java — Tue, 07 Apr 2026 15:07:00 +0000

AI Without Plumbing Is Just a Party Trick 🔧

You typed a prompt. You got a stunning image. Cool, right?

Now do it 50 more times. For 200 SKUs. With consistent brand colors. In three ad formats. By Friday.

That's where the party trick falls apart.

Nate B. Jones recently analyzed the massive leak of Claude Code’s source code (Anthropic accidentally published it) and found something that every seller should hear: the actual AI call is only 20% of the system. The other 80%? Plumbing. Session management, error handling, tool coordination. Infrastructure.

Here's why this matters to you.

When we build any AI creative at PPC Ninja, the "generate image" step is ONE step in a pipeline of many.

Extract brand colors for consistency.
Generate an AI background.
Detach elements.
Positoin product.
Add text.
Resize for the ad format.
Upscale for quality.
…..

We’ve made tools for every conceivable step of the way. 80% of the steps are plumbing.

And they're the reason we can easily deliver 100+ RUFUS audits a week instead of weeks or months.

The rule we follow now:

Invest 80% of your AI effort into building the workflow, the plumbing tools. The other 20%, the actual "hit generate" moment, is the easy part.

A prompt is a party trick. A pipeline is a business.

So next time you're tempted to spend an hour perfecting one prompt, ask yourself: do I have a system that could run this 50 times without me?

If not, you don't have an AI advantage. You have a hobby.

Build the pipes!

Google AI Studio

One of the best way to invest in plumbing is to build your own tools. If you build once you can use them over and over again. This is the basic principle of automation.

Google AI Studio makes all of this super easy.

Step 1: Describe your app. Type what you want to build in the prompt box. Be as specific as possible. For example: "A tool that takes a product image and lets me choose from 5 AI background styles, then composites them together." or “Create an app where I can upload a script and receive a sequence of generated storyboard images.”

Step 2: Add capabilities with the chips. See those buttons at the bottom (Generate music, Add database and auth, Create & edit images, Add voice conversations)? Those are AI Chips. Click any that apply to your app and they get folded into your prompt automatically. For an image processing tool, you'd tap "Create & edit images."

Step 3: Hit enter (or "I'm feeling lucky" if you want Gemini to surprise you). Gemini generates the full code and shows a live preview on the right side. It defaults to a React frontend.

Step 4: Iterate via chat or annotations. Don't like something? You can either type a follow-up in chat ("make the background selector a grid instead of a dropdown") or use annotation mode to highlight a specific part of the UI and describe what you want changed.

Step 5: Deploy. When it's ready, you can deploy to Google Cloud Run or share with collaborators.

The whole thing is basically vibe coding inside Google's environment. No local setup, no terminal. You describe, it builds, you refine.

One thing to note: this is the same Google AI Studio you already use for image processing (resize, upscale, background removal). The "Build" mode just adds app generation on top of it.

Scheduled Tasks Are the Plumbing You Didn't Know You Needed 🤓

You know what's even better than a great AI workflow? One that runs without you.

The Claude desktop app now supports scheduled tasks. You describe a task once, set a cadence (daily, weekly, hourly, whatever you want), and it runs automatically in its own session with full access to your tools and data.

In Cowork (the desktop UI), click "Scheduled" in the sidebar, write your instructions, pick a frequency, and you're done. No deployment pipeline. No terminal.

If you're a developer using Claude Code (the CLI version), you get the same feature with /schedule and cron expressions for precise timing.

Here's where this ties back to the plumbing thesis.

A scheduled task IS plumbing. It's the part of your workflow that says "run this pipeline every Monday at 9am whether I remember to or not." I use one to collect newsletter topic ideas every week. It searches the web for AI and e-commerce news, curates the top stories, and drops a markdown file into my folder. By the time I sit down to write, the research is already done.

That's one task. Now imagine stacking five of them. Weekly competitor price checks. Daily search term report analysis. Monday morning ad performance summaries. A Friday creative audit.

Each one takes 10 minutes to set up. Then it runs forever (well, as long as your computer is awake and Claude is open).

Example:
"What's new in AI" briefing. You tell Claude "Every Monday at 8am, search for the latest AI tools and news relevant to e-commerce, summarize the top 5 stories, and save a report to my folder." You wake up Monday, coffee in hand, and your research is already sitting there.

The catch: scheduled tasks only run locally. No cloud execution yet. Your machine needs to be on. But for most sellers working from a desk? That's not a dealbreaker.

Pro Tip: Start with one. Pick the task you do every week that's mostly "go gather this information and summarize it." That's your first scheduled pipe.

The tools are there. The automation is there. Most people just don't know where to look.

News worth following 📡 ⚡ 🌐 📰

🔶 Anthropic Accidentally Leaks Claude Code's Entire Source Code
🔶 Anthropic's Leaked "Claude Mythos" Model Shows a "Step Change" in AI Capabilities
🔶 Amazon Launches "Dynamic Canvas" AI Dashboard in Seller Central
🔶 Shopify's Agentic Storefronts Let ChatGPT and Gemini Buy From Your Store
🔶 OpenAI Fully Retires GPT-4o From All Plans
🔶 Google's Universal Commerce Protocol Makes Search Shoppable via AI Agents
🔶 Gmail is entering the AI Era

🗞️ I Made a Video Without Opening a Video Editor💥🚀

Ritu Java — Wed, 01 Apr 2026 13:07:00 +0000

🗞️ I Made a Video Without Opening a Video Editor💥🚀

You know those animated explainer videos on Instagram and TikTok? The ones with clean text animations, progress bars that fill up, and a branded outro with a glowing call to action?

I made one last week. In about 10 minutes. Without touching After Effects, Canva, or any video editor at all.

I typed a prompt. Claude wrote the code. I hit render. Done.

The tool that makes this possible is called Remotion. And when you pair it with Claude Code, you get something that feels almost unfair: a full video production pipeline controlled entirely by conversation.

Let me show you how it works, how to set it up, and where this is headed.

Wait, Code? I Thought This Was About Video.

Here's the concept. Remotion is an open source framework that treats video like a React application. Every frame of your video is a React component. Your animations are JavaScript functions.

Your timeline is code. That sounds technical, but here's why it matters for you: Claude already knows how to write React. So when you install the Remotion skill into Claude Code, you're basically giving your AI assistant the ability to produce broadcast quality animations from a text prompt.

No timeline dragging. No keyframe tweaking. No export settings rabbit hole. You describe what you want, Claude builds it, and Remotion renders it to MP4.

This is what my experimental video looks like (remember, I spent very little time on it). Honestly, I was pretty impressed.

How to Set This Up (Three Ways)

Here's the good news. Getting started is straightforward. You need Claude Code (which requires a paid Claude subscription) and Node.js installed. After that, you have options:

Option 1: One Command Install (Easiest)

Open your terminal in a Remotion project and run:

npx skills add remotion-dev/skills

This installs the official Remotion skill into Claude Code. It teaches Claude best practices for animations, timing, text rendering, transitions, captions, audio, and more. Think of it as giving Claude a crash course in video production.

Option 2: Use a Starter Template

If you want a project that's already wired up and ready to go, clone one of the community starter repos:

npx create-video@latest

This scaffolds a fresh Remotion project. Then add the Claude skill on top of it.

Option 3: Let Claude Do Everything

Open Claude Code in any folder and say: "Set up a Remotion video project." Claude will search for the Remotion skill, install it, create the project structure, and configure everything. You just approve the steps.

Once you're set up, the workflow is:

Start the preview server (npm run dev)
Tell Claude what video you want
Preview it in your browser at localhost:3000
Render it (npx remotion render)

That's the whole process. Describe, preview, render.

The Skill Is the Secret Sauce

When I looked under the hood of my project, I found something that surprised me. The Remotion skill that Claude installs isn't just a basic setup guide. It's a comprehensive library of 30+ specialized rules covering:

🔶 Animations: Spring physics, interpolation, staggered timing
🔶 Text: Typewriter effects, word highlighting, measuring text for layout
🔶 Video: Embedding existing footage, trimming, transparent video support 🔶 Audio: Background music, sound effects, voiceover integration, audio visualization
🔶 Captions: Auto transcription, SRT import, animated subtitle display
🔶 Charts: Animated bar charts, data visualization
🔶 Advanced: 3D rendering, Lottie animations, light leak effects, maps

This is what makes the output good on the first try. Claude isn't guessing. It's following production tested patterns.

But Can It Edit Existing Video?

The answer is yes, with some nuance.

Remotion has a component that lets you import existing video files. You drop your MP4 into the public/folder, reference it with staticFile(), and now it's a layer in your composition. From there you can:

🔶 Overlay animated text on top of your footage (think: captions, titles, lower thirds)
🔶 Trim and sequence clips using startFrom and endAt props
🔶 Add animated graphics that appear over your talking head video 🔶 Composite multiple videos together with different timing
🔶 Add captions from an SRT file or auto generated transcript

The skill even has specific rules for avatar overlays (animated graphics on top of talking head video) and caption display. So you could take a recording of yourself, import it into Remotion, and ask Claude to add animated lower thirds, data callouts, or branded intros and outros.

For heavier editing (merging clips, adding audio tracks, format conversion), there's also an open source mcp-video server that wraps FFmpeg and works alongside Remotion for the post production pipeline.

The bottom line: Remotion isn't a replacement for DaVinci Resolve or Premiere. But for the kind of quick, branded, animated content that sellers need for social media? It's faster than anything I've used.

Prompt Templates: Your Starting Point

My project came with five ready to use prompt templates that you can hand to Claude for instant results:

#	Template	What It Creates
1	Educational Explainer	30s animated explainer on any topic
2	Product Demo	25s product showcase from any URL
3	Google Reviews	20s testimonial spotlight
4	Avatar Overlays	Animated graphics on talking head video
5	Data Visualization	15s animated dashboard from CSV data

Each template includes specs for safe zones (so your text doesn't get hidden behind social media UI), minimum font sizes, color schemes, and scene structure. You swap in your topic and go.

Why This Matters for E-Commerce Operators

Let's be honest. Most Amazon sellers aren't making enough video content. The ones who are usually outsource it (expensive) or use Canva templates (generic).

Remotion plus Claude flips both of those problems:

🔶 Cost: Free for individual sellers and small teams (under 4 employees). No monthly subscription
🔶 Speed: Concept to rendered MP4 in minutes, not days 🔶 Customization: Every pixel is code, so everything is customizable. Colors, timing, layout, data
🔶 Scale: Once you have a template, generating 50 variations (different products, different data, different languages) is trivial
🔶 Brand consistency: Your design system lives in code. Every video matches your brand automatically

That last point about scale is the one that gets me excited. Someone recently shared that they generated 50 personalized demo videos from a single Remotion component in 15 minutes. That's the kind of leverage that used to require a production team.

The Licensing Fine Print

One thing to know: Remotion is free for individuals, nonprofits, and companies with three or fewer employees. If your company has four or more employees, you'll need a commercial license. Worth checking before you scale up.

The Bigger Picture

Six months ago, making a professional animated video meant either learning After Effects or paying someone who had. Today, you can describe what you want in plain English and have a rendered MP4 before your coffee gets cold.

This is another example of the pattern we keep seeing. The tools are getting powerful enough that the bottleneck isn't technical skill anymore. It's knowing the tools exist and taking 20 minutes to try them.

Remotion plus Claude isn't perfect. Complex animations sometimes need a few rounds of iteration. The preview server takes some getting used to. And you'll want to invest time in learning what makes a good video prompt versus a vague one. But the tradeoffs are worth it.

But the floor has been raised. The minimum viable video for your brand just got a lot easier to create.

And it's only going to get better from here.

Have you tried making videos with Remotion and Claude? Or are you using something else for your social media content? Hit reply and let me know. I'm collecting workflows.

Know a seller who keeps saying "I should be making more video content"? Forward this to them. No excuses left.

~Ritu

I build these AI proof-of-concepts so you can implement them in minutes, not days. Did this one hit the mark? Reply and tell me "Yes" or "Not quite". Your feedback literally shapes next week's edition.

Spread the word: If you enjoyed this, share it with one person who needs an AI edge. They’ll thank you for the head start!

— Ritu 😃

🗞️ Stop Spying With Your Eyes. Let Claude Do It.💥🚀

Ritu Java — Tue, 24 Mar 2026 15:08:00 +0000

Claude Code and Cowork can now use your computer

You know that feeling. You're scrolling through your competitor's listings at 10 PM because you need to know what keywords they're ranking for, what their price point is, what reviews are saying about their product variants. So you dig. Tab after tab. Click after click. Note-taking friction. Browser history looking like an archeological dig site.

What if you just... told Claude to do it?

Claude's computer use feature (available to Max subscribers as of March 2026) means you can hand an AI agent a task ("analyze my top 5 competitors' Amazon listings and pull their pricing, keywords, review sentiment, and product variants") and come back in 30 seconds to a structured report.

No clicking. No manual note-taking. No pretending you're not getting distracted by the comments on their listing again. Watch this video from Anthropic.

Here's exactly how to set it up, what you can actually pull, and the kinds of competitive insights that will move the needle on your product strategy.

What Computer Use Actually Is (And Isn't)

Before we jump in, let's be clear about what we're talking about.

Computer use is Claude's ability to see your screen, click buttons, fill forms, and navigate websites on your behalf. It's like having someone sitting at your computer, taking direction from you in plain English, and doing the boring clicking work while you think about strategy.

It's not a 24/7 agent. It's not autonomous. You stay in control.

Here's what matters: you tell Claude what you want ("find me the price history for this product"), Claude sees your screen in real-time, figures out what to click, does the thing, and reports back with structured data. All within a single conversation.

This is different from APIs or browser automation tools. You're not building code. You're not hiring a dev. You're just... talking to an AI that can use a computer.

Setting Up Computer Use for Competitor Research

You'll Need:

🔶 Claude API Max plan or Claude.ai with Computer Use access (subscription feature, available to paid users as of March 2026)
🔶 A browser (Chrome, Safari, Firefox... doesn't matter)
🔶 A list of competitors (ASINs, URLs, or just names)
🔶 A specific goal (pricing tracking, keyword analysis, review sentiment, product variant mapping)

The Setup (5 Minutes):

The Setup (2 Minutes):

Open the Claude desktop app on your Mac
Go to Settings (gear icon) —> General
Find the Computer use toggle. It says "Let Claude take screenshots and control your keyboard and mouse in apps you allow"
Flip that toggle on
macOS will ask you to grant two permissions: Accessibility (so Claude can click, type, and scroll) and Screen Recording (so Claude can see your screen). Approve both in System Settings
Start a new conversation and ask Claude to do something on your computer. The first time it needs to touch a specific app, it'll ask your permission

That's it. No API keys. No terminal. No code.

Pro Tip: Start simple. Test with one competitor URL before building out a full competitive analysis workflow. Let Claude fail on something small so you understand how it thinks about navigation. More safety tips.

Real Example: The 3-Minute Competitor Pull

Let's say you sell face serums on Amazon. You want to know what your top 3 competitors are doing. Here's the prompt:

I'm running a competitive analysis. For each of these ASINs,
navigate to their Amazon listing and pull:
- Current price and any discounts
- Keywords from title and first 3 bullets
- Star rating and review count
- Top 5 review themes from recent reviews

Compare all three side-by-side in a table. Then give me
3 strategic insights about where I can differentiate.

Competitors: [ASIN 1], [ASIN 2], [ASIN 3]

Claude opens your browser, navigates to each listing, scrolls through, reads the content, and comes back with something like this:

	SeraLux	GlowTech	EssenceMax
Price	$54.99 (S&S 20% off)	$32.00 (Prime Deal)	$14.99
Rating	4.6 (2,341 reviews)	4.4 (8,720 reviews)	4.3 (15,400 reviews)
Top keyword	Anti-aging	Lightweight formula	Budget-friendly
Top review theme	"Hydration" (73%)	"Non-greasy" (58%)	"Unexpected quality" (35%)

Strategic insight from Claude: all three are competing on hydration. But nobody owns "sensitive skin" or "acne-prone." If you position around a specific skin concern, you could own a vertical they're not fighting for.

The whole analysis took about 12 minutes for 3 competitors. Manually? That's a two hour job.

Important to know: Claude can only see what a normal shopper sees. It can pull pricing, keywords, reviews, variants, A+ content, badges, and ranking signals. It cannot see backend keyword data, competitor ad spend, or Seller Central metrics. The advantage isn't secret data. It's reading a lot of visible data really fast and organizing it into decisions.

Quick Tips

🔶 Be specific. "Analyze my competitor" is vague. "Pull the price, first 3 bullets, star rating, and top 3 review complaints" is actionable. Specificity helps Claude click faster and extract cleaner.
🔶 Don't forget variants. Competitors often hide their best sellers in size/color variants. Ask Claude to click through each one and pull pricing per variant.
🔶 Run it monthly. Save your prompt. Reuse it. After 3 months you'll have trend data on price shifts, review velocity, and keyword changes. That's competitive intelligence, not just research.
🔶 Combine with your own data. Claude sees public listings. You see your Seller Central data. Together you can map competitor positioning against your actual conversion rates.

The Ethical Line (And Where It Is)

One question always comes up: Is this scraping? Is this against Amazon's ToS?

Here's the honest take.

You're not automating scraping. You're not building a bot. You're not extracting data at scale for resale. You're looking at public information (the same information your customers see) and making strategic decisions about your own product.

That's fine. That's what competitors do.

What would not be fine:

🔶 Using this to scrape and resell competitor data to third parties
🔶 Automating it at scale to feed a price-monitoring tool you're building to monetize
🔶 Accessing non-public information (like Seller Central data from someone else's account)

But pulling data from 5 to 10 competitor listings to inform your own strategy? That's just being a smart operator.

"But Wait... Is Claude Watching Me?"

Let's address the elephant in the room. You're giving an AI control of your mouse and keyboard. That sounds terrifying. So let's be really clear about what's actually happening.

🔶 Computer use is not always on. Claude only sees your screen and takes action when YOU start a session and give it a task. Close the conversation? It stops. It's NOT running in the background watching what you do.
🔶 It asks permission per app. The first time Claude needs to touch a new application, a prompt pops up asking you to allow or deny access for that session. You're in control of what it can and can't touch.
🔶 It won't access your bank. Anthropic has built in hard safety constraints. Claude will refuse to log into banking or financial sites, input sensitive credentials, engage in stock trading, or scrape facial images. I tested this myself. Asked it to go to a bank website and log in. It flat out said no.
🔶 It doesn't store what it sees. The screenshots Claude takes during a session are used to navigate in real time. They're not being saved to some server for training or future reference.

If you've heard of Microsoft's Recall feature (the one that continuously screenshots everything you do on Windows), this is not that. Recall was designed to be ambient and always on. The backlash was immediate and Microsoft had to delay and rework the entire feature. Claude's computer use is the opposite: it only activates when you tell it to, and it stops when you're done.

The Catch: Claude Takes Over Your Screen

One thing to know upfront. When Claude is using your computer, it is literally moving your cursor, clicking your apps, and scrolling your windows. You can't really use the computer at the same time. If you grabbed the mouse mid task, you'd be fighting each other.

For quick tasks (pulling data from one listing, checking a price), this is fine. You sit back for 90 seconds and watch it work.

For longer workflows (analyzing 5 competitors, pulling review data across multiple ASINs), you probably don't want to just sit there. This is where Dispatch comes in. Dispatch lets you assign Claude a task from your phone, walk away, and let it work on your Mac while you do something else entirely. When it's done (or needs permission for something), it pings your phone.

My recommendation: For anything that takes more than a couple of minutes, use Dispatch. Kick off the task from your phone, go do actual work, and come back to structured results. That's the workflow that makes this practical for busy sellers, not sitting at your desk watching a cursor move around on its own.

How This Compares to What Others Have Tried

If you've been following the "AI agent" space, you know Anthropic isn't the first to try this. Four players have taken a swing at "AI that controls your computer." Each one made very different choices.

Microsoft went big. They embedded AI directly into Windows with Recall, a feature that continuously screenshots everything you do and builds a searchable history. The backlash was immediate. Users called it surveillance. Microsoft had to delay, switch it to opt in, and add encryption. The trust damage was already done.

OpenAI launched Operator, a standalone agent that operates inside a sandboxed browser. You give it a task, it clicks through websites on your behalf. Cleaner than Microsoft's approach, but limited to browser tasks only. It can't touch your desktop apps.

OpenClaw is the one everyone's been talking about. If you've already dived into OpenClaw, you'll see similarities with Claude's new features. But Claude is not trying to be OpenClaw. OpenClaw is a personal assistant with a friendly chatbot vibe. It runs 24/7, browses the web, manages your calendar, controls smart home devices. Sounds amazing until you remember that a recent security audit found over 820 malicious skills in OpenClaw's community hub. Over 20% of available skills were flagged. That's a real concern if you're running a business.

Anthropic's Claude takes a different path entirely. Their releases are deliberate, targeted at knowledge work, not personal assistance. Like Operator, computer use is task invoked (not always on). Like OpenClaw, it can control any appon your Mac, not just a browser. But unlike Microsoft, it doesn't record anything. Unlike Operator, it's not sandboxed into a single window. And unlike OpenClaw, there's no community marketplace full of unvetted plugins. Claude asks permission per app, refuses to touch your bank, and stops the moment you close the conversation.

The bottom line: Anthropic watched what didn't work and built accordingly. No ambient monitoring. No OS level integration. No persistent recording. No sketchy plugin ecosystem. Just a capable work tool that uses your computer when you ask it to, and stops when you don't.

The Real Payoff

Here's why this matters beyond just "saving time."

Speed creates advantage. Not in copying (anyone can copy a title). But in decision velocity.

A seller who runs a 15-minute competitive analysis every month can respond to market shifts, keyword gaps, and positioning opportunities before they become obvious.

A seller who's doing this manually? They're still clicking through listings at 10 PM.

The learning window for early adopters of computer use is closing fast. In six months, this will be table stakes. The sellers who figured out how to pull competitive intel from Claude in March 2026 will be ahead of the sellers still clicking tabs in August.

That's the quiet advantage. Not the tool itself. The speed at which you can act on what you learn.

Have you tried Claude Computer Use or other agentic systems yet? What's the first task you'd hand off? Hit reply and tell me.

Know a seller who's curious about AI agents but worried about giving up control? Forward this to them. The safety section alone might change their mind.

~Ritu

PPC Ninja helps brands future proof their listings for AI, helps you build RUFUS enabled, stunning images and videos with AI. Hit reply on this to chat with us. Explore how we can scale your content production across Social media, Amazon ads, Amazon Posts efficiently and affordably.

Super excited to be speaking at the Billion Dollar Seller Summit BDSS 13, and AI-focused event hosted by Kevin King. RSVP here.

🚀 Never Pay for Another App! Learn How to Build Custom Features Yourself📍 Event: Seller Summit 2026
📅 Dates: April 21–23, 2026
Save your spot here: Link

🗞️ The AMC Unlock: AI Writes SQL So You Don't Have To💥🚀

Ritu Java — Tue, 17 Mar 2026 15:19:43 +0000

If you’re new here, welcome! If you’ve been reading for a while, thank you for sticking around as we navigate this wild AI shaped shift happening across eCommerce. Each week I share what I’m experimenting with, what’s actually moving the needle, and the trends that deserve your attention before they hit your competitors’ playbooks.

A quick heads up. I’ve organized all previous editions into one searchable hub. If you want the full journey, it’s all here.

The AMC Unlock: How AI Does the SQL So You Don't Have To

Most Amazon advertisers know AMC (Amazon Marketing Cloud) exists, that gold mine of insights from Amazon. Very few actually use it.

Not because Amazon’s data isn't valuable. It's some of the most powerful first-party data available on the platform. Path-to-purchase, new-to-brand attribution, audience overlap, cross-channel touchpoints. The insights are genuinely unfair.

The problem has always been the SQL barrier. AMC runs on query-based access, and if you're not a data analyst, that's where most brands stop cold.

The conversation around AMC has shifted. What used to require a data science team or a $50K agency engagement is now accessible to any brand willing to spend 20 minutes with the right AI prompt.

AMC Is Finally for Everyone

First of all, you’ll need to access AMC within your Amazon Ads Console. It’s a little tucked away:

Here's what's changed: The new Ads Agent is an AI powered conversational tool that allows you to describe what you want from Amazon’s data in plain English and the tool converts it into SQL.

Now this SQL query is likely going to be UNUSABLE, unless you are already experienced in SQL and know what to change. You will be asked to swap out sections of the query with your actual Campaign IDs etc. This task is often intimidating for normal people.

But have no fear, because AI models like Claude and ChatGPT can write AMC-compatible SQL queries from plain English descriptions. You describe what edits you want, and the AI writes the query, and you copy it into AMC. No coding background required.

Three AMC Queries You Can Generate with AI Today

You don't need to understand SQL to use these. Just copy the plain-English prompt into your AI tool of choice, then paste the output into AMC.

1. Path to Purchase Ask AI: "Write an AMC SQL query that shows me how many click based touch points customers had before converting.

2. New-to-Brand Attribution Ask AI: "Write an AMC query that shows new-to-brand orders by campaign type over the last 90 days."

Your standard Amazon Ads reports give you NTB data, but AMC lets you cut it by campaign, ASIN, and time window simultaneously. Completely different level of clarity.

3. Ad Type Overlap Ask AI: "Write an AMC query to show me the AOV of shoppers that were exposed to different Ad types"

I use Claude because it is great at handling complex, multi-table SQL well and is good at explaining what the query does line by line (helpful when you need to edit it)

One thing I see brands miss consistently: they pull AMC data once and never act on it. AMC is most powerful as a recurring diagnostic, not a one-time audit. Build a habit of running your top two or three queries monthly, and you'll start spotting patterns that change how you allocate budget.

The brands pulling ahead on Amazon right now aren't necessarily spending more. They're making better decisions faster. AMC plus AI is one of the clearest ways to get there.

Have you tried AI for AMC yet?

Amazon Creative Agent: The AI Creative Studio Your Competitors Are Sleeping On

Most Amazon advertisers know Creative Studio exists. Very few are using it to its full potential.

That just got a lot harder to ignore because Amazon released Creative Agent in February 2026, an agentic AI tool built directly inside Creative Studio. It supports the full creative process, from ideation and scriptwriting to images, video, animation, voiceovers, music, and final ad delivery, powered by Amazon's own retail insights. Amazon

This is not a generic image generator. Within Creative Studio, advertisers can click "chat" to access a conversational AI creative partner that conducts product and audience research, brainstorms ideas, develops creative concepts in storyboard format, and produces compelling video and display ads. It analyzes audience signals, product and brand details, and brand assets to create ad themes that resonate with a brand's audience. Amazon Ads

What makes it different from third-party tools: by combining customer shopping signals with information from the advertiser's product pages, Brand Store, and website, the tool generates ideas that truly bring a brand to life. Amazon Ads It is working from the same data that powers Amazon search and recommendations. No other platform has that.

The finished output can run across Sponsored Brands, Sponsored Display, Amazon DSP, and Brand Stores. Assets are saved directly to the Creative Asset library.

Honestly, it does not beat Nano Banana, but the price point is hard to argue with. Creative Agent is available at no additional cost to advertisers.

The numbers worth knowing:

On average, advertisers saw a 10.3% higher ROAS on Sponsored Brands campaigns that used AI-generated images compared to campaigns without. Brands using AI-powered creative solutions advertised five times more products and used twice as many images per advertised product. Amazon Ads

HERO IMAGE →

LIFESTYLE

Where it stands right now:

Creative Agent is currently in open beta for US advertisers. Image Generator is available in Canada as well. Video Generator is available for Sponsored Brands video in Canada, the US, and several other markets. Amazon Ads Check your Amazon Ads Console for access.

The bottom line: The biggest shift here is not the tool itself. It is that the creative barrier on Amazon is disappearing. Brands that used to be limited by budget or production capacity can now test more creative, more often, across more placements. The brands that will win are the ones who actually use it.

Claude Code ‘Hooks’ and ‘Skills’: How to Make AI Follow Your Rules

If you use Claude Code and you are only typing prompts, you are missing two features that change how the whole thing works.

Hooks are automatic rules you set once and Claude follows every single time without being asked. Think of them like a store policy. You do not remind every employee about the return policy every morning. You write it down once and it applies to everyone, always.

Some things you can automate with hooks:

🔶 Never let Claude touch certain files or run dangerous commands
🔶 Automatically run a quality check every time Claude makes a change
🔶 Get a desktop notification when Claude finishes a task so you can step away
🔶 Auto-save a backup before Claude makes any big edits

The difference from just asking Claude nicely in a prompt? A prompt is a suggestion. A hook is a guarantee.

Skills are reusable playbooks written in plain text files that tell Claude how to handle specific types of work.

At PPC Ninja we use skills for things like:

🔶 Conduct a PPC Audit
🔶 Creating Amazon bulk files with the right campaign structure
🔶 Write SQL queries for AMC

Instead of re-explaining your preferences every session, Claude reads the skill file and already knows your style, your format, and your standards before you say a word.

Together, hooks enforce the rules and skills carry the standards. Set them up once and the quality shows up every time.

Jensen Huang at GTC 2026: "Every Company Needs an OpenClaw Strategy"

Jensen Huang took the stage in San Jose yesterday at Nvidia's GTC 2026, and agentic AI was front and center.

Huang made his position clear: "For the CEOs, the question is, what's your OpenClaw strategy? We need it. We all have a Linux strategy. We all needed to have an HTTP HTML strategy, which started the internet." TechCrunch

That is not a subtle statement. He is putting OpenClaw, the open-source AI agent framework that has gone viral in recent weeks, in the same category as the technologies that built the modern web.

Huang framed it this way: "Mac and Windows are the operating systems for the personal computer. OpenClaw is the operating system for personal AI." The Motley Fool

The big announcement at GTC was NemoClaw. Nvidia has developed NemoClaw, an enterprise-grade AI agent platform built on top of OpenClaw, combining the ease of OpenClaw's agent-building capability with Nvidia's privacy and security controls needed to run agents within an enterprise. TechCrunch

NemoClaw serves as the policy engine for SaaS companies worldwide, combining policy enforcement, network guardrails, and privacy routing so enterprises can deploy autonomous agents across proprietary workflows with governance, isolation, and control. NVIDIA Blog

Beyond agents, Huang told attendees that Nvidia expects its flagship AI processors to help generate $1 trillion in sales through 2027, doubling down on an earlier forecast of $500 billion. CNBC

What this means for e-commerce operators:

The shift Huang is describing is not distant. Post-OpenClaw, he says, will turn IT from SaaS into GaaS, with AI acting as the service itself. Tom's Guide For e-commerce brands, that means the tools you are using today, your ad platforms, your listing optimization tools, your inventory systems, will increasingly have AI agents running inside them, not just AI features.

The question Huang is really asking you is not a technical one. It is a strategic one. If agents are the next interface, what tasks in your business are you going to hand them first?

Super excited to be speaking at the Billion Dollar Seller Summit BDSS 13, and AI-focused event hosted by Kevin King. RSVP: https://www.bdss13.com

🚀 Never Pay for Another App! Learn How to Build Custom Features Yourself📍 Event: Seller Summit 2026
📅 Dates: April 21–23, 2026
Save your spot here: Link

Attio is the AI CRM for modern teams.

Connect your email and calendar, and Attio instantly builds your CRM. Every contact, every company, every conversation, all organized in one place.

Then Ask Attio anything:

Prep for meetings in seconds with full context from across your business
Know what’s happening across your entire pipeline instantly
Spot deals going sideways before they do

No more digging and no more data entry. Just answers.

Start your free trial →

🗞️ The Nano Banana Switcheroo💥🚀

Ritu Java — Tue, 03 Mar 2026 16:07:00 +0000

A quick heads up. I’ve organized all previous editions into one searchable hub. If you want the full journey, it’s all here.

OMG, Google Swapped Your AI Image Model and You Didn't Notice

Did you generate an image in Gemini this week?

Then you used a different model than last week. And Google didn't exactly send a memo.

On February 26th, Google quietly replaced Nano Banana Pro (the model behind all those viral AI product images) with a brand new model called Nano Banana 2. No popup. No "hey, we changed things." Just... a swap.

If your images looked a little different and you couldn't figure out why, now you know.

Here's what actually changed, what it means for your product photography workflow, and how to switch back to Pro if you want to.

First, Let's Untangle the Version Soup

Google's naming makes this confusing. So let me lay it out clearly.

There have been three Nano Banana models. Each one is built on a different version of Google's Gemini AI. Here's the timeline:

Model	Gemini Version	Tier	Launch Date
Nano Banana (original)	Gemini 2.5 Flash Image	Flash (light & fast)	August 2025
Nano Banana Pro	Gemini 3.0 Pro Image	Pro (heavy & precise)	November 2025
Nano Banana 2	Gemini 3.1 Flash Image	Flash (light & fast)	February 2026

Notice the pattern?

The original Nano Banana went viral in August. Built on Gemini 2.5 Flash, it was fast and fun, but rough around the edges. Then in November, Google released Nano Banana Pro on Gemini 3.0 Pro. Slower, more expensive, but with studio-level quality, better text rendering, and real-world knowledge.

That was the model people fell in love with. The one generating those jaw-dropping product shots.

So Did Google Downgrade Everyone?

Here's the part that has people confused. And a little frustrated.

On February 26th, Google replaced Nano Banana Pro (Gemini 3.0 Pro) with Nano Banana 2 (Gemini 3.1 Flash) as the default.

The version number went up, from 3.0 to 3.1. But the tier went down, from Pro to Flash.

It's not a version downgrade. It's a tier swap. Google moved everyone from the heavy, premium Pro architecture to the lighter, faster Flash architecture. The new Flash is better than the old Flash. But it's still Flash, not Pro.

Think of it like going from a BMW 3 Series to a newer Toyota Camry. The Camry is newer, more efficient, and honestly great for daily driving. But it's not a BMW.

The Silent Downgrade Nobody Asked For

And it gets worse for free users. Even before Nano Banana 2 launched, people on the Gemini community forums were reporting silent downgrades. When free users hit their daily image quota, Google would quietly drop them from Nano Banana Pro to the original Nano Banana. No notification, no warning.

Resolution would tank. Text rendering would degrade. And users had no idea why their images suddenly looked worse.

Now with Nano Banana 2 as the default, free users don't even get the option to use Pro. It's reserved for Google AI Pro and Ultra subscribers only.

Is Nano Banana Pro Going Away?

With all the changes happening, I had to check. Good news: Nano Banana Pro is not being shut down. Google's official deprecation page shows no sunset date for the image generation model.

So why the switcheroo? I think this is a cost play, not a quality play. Nano Banana Pro runs on the heavy Gemini 3.0 Pro architecture. It's expensive to operate at scale. Giving hundreds of millions of free users unlimited access to your most expensive model isn't sustainable. So Google did what makes sense: they moved the masses to the cheaper, faster Flash model and reserved Pro for paying customers.

For most casual users generating memes and profile pics? Nano Banana 2 is more than enough. But for eCommerce sellers who need that extra finesse (accurate product details, precise text on packaging, photorealistic lifestyle shots) Pro still matters. And now it's behind a paywall.

What Is Nano Banana 2, Really?

So where does this leave us? Three models, simplified:

🔶 Nano Banana (original). The OG. Gemini 2.5 Flash. Fast, fun, viral. But limited quality. Think of this as V1.

🔶 Nano Banana Pro. The premium upgrade. Gemini 3.0 Pro. Slower, pricier, but the best image quality Google offered. The craftsman.

🔶 Nano Banana 2. The new default. Gemini 3.1 Flash. Takes the best features from Pro (text rendering, world knowledge, sharper details) and runs them on the faster Flash architecture. The workhorse.

Google decided the workhorse was good enough for most people. So they made it the default and tucked Pro behind a menu.

The Numbers Don't Lie

Here's the side-by-side breakdown:

	Nano Banana Pro	Nano Banana 2
Speed	~30 seconds	~10-15 seconds
Cost per image	~$0.06	~$0.03
Quality	Best available	~95% of Pro
Text rendering	Good	Better (readable signs, labels)
Content filters	More relaxed	Stricter
Architecture	Gemini Pro (heavy)	Gemini Flash (light)
Best for	Precision, fine detail	Volume, speed, production

Half the price. 2-3x faster. 95% of the quality.

For most eCommerce workflows? That math is hard to argue with.

I Tested It. Here's What I See.

I ran the same simple product image through both models (a Focallure Face compact) to see the difference with my own eyes.

Make a model hold this product in her hand

Original Photo

👇 This is what I got with Nano Banana 2:

With Nano Banana 2

👇 And this is what I got with Pro:

Nano Banana Pro

The difference? Nano Banana 2 still has that telltale AI look. You know the one. An extra finger sneaking into frame, skin that's just a little too smooth. Pro? It looks like a real photo.

Pro gives you slightly cleaner edges, a touch more realism in the skin texture, better placement and a bit more polish overall. It's the difference between a really good photo and a really good photo

Nano Banana 2 is warm, clean, and production-ready. For a product listing? For a social media ad? For 90% of what we do as sellers? You'd never notice the gap.

But here's the thing. If you're doing high-end brand photography or need absolute pixel-level precision, Pro still has the edge.

How to Access Each Model

This is the part most people are missing.

Nano Banana 2 (Default. You're Already Using It)

Just generate an image in Gemini. That's it. Nano Banana 2 is now the default across:

🔶 Gemini App: select "Create images" from tools, works in Fast, Thinking, or Pro mode
🔶 Google AI Studio: via the Gemini API (paid key required)
🔶 Google Search: through AI Mode and Lens
🔶 Vertex AI: for enterprise deployment

Nano Banana Pro (Hidden, But Still There)

Here's the trick nobody's talking about:

Generate your image normally in Gemini (it uses Nano Banana 2)
Click the three-dot menu (⋮) on the generated image
Select "Regenerate with Nano Banana Pro"

That's it. Pro is one click away, but only if you know it's there.

Important: You need a Google AI Pro or Ultra subscription to access Nano Banana Pro. Free tier users only get Nano Banana 2.

The "So What" for Sellers

Let's do some quick math.

Say you're generating 50 product images a week for listings, A+ content, and social. That's a realistic number for an active seller.

🔶 With Pro: 50 images x $0.06 = $3.00/week, plus ~25 minutes of generation time
🔶 With Nano Banana 2: 50 images x $0.03 = $1.50/week, plus ~10 minutes of generation time

Over a month that's $6 in savings and an hour of your life back. Not earth-shattering for small volumes. But scale that to 500 images for a catalog refresh or a seasonal push? Now you're looking at $15 saved and 2+ hours recovered per batch.

Speed matters more than cost here. When you're iterating (trying 10 different angles, tweaking prompts, A/B testing lifestyle shots) a 10-second generation time versus 30 seconds changes how you work. You stay in flow. You experiment more. You get to "good enough" faster.

That's the real advantage.

The One Catch

Content filters.

Nano Banana 2 is noticeably stricter than Pro. Google tightened the safety rails around:

🔶 Famous figures and celebrities
🔶 Character face/outfit swapping
🔶 Anything remotely suggestive
🔶 Financial information in images

For most product photography, this won't matter. But if you're doing creative lifestyle content, brand storytelling, or anything that pushes boundaries, you might hit walls that Pro wouldn't throw up.

Also worth noting: character consistency across multiple images is still a weakness. If you need the same model holding your product across five different scenes, both models will struggle, but Pro handles it slightly better.

Pro or 2? My Take.

Here's how I'd think about it:

Use Nano Banana 2 (default) when:

You're generating product-on-white or simple lifestyle images
You're iterating quickly and need volume
You're running A/B tests on creative
Budget matters (it always does)
You need readable text on packaging, labels, or signage

Switch to Pro when:

You're creating hero images for brand storefronts
You need the absolute best photorealism
Content filters are blocking your creative direction
You're doing a small batch where time doesn't matter but quality does

For 90% of what eCommerce sellers do every day? Nano Banana 2 is the move. It's fast, it's cheap, and the quality is close enough that your customers will never know the difference.

Pro is your secret weapon for the shots that really matter.

The Bigger Picture

Here's what I keep thinking about.

Six months ago, Nano Banana Pro was the cutting edge. People were blown away. It topped the image generation leaderboards.

Now it's the backup option behind a three-dot menu.

That's how fast this moves. The premium model of yesterday becomes the free tier of tomorrow. The tool you spent weeks mastering gets replaced by something faster and cheaper before you've even optimized your workflow.

This is why the learning window matters. Not because any single tool will last forever, but because the skill of adapting to new tools is the real advantage. The sellers who tested Pro last month are already comfortable with 2 this month. The ones who waited are still figuring out what changed.

Nano Banana 2 isn't perfect. Pro isn't dead, yet. But the direction is clear: faster, cheaper, and good enough is eating slower, expensive, and slightly better.

For eCommerce sellers, that's not a threat. That's an opportunity.

Claude Remote Control 🤖 The Safe (But Limited) Alternative to OpenClaw

If you've been following the OpenClaw hype, you know the promise: an AI agent that runs on your computer 24/7, browses the web, manages your calendar, controls smart home devices, and basically acts like a digital assistant on steroids.

Sounds amazing. Until you hear that a recent security audit found over 820 malicious skills in OpenClaw's community hub. That's not a typo. Over 20% of available skills were flagged.

So when Anthropic dropped Claude Remote Control on February 25th, a lot of people (myself included) got excited. It's not the same thing as OpenClaw (more on that in a second), but it scratches a similar itch: controlling an AI agent from your phone. And it's way safer.

What It Actually Does

Claude Remote Control lets you run a Claude Code session on your computer and then control it from your phone, tablet, or any web browser. Think of it like screen-sharing, but for your AI coding assistant.

Here's the setup:

Open your terminal and run claude remote-control
A session URL appears. Press spacebar to show a QR code
Scan the QR code with the Claude app on your phone (or paste the URL in any browser)
You're now controlling your local Claude session from anywhere

Already inside a Claude Code session? Just type /rc to start Remote Control without leaving.

Want it on by default? Type /config inside Claude Code and set "Enable Remote Control for all sessions" to true. After that, every session you start is automatically available from your phone. No extra commands needed.

Pro Tip: Before going remote, type /rename to give your session a descriptive name (like "product page redesign"). It makes finding the right session on your phone way easier.

Your files, your tools, your project configuration. All still running locally on your machine. Nothing moves to the cloud. The phone is just a window into your local session.

Why It's Safer Than OpenClaw

🔶 No inbound ports. Your machine only makes outbound HTTPS requests. Nobody can connect toyour computer
🔶 Everything runs locally. Your code, your files, your data stay on your machine
🔶 No community skill marketplace full of unvetted (and potentially malicious) plugins
🔶 Short-lived credentials. Sessions are secured via TLS through Anthropic's API
🔶 You stay in control. It doesn't run unattended. You initiate, you guide, you approve

But Here's Where It Falls Short

Let's be honest about the limitations:

🔶 Not a 24/7 agent. Your terminal has to stay open. Computer goes to sleep? Session dies.
🔶 One session at a time. No running multiple remote sessions in parallel
🔶 10-minute timeout.Lose network for more than 10 minutes and you have to restart
🔶 No browser automation. It can't browse the web, check flights, or manage your email like OpenClaw can
🔶 No persistent memory.OpenClaw saves context to local files and can reference conversations from weeks ago. Remote Control starts fresh each session
🔶 Expensive. Requires a Max plan ($100-$200/month). The official docs say Pro support is "coming soon," but it's not there yet. Not available on Team or Enterprise plans either

So Who Is This For?

If you're a seller who also vibe codes (and I know a lot of you are), this is genuinely useful. Kick off a coding task on your laptop, then check on it from your phone while you're at the warehouse or on a call. Review code, approve changes, monitor progress.

It's not OpenClaw. It's not trying to be. OpenClaw wants to be your autonomous digital employee. Remote Control wants to be a really long extension cord for your existing Claude Code workflow.

My take: If you're already paying for Claude Max and using Claude Code, this is a no-brainer. If you're not in that world yet, this alone isn't worth the $100/month. But keep it on your radar. Anthropic is building toward something bigger and safer here.

Have you tried Remote Control or OpenClaw? I'm curious which camp you're in. Hit reply and let me know.Have you tried Remote Control or OpenClaw? I'm curious which camp you're in. Hit reply and let me know.

The Vertex AI Back Door to Nano Banana Pro 🤓

Now this section is for the Nerds. Remember how I mentioned above that Google moved Nano Banana Pro behind a paywall? That free users are stuck with Nano Banana 2, and Pro now requires a Google AI Pro or Ultra subscription?

Well, there's a back door. And it's sitting right there in plain sight.

Vertex AI: The Access Nobody Talks About

Most sellers interact with Nano Banana through the Gemini app. That's the front door. Limited daily generations, resolution caps, silent downgrades when you hit your quota.

But Vertex AI is Google's enterprise AI platform. And it gives you direct, unrestricted access to both Nano Banana Pro and Nano Banana 2. No three-dot menus. No praying your daily quota hasn't expired. Just the model you want, when you want it.

Here's the part that gets interesting:

🔶 Google Cloud gives new accounts $300 in free trial credits. At roughly $0.13 per 2K image, that's over 2,200 Nano Banana Pro images before you spend a dollar
🔶 Google AI Studio (a separate tool) gives free users 50 API requests per day with just a Google account. No credit card needed
🔶 On paid Vertex AI tiers, there are no daily caps. Generate as many images as your budget allows

Why This Matters for Sellers

While everyone else is fighting over 2-3 free images per day in the Gemini app, you could be generating hundreds through the API with trial credits. No subscription. No waitlists. Just direct access to the model you actually want.

This is especially powerful if you're doing catalog refreshes, seasonal creative pushes, or A/B testing across dozens of product variations. The Gemini app was built for casual users. Vertex AI was built for production.

How to Get Started

Sign up at cloud.google.com (you'll need a credit card for verification, but you won't be charged during the trial)
Enable the Vertex AI API in your Google Cloud console
You now have $300 in credits and full access to Nano Banana Pro

That's it. Three steps and you've bypassed every limitation the free tier throws at you.

Pro Tip: If you're not comfortable with APIs yet, start with Google AI Studio (aistudio.google.com). It gives you a visual interface to test prompts with 50 free requests per day. No coding required. Think of it as the middle ground between the Gemini app and full Vertex AI.

Go With The Flow Podcast: Cowork Deep Dive - Skills, Guardrails, and Avoiding Disaster

In this episode Danny McMillan and I dive deep into Anthropic's new Cowork feature in Claude Desktop. The good, the bad, and the ugly.

I share a cautionary tale about file deletion gone wrong (yes, it happened to me), and Danny demos his custom skills pack that protects users from common pitfalls.

What we cover:

🔶 What Cowork is and how it differs from Claude Code and Claude Desktop
🔶 Why rm -rf can permanently delete your files (and how to prevent it)
🔶 How to set up deny lists in your Claude settings to protect critical files
🔶 The power of skills and bootstrap files for consistent, reliable outputs
🔶 2026 predictions: Ambient AI and pixel-free interfaces

Have a listen!

Convert More Clicks Summit
RSVP here: https://maximizingecommerce.com/CMC26-Ritu-Java
It’s FREE!

Super excited to be speaking at the Billion Dollar Seller Summit BDSS 13, and AI-focused event hosted by Kevin King. RSVP: https://www.bdss13.com

🚀 Never Pay for Another App! Learn How to Build Custom Features Yourself📍 Event: Seller Summit 2026
📅 Dates: April 21–23, 2026
Save your spot here: Link

Attio is the AI CRM for modern teams.

Connect your email and calendar, and Attio instantly builds your CRM. Every contact, every company, every conversation, all organized in one place.

Then Ask Attio anything:

Prep for meetings in seconds with full context from across your business
Know what’s happening across your entire pipeline instantly
Spot deals going sideways before they do

No more digging and no more data entry. Just answers.

Start your free trial →

🗞️ What is Prompt Injection?💥🚀

Ritu Java — Tue, 17 Feb 2026 16:07:13 +0000

A quick heads up. I’ve organized all previous editions into one searchable hub. If you want the full journey, it’s all here.

What is Prompt Injection?

You've probably heard the term floating around. Prompt injection. It sounds technical. It sounds like a developer problem.

It's not. If you're using AI in your business (and you should be), this is something you need to understand.

The simplest way to think about it:

Prompt injection is when someone hides instructions inside content that your AI reads, and the AI follows those hidden instructions instead of yours.

Remember writing secret messages with orange juice as a kid? The paper looked completely blank. But hold it over heat and the hidden message appeared.

Prompt injection works the same way. Attackers hide instructions inside normal-looking content (product reviews, emails, documents) that are invisible to you but get "revealed" when an AI reads them.

The AI doesn't know the difference between your instructions and theirs. It just follows what it reads.

I built a little Prompt Injection Detector simulation to show you what I mean. It works like holding invisible ink over a flame.

In the left pane you see a perfectly normal product review. Click the reveal button and the right side lights up with hidden instructions that were lurking inside the whole time.

That entire red block was hiding in the review, sandwiched between invisible Unicode characters. You couldn't see it. Your customer couldn't see it. But an AI reading that review? It sees everything, and it follows the instructions without question.

Here’s a example of a potential prompt injected message from a supplier. See what happens when we do the flame test:

That's bad enough with one AI. Now picture a whole team of them.

The industry is moving fast toward multi-agent systems. Multiple AI agents working together autonomously. One handles your ad bids, another writes your listing copy, another manages your customer service queue. The potential for e-commerce sellers is massive.

But prompt injection doesn't just scale with multi-agents. It compounds. Here are three risks worth knowing:

🔶 The Domino Problem. In a multi-agent setup, agents hand off tasks to each other. If one agent picks up a poisoned instruction (say, from a tampered product data feed), it doesn't stop there. It passes the bad instruction down the chain. Your inventory agent feeds bad data to your pricing agent, which feeds bad data to your ad agent. One contaminated input, the whole workflow is compromised.

🔶 The Backdoor Problem. Your agents don't all have the same power. One might only read reports. Another might send emails. Another might authorize payments. But if they share a common memory layer (and most multi-agent frameworks do), then compromising the weakest agent can give an attacker indirect influence over the most powerful one. It's like leaving your warehouse unlocked because "it's just storage," until someone uses it to access the office safe.

🔶 The Multiplier Problem. With a single AI, one prompt injection affects one system. With multi-agents, a single malicious instruction buried in an email or document can get picked up and re-executed by every agent in your network. What used to be a one-off trick becomes an automated chain reaction.

The good news? The big players are taking this seriously.

Sam Altman just brought the creator of OpenClaw, Peter Steinberger, into OpenAI specifically to tackle multi-agent coordination and security. That's a strong signal that these problems aren't being ignored.

But you don't need to wait for OpenAI to protect you. Some practical guardrails you can put in place today:

🔶 Treat every external input as untrusted. Product reviews, supplier emails, customer messages. Anything an AI agent reads from the outside world should be filtered before it enters your workflow.

🔶 Lock down agent permissions. Every agent should have access to only what it needs. Your analytics agent shouldn't be able to write to your payment system. Period.

🔶 Keep a human in the loop for high-stakes actions. Automated bidding? Fine. Automated refunds above a threshold? That needs a human approval step.

🔶 Audit the handoffs. The riskiest points in a multi-agent system are where one agent passes information to another. Know where those handoffs are. Log them. Review them.

The sellers who build with guardrails from day one won't need to wait for someone else to protect them. The learning window is open. Understanding this risk early is the competitive advantage.

Claude Code's Playground: Build Interactive Tools in Seconds

A new badass functionality in Claude opened up. It’s called “Plugins” and you can literally add on hundreds of Plugins that come with certain pre-trained skills. Today I am talking about a Plugin called “Playground”

To set it up, just access Plugins in the bottom left…

Then Browse available plugins…

Look for Playground and click Install:

Now you can just access this plugin from your Claude Code chat bar like so…

The Playground skill lets Claude generate self-contained interactive tools directly in your browser. Sliders, presets, live previews, all running locally in a single HTML file. No frameworks, no dependencies, no deployment. You describe what you want, Claude builds it, and it opens in your browser ready to use. Like this one I built (feel free to click and check it out):

That's an Amazon PPC Dashboard playground. KPI cards, budget alerts, hourly heatmaps, chart layouts. All configurable with the controls on the left, all updating in real time on the right. Built with one prompt. Now, remember that this is a playground, and each time you make an adjustment it will translate into a prompt that gets fed right back into Claude.

Five ways people are using Playground:

🔶 Visual dashboards. Turn your campaign data into interactive HTML reports you can actually explore instead of staring at spreadsheets.

🔶 UI prototyping. Need a landing page layout? Describe it, get a live preview with sliders for spacing, colors, and typography. Tweak until it's right, then send it back to Claude to build the real thing.

🔶 Product image prompt builders. Create a playground with controls for lighting, angle, background, and mood. See your prompt update live, copy it, paste it into your image generator.

🔶 Document review tools. Load any text file and get inline suggestions you can approve, reject, or comment on. Like a visual editor for your listing copy.

🔶 Data exploration. Got a complex search term report? Turn it into an interactive page where you can filter, sort, and spot patterns visually.

Here are Five prompts to try right now with Playground:

🔶 PPC Dashboard

Create a playground for an Amazon PPC Dashboard. Controls: 
date range selector, KPI card picker (ACoS, ROAS, impressions, clicks, CTR, CPC, CVR), chart layout toggle. Preview shows KPI cards with sparklines, budget alerts, ad spend vs sales chart,and hourly performance heatmap. Dark theme.

🔶 Product Landing Page Builder

Create a playground for a product landing page layout. Controls: hero image position (left, right, full-width), feature columns(2, 3, 4), color scheme presets (warm, cool, neutral), CTA button style, font pairing selector. Preview shows a live product page for a premium kitchen appliance brand.

🔶 AI Product Photo Prompt Builder

Create a playground for building AI product photography prompts. Controls: product type (bottle, box, pouch, jar), surface material, lighting style, camera angle, mood. Preview shows the assembled prompt updating live. Include presets: Clean Minimal, Luxury Editorial, Lifestyle Warm.

🔶 Listing Copy Reviewer

Create a playground for reviewing Amazon listing copy. Load a sample listing with title, bullets, and description. Show inline AI suggestions with approve, reject, comment buttons. Include readability score and keyword density meter in the sidebar.

🔶 Search Term Report Explorer

Create a playground for exploring Amazon search term data. Sample data with 50 terms including impressions, clicks, spend, sales, ACoS. Controls: sort by column, filter by ACoS threshold, spend range slider. Highlight high-spend zero-sale terms in red. Scatter plot of spend vs sales.

The OpenClaw Ecosystem

Geo from one of my Whatssap groups on AI recently shared this cool OpenClaw ecosystem and I thought it was worth passing along. Thanks Geo!

It’s crazy to think that less than a month ago, OpenClaw did not even exist!

Want to learn the science behind RUFUS?

Andrew Bell put together a neat resource called Generative Artificial Intelligence Model Streaming: An interactive educational guide to Amazon's patent on efficient AI response streaming with retrieval-augmented generation and intermediate token processing.

As you might now, Andrew eats patents for Breakfast, Lunch and Dinner. So you can expect this to be a solid resource if you want to satisfy your intellectual curiosities.

This is an interactive resource that let’s you learn at your own pace. You can even attempt a quiz at the end to test your understanding. Here’s the link to check it out.

When Kevin King puts up an event, you know it’s going to be stellar. I am SUPER excited to be speaking at the ECOM Mastery AI featuring Billion Dollar Sellers Summit in Nashville. When: April 8-12, Nashville. Get your tickets here.

🚀 Never Pay for Another App! Learn How to Build Custom Features Yourself📍 Event: Seller Summit 2026
📅 Dates: April 21–23, 2026
Save your spot here: Link

World’s First Safe AI-Native Browser

AI should work for you, not the other way around. Yet most AI tools still make you do the work first—explaining context, rewriting prompts, and starting over again and again.

Norton Neo is different. It is the world’s first safe AI-native browser, built to understand what you’re doing as you browse, search, and work—so you don’t lose value to endless prompting. You can prompt Neo when you want, but you don’t have to over-explain—Neo already has the context.

Why Neo is different

Context-aware AI that reduces prompting
Privacy and security built into the browser
Configurable memory — you control what’s remembered

As AI gets more powerful, Neo is built to make it useful, trustworthy, and friction-light.

Try Norton Neo Now

🗞️ When AI Writes Itself💥🚀

Ritu Java — Tue, 10 Feb 2026 16:07:14 +0000

A quick heads up. I’ve organized all previous editions into one searchable hub. If you want the full journey, it’s all here.

🗞️ AI is now Writing Itself

For the last two years, we treated AI like a very fast intern. Give it a task. Get a draft. Move on. Output was the end of the interaction.That model is already obsolete.

When machines started operating on their own instructions, something fundamental shifted. Anthropic recently said that Claude is now writing Claude code. Claude Cowork was written entirely by Claude.

When an agent distills documentation into a compact Skill and feeds it back into itself, we are no longer using AI. We are managing recursive systems.

This is the Recursive Feedback Loop. Output is no longer an endpoint. It becomes structured input for the next run. Every task leaves the system sharper than before.

From Context Rot to Context Compaction

For a while, we believed more context meant better results. Bigger prompts (pages and pages of it). More PDFs. Longer threads.

What we got instead was context rot. The model knew more, but understood less. Signal buried in noise.

The breakthrough was not bigger context windows. It was Context Compaction.

Instead of feeding chaos into models, operators force the model to distill what matters. The result is dense, reusable logic rather than raw information.

This is not summarization. It is compression with intent.

What “Skills” Actually Are

When Claude introduced the concept of “Skills” the idea was simple. Skills are the unit of compaction. As of today, Anthropic is not the only AI that supports skills. Gemini and Codex support it too. In practice, they are simple markdown files. Plain text. High density. Deliberate structure.

They live outside the model. In repos. In folders. In version control. They are loaded when needed, not rediscovered through prompting.

That is why Skills persist across sessions, agents, and models. They survive context limits and prompt churn. They allow intelligence to accumulate.

As a Claude user, you can add Skills under Capabilities in your Settings. Claude can even help you create Skills out of work that you are already doing.

If you work at the command line or with Claude Code, the folder structure can contain a Claude.md file that Claude will automatically look inside for any instructions you left it. No need to even say “Go to Claude.md” or restate context every time.

What can e-Commerce Operators Extract from this?

Every serious e-commerce team already has Skills. They just live in people’s heads, Slack threads, and forgotten Notion docs.

Start pulling them out.

Turn your best PPC audits into Skills. One markdown file that encodes how you typically diagnose wasted spend, when you pause, when you scale, and what signals you trust. And then turn that into a Skill.

Turn your brand voice into a Skill. One markdown file that encodes how your brand speaks, what it prioritizes, which claims it allows, which comparisons it avoids, and what “off brand” actually means. Load it once, and let every listing, ad, image, and response inherit it by default.

Turn image generation into a Skill. Encode your visual rules, composition preferences, lighting, backgrounds, do’s and don’ts, and what a “good” image looks like for your brand. Load it once, and let every creative inherit consistency instead of starting from scratch.

Store these Skills where your tools already look. Project folders. Repos. Campaign directories. Let your agents inherit them automatically instead of re-explaining them every time.

The goal is not more automation or better prompting. It is fewer resets.

If your AI needs you to re-explain how you run your business every week, you are still operating in 2024. The teams winning in 2026 are the ones turning experience into compact, persistent memory.

What Skills are you creating with AI? Hit reply and let us know!

Google Flow

A notable shift at Super Bowl LX is how AI became both the subject and the engine of ads. About 23 percent of commercials (15 of 66) referenced or featured AI in some way, with tech players and consumer brands alike pushing the narrative that AI was now part of mainstream life. That includes giants like OpenAI, Anthropic, Google Gemini, Amazon, Meta, Wix, and others buying premium Super Bowl airtime to frame their products through artificial intelligence messaging.

The most AI-heavy commercial of the night was Svedka’s “Shake Your Bots Off”. The vodka brand claimed it was the first primarily AI-generated Super Bowl spot, featuring its robot mascots animated using generative tools. That ad leaned fully into the technology as a creative statement even if reactions were mixed.

The Svedka spot mattered not because it was brilliant, but because it crossed a line. AI was no longer hiding behind polish or pretending to be invisible. It was the point. The tool became the headline. And once that happens on the Super Bowl stage, the question stops being whether brands should use AI and starts being whether they know how.

Of the many film making tools out there, Google Flow is a far choice that is aimed far above quick clips. This is text to cinema. You describe scenes, characters, tone, and pacing, and Flow generates short narrative films using Google’s Veo models. The output feels directed, not stitched together. Camera movement is deliberate. Lighting holds. Characters stay expressive instead of melting between frames.

Most Flow videos today live in the short film range. Think several seconds per scene, often extended into one to two minute narratives by chaining scenes together. The quality jump is what stands out. This is no longer novelty AI video. Depth, shadows, facial emotion, and cinematic framing hold up on large screens. You can still tell it is generated, but it no longer looks disposable.

One of the standout features is Ingredients. Ingredients let you define and reuse the building blocks of a story. Characters, locations, props, visual style, mood, lighting, even camera behavior can be saved and reused across projects. This solves the biggest pain in generative video which is drift. Your character stays your character. Your world stays intact. Your brand look does not reset every time you prompt.

Ingredients unlock something most product videos never touch. Story continuity. A character that appears across launches. A customer avatar who returns in different moments. A consistent visual universe that feels intentional instead of promotional.

Flow is not built for feature demos. It is built for narrative. Origin stories. But with some creativity and a bit of Canva magic, you can convert it into a decent video that can be displayed in your storefront.

PPC Ninja is helping brands future proof their listings for AI, helping you build RUFUS-enabled, stunning images and videos with AI. Hit reply on this to chat with us. Explore how we can scale your content production across Social media, Amazon ads, Amazon Posts efficiently and affordably.

Gemini Gems

I have always been a fan of Gemini Gems. They are closest thing Google has to CustomGPTs, but built to live where the work already happens. Over time, I have moved many of my CustomGPTs over to Gems, mainly because they stay inside the Google ecosystem instead of sitting beside it.

You will find a ton of premade Gems by Google that are ready to use out of the box. Or, you could create your own from scratch with your own special instructions to perform certain tasks consistently.

Gems are not floating chat personas. They can reference Google Drive files, work inside Workspace apps via Gemini, and increasingly plug into Notebook LM as a structured source of truth. Instead of pasting docs, copying links, or reminding the model how to behave, you design a Gem once and let it operate on curated knowledge. When that knowledge changes, you update the source, not the prompt.

One smart workflow I saw recently, shared by John Aspinall, pairs Gems with Notebook LM. Notebook LM holds the long term memory, things like PPC frameworks, CTR research, DSP notes, brand strategy docs. The Gem handles execution, turning that material into SOPs, drafts, or structured documents when needed.

Notebook LM becomes the library.
Gems become the operator.
Google handles the wiring.

🚀 Never Pay for Another App! Learn How to Build Custom Features Yourself📍 Event: Seller Summit 2026
📅 Dates: April 21–23, 2026

Save your spot here: Link

Tech moves fast, but you're still playing catch-up?

That's exactly why 100K+ engineers working at Google, Meta, and Apple read The Code twice a week.

Here's what you get:

Curated tech news that shapes your career - Filtered from thousands of sources so you know what's coming 6 months early.
Practical resources you can use immediately - Real tutorials and tools that solve actual engineering problems.
Research papers and insights decoded - We break down complex tech so you understand what matters.

All delivered twice a week in just 2 short emails.

Join 100K+ engineers

Still want to hear from us?

Ritu Java — Tue, 03 Feb 2026 22:56:33 +0000

Hey there,

I noticed you haven't opened our emails in a while, and I wanted to check in.

Every week, we break down the AI tools and strategies that are actually moving the needle for e-commerce brands — no hype, just practical insights you can use right away. You can read all our past posts here: https://www.ppc-ninja.com

Our readers tell us these updates help them:
• Discover AI tools before their competitors do
• Save hours on content, creative, and operations
• Cut through the noise and focus on what actually works

If that still sounds valuable to you, just click below to stay on the list:

Yes, keep me subscribed

No hard feelings if your priorities have shifted — I'll remove you from the list in a few days if I don't hear back.

Thanks for being part of the AI for E-Commerce community.

– Ritu