🔥🔥🔥 Three Highly Curated AI Stories

1. Claude just got a major upgrade with Opus 4.7

Anthropic released Claude Opus 4.7 this week, and it's noticeably better at the stuff people actually use Claude for: writing code, analyzing images, and handling complex multi-step tasks without losing the plot.

It claims to solve three times as many real-world work tasks as the previous version. On reading images and charts, it went from getting about half right to getting nearly all of them right.

That's a big deal if you've ever uploaded a screenshot and gotten a confidently wrong answer back.

Oh, and Anthropic casually mentioned their internal model is even better. We just can't use it yet. If you've been using ChatGPT or Gemini and haven't tried Claude lately, this is the version worth testing. Same price, meaningfully smarter.

2. Gemini starts remembering what you actually like

Google is quietly making Gemini more personal. With “Personal Intelligence,” the model now understands your preferences and interests when generating images. But first, if you’re wondering what personal intelligence means in artificial intelligence, watch this:

In practice, that means less prompting, less tweaking, and fewer “almost right” outputs. You don’t have to keep explaining your taste every single time.

This is a subtle shift, but an important one. The more context these systems retain about you, the more they move from tools to collaborators. Over time, this compounds into faster workflows and more consistent creative output.

3. HeyGen open-sources a new way to generate videos

HeyGen just released HyperFrames, and this one’s for anyone who’s tried (and struggled) to automate video production.

The idea is different from most tools. Instead of dragging elements into a timeline, you describe the video. Your coding agent generates HTML, CSS, and animations. HyperFrames then renders it into a finished video file.

The setup is surprisingly simple. One command and you’re in:
npx skills add heygen-com/hyperframes

It supports outputs like MP4, MOV, and WebM, and integrates directly with tools like Claude Code, Cursor, and Codex using slash commands. They also shared a super useful prompt guide for Claude Code, Cursor, Codex, and other AI agents to author Hyperframes compositions.

🔥🔥 Two AI Tools Worth Trying

1. 🎬 OpenArt + Seedance 2.0

If you’re creating video content for marketing, ads, or social, this combo is worth a try. Seedance 2.0 inside OpenArt can generate multi-shot videos with consistent characters and multiple camera angles from simple inputs.

You can upload reference audio, images, or clips, add a prompt, and generate full video assets without a production team. For small teams or solo creators, this cuts both time and cost dramatically.

2. 📂 Microsoft Copilot Notebooks

If your work lives inside documents, this is a quiet game-changer. With a Copilot license, you can now turn a collection of files into a structured “project hub” inside a notebook.

It automatically builds summaries, lets you ask questions across all documents, and can convert that information into slides, summaries, or even podcast-style outputs. Because it’s grounded in your own files, the outputs are far more reliable. You’re essentially working off your own data, which reduces the usual AI guesswork and hallucinations.

🔥 1 Pro AI Tip To Try Today

Build your own AI Chief of Staff

You don’t need to be the president to have a chief of staff. With the right setup, you can run your day with a lot more clarity (and yes, prove your mom wrong). If your time disappears into emails, meetings, and documents, this workflow helps you offload the busywork and focus on what actually moves things forward.

1. Connect your AI tool (we tested this exclusively using Claude) to your email, calendar, and document storage.
2. Use it to triage emails into priority buckets and draft replies.
3. Generate meeting briefs before calls using calendar context and past notes.
4.Summarize documents and extract key actions automatically.
5. Run a daily “check-in” prompt to plan your day based on everything above.

We break it down in a detailed guide with steps here.

TL;DR — Too Long Didn’t Read

• Understand what makes it different. Hermes is the only open-source agent with a closed learning loop — it creates reusable skills from completed tasks and remembers them forever. Most agents forget everything when a session ends. Hermes doesn't.

• Know what it isn't. It's not a consumer app. It runs on a server (a $5 VPS works), not your laptop. Windows users need WSL2. You'll be pasting commands into a terminal to set it up.

• Pick it if you have a recurring workflow. Daily reports, content automation, research digests, trading signals, weekly client summaries — any task you do the same way over and over is where Hermes compounds. The more you use it on one thing, the better it gets at that thing.

• Skip it if you want polish. It's on version 0.9.0. There are rough edges. If you want something that just works on your MacBook, you're not the target user yet.

• Budget realistically. The software is free. A VPS is ~$5/month. LLM API costs run roughly $10–30/month for light users on budget models (DeepSeek, Kimi, MiniMax) and $100+ for heavy users on premium models. Local models via Ollama are free but need a capable machine.

• The smart play is to use it alongside OpenClaw, not instead of it. They're built for different things. Hermes is the brain that learns. OpenClaw is the broader set of arms. Running both gives you the best of each.

1. What Is a Hermes Agent?

Hermes Agent is a free, open-source AI agent built by Nous Research — the same research lab behind the Hermes family of open-weight language models. It launched on February 25, 2026. By mid-April, the GitHub repo had passed 91,000 stars and 12,000 forks.

(Refer to this quick installation guide to get the Hermes Agent up and running in under two minutes)

Here's the clearest way to describe it: most "AI agents" you've used are chat wrappers. You type something, the AI replies, you close the tab, it forgets you existed. Hermes is a persistent process that runs on a server you control. It talks to you through messaging apps you already use — Telegram, Discord, Slack, WhatsApp, Signal, email, and about a dozen others. It remembers every conversation. And it creates new capabilities for itself as it works.

A few key facts to anchor this:

The tagline Nous Research uses is "the agent that grows with you." That's not marketing — it's a literal description of how the architecture works. Which brings us to the one feature that separates it from everything else.

2. What Is the Self-Learning Loop?

The self-learning loop is the single feature that separates Hermes from virtually every other personal AI agent.

Here's the one idea you need to understand. If you get this, the rest of Hermes makes sense.

Every AI agent you've used works roughly the same way: you give it a task, it reasons through the task, it produces an output, the session ends. Each new conversation starts from zero. The agent has no memory of what it's done before, what worked, or what didn't.

Hermes works differently. Every 10–15 actions it takes, it stops and asks itself four questions:

• What did I just do?

• Did it work?

• Is any of this worth remembering?

• Should I turn this into a reusable procedure?

If the answer to the fourth question is yes, Hermes writes a skill document — a plain markdown file describing how to do the thing it just figured out. Next time a similar task comes up, it loads that skill instead of solving the problem from scratch.

That's the loop.

The agent working with you in month three is meaningfully better at your specific work than the one you started with in month one, because it has built a personal library of procedures that work for you.

This isn't speculative. One Reddit user reported a 40% speedup on repeated research tasks after the agent had generated three skill documents over about two hours of use. Skills also self-improve during use — if Hermes finds a better way to do something mid-task, it updates the skill document automatically.

The critical distinction: most "persistent memory" in other AI tools remembers what you told it. Hermes learns from what it did. Those are different things, and the second one compounds..

3. The three types of memory that make it work

The learning loop needs somewhere to store what it learns. Hermes uses three layers, and each does something specific.

1. memory.md — the file that knows you

A single markdown file stored on your machine. It captures preferences, context, and key facts about you and your work. Things like: "the user prefers bullet points over prose in Slack updates," or "the client's reporting period ends on the 15th of each month." You never edit this manually — the agent maintains it itself.

2. Skill documents — procedural memory

When Hermes completes a complex task (typically five or more tool calls), it writes a skill document describing the procedure. These follow the open agentskills.io standard, which means they're portable to other agents that support the same format. Skills live on your disk and are searchable. The community Skills Hub has a growing library of pre-built ones you can install.

3. SQLite session database — searchable history

This is the layer most competing agents don't have. Every conversation, every tool call, every outcome is stored in a local SQLite database with full-text search. When Hermes encounters a situation it vaguely recognizes, it can search its own history to find how it handled something similar before. Combined with LLM-powered summarization, it can pull relevant context from months ago without burning through your current conversation's token budget.

All three layers work together. The markdown file captures who you are. Skills capture how to do things. The database captures what happened. Together they mean Hermes doesn't just have memory — it has context, procedures, and history, and it can use all three simultaneously.

4. What you can actually do with it (5 real use cases)

TL;DR — Too Long Didn’t Read

• The problem: Managers spend an estimated 28% of their workweek on email alone, and a significant chunk of the rest on meeting prep and admin work that AI can now handle.

• The solution: A small stack of AI tools — primarily Claude connected to your Google Workspace — that triages your inbox, preps your meetings, summarizes your docs, and organizes your day before you've had your first coffee.

• The core tools: Claude (with Google Workspace connectors), Reclaim.ai for calendar protection, and a meeting notetaker like Read.ai or Fireflies for post-meeting summaries.

• The workflow: Run a morning triage prompt. Check your meeting prep brief. Upload any docs that need reviewing. Let AI draft responses you approve. Protect your focus time.

• The honest caveat: AI handles the information load. The judgment, relationships, and decisions still sit with you.

1. What Is an AI Chief of Staff Workflow

An AI chief of staff workflow handles information processing tasks such as email triage, meeting preparation, document summarization, and daily planning before human decision-making.

Precisely this covers four key areas of work:

1. Email triage — scanning your inbox, categorizing by urgency, drafting responses, and surfacing only what actually needs your attention.

2. Meeting prep — pulling context from emails, documents, and past notes before a meeting so you walk in with the background you need.

3. Document summarization — compressing lengthy reports, proposals, and threads into the key points and actions, without you reading every word.

4. Day planning — translating your tasks, meetings, and priorities into a workable daily structure.

None of these require a dedicated AI assistant product. With Claude connected to your Google Workspace, most of this is available today through a standard subscription.

2. The AI Chief of Staff Tech Stack

Claude (with Google Workspace connectors)

Claude is the central tool in this stack. It connects directly to Gmail, Google Calendar, and Google Drive — meaning you can ask it to read your emails, check your calendar, and pull documents without leaving the conversation. This is what makes it function as a real chief of staff rather than just a writing assistant.

Claude's Google Workspace connectors are available on Claude.ai and Claude Desktop for all users. To connect them, go to your Claude settings, find the connectors section, and authenticate your Google account. Once connected, Claude can read emails, surface calendar events, and access files in Drive when you ask it to.

What Claude does well: Long-form summarization, drafting in your voice, pulling context across multiple documents or threads, and holding a lot of information in a single conversation.

Reclaim.ai

Reclaim connects to Google Calendar (or Outlook) and automatically blocks time for your priorities — deep work, habits, and tasks — working around your meetings. It defends those blocks when new meetings try to take over, and reschedules when your calendar shifts.

It's not an AI assistant in the conversational sense. It's more like a calendar optimizer that runs in the background. Think of it as protecting the time you free up by offloading email and prep to Claude.

Pricing: Free tier available. Pro starts at $8/user/month.

A Meeting Notetaker (Read.ai, Fireflies, or Fellow)

These tools join your Zoom, Google Meet, or Teams calls, transcribe the conversation, and generate summaries and action items afterward. They feed back into your chief of staff workflow — you can paste their summaries into Claude to draft follow-up emails or track open items.

Read.ai is particularly strong for managers because it tracks engagement patterns and surfaces insights across meetings over time.

Fireflies is solid for teams who want searchable transcripts and keyword tracking.

Fellow adds agenda collaboration and manager-specific features like one-on-one templates.

Optional: Superhuman (for email-heavy roles)

If email volume is your primary bottleneck, Superhuman is worth considering. It's an email client built for triage speed — keyboard shortcuts, split inbox, AI-drafted replies — and cuts the mechanical overhead of email processing. It doesn't replace Claude for analysis or summarization, but it makes the email interface itself faster.

Pricing: $30/month. Gmail and Outlook only.

3. How to Use AI for Email Triage

What to do

Connect Claude to Gmail using the Google Workspace connector. Then, instead of opening your inbox first thing in the morning and reading every subject line, start a conversation with Claude and run your triage prompt.

The triage prompt

This gives you a structured view of your inbox in under two minutes of reading, instead of thirty minutes of clicking.

What to do with the output

Review Claude's categorization. Correct anything it got wrong — over time, being specific in your prompt about what counts as noise vs. important trains you to give better instructions. Approve and send the drafted replies you're happy with. Add your edits to anything that needs a personal touch.

The key rule here: Claude drafts, you approve. Never set up any workflow where AI sends email on your behalf without your review. Beyond the obvious risk of errors, it removes the signal that a real person is behind the message.

Best practices

Be specific about what counts as urgent. Your triage prompt works better when you tell Claude who your most important contacts are, what projects are live right now, and what types of requests need same-day responses. You can paste this context at the start of a new conversation or save it in a Project in Claude.

Run triage twice a day, not continuously. Checking email constantly is the habit you're trying to break. A morning triage and an afternoon check is enough for most roles.

Use email aliases if your domain supports them. Routing newsletters to a separate alias (updates@yourdomain.com) makes the triage dramatically more accurate because the noise is already separated before Claude sees it.

What this looks like in time

The goal is to get from 60+ minutes of scattered email checking to a 10-15 minute structured triage session. The actual reading and responding gets compressed. The decision-making — which emails get real attention — stays with you.

4. How AI Prepares You for Meetings

What to do

Before any significant meeting, run a prep prompt in Claude that pulls context from your email and calendar. The output is a two-minute brief you read before joining.

The prep prompt

If you have a document related to the meeting — an agenda, a proposal, a project brief — upload it to Claude alongside this prompt and ask it to incorporate the file into the brief.

For recurring meetings

One-on-ones and weekly team standups benefit from a slightly different prep approach. Keep a running document in Google Drive or Claude Projects with notes from previous sessions. Before each meeting, ask Claude to pull that document and summarize what was discussed last time, what was committed to, and what's been resolved vs. still open.

This means you walk into a one-on-one actually knowing what you said you'd follow up on two weeks ago, without having to dig through notes.

Best practices

Run prep 15 minutes before the meeting, not the night before. Context is most useful when it's fresh. A brief you read at 9:45am for a 10am meeting is more useful than one you forget you read the night before.

Don't skip the document upload step. Claude's meeting prep gets meaningfully better when it has the actual agenda or background document to work from, not just email history. If there's a proposal on the table, upload it.

After the meeting, use your notetaker's summary as input. Once Read.ai or Fireflies gives you the transcript summary, paste it into Claude and ask: "Based on this meeting summary, draft a follow-up email to the attendees with the decisions made and action items. Keep it under 150 words." This closes the loop without you writing the recap from scratch.

5. How to Use AI for Document Summarization

What to do

For any document you need to engage with — a report, a proposal, a lengthy email thread, a contract — upload it to Claude and ask for a targeted summary. The key is being specific about what you need out of it.

The summarization prompt (better than just asking Claude to "summarize")

This is meaningfully better than asking Claude to "summarize" because it tells Claude what dimensions matter to you. A summary of a financial report looks different from a summary of a vendor proposal, even if they're the same length.

For email threads

Long email chains are one of the most time-consuming things managers deal with. Claude handles these well. Paste the thread into a conversation (or let Claude pull it via Gmail if it's recent) and ask:

For documents in Google Drive

If the document lives in Drive, you can ask Claude directly: "Find the document called [Title] in my Google Drive and summarize it using the framework above." This works with the Google Workspace connector active.

Best practices

Be specific about your role in the document. If you're a reviewer vs. a decision-maker vs. just being kept in the loop, your summary needs will be different. Tell Claude upfront: "I'm being asked to approve this proposal — tell me what I need to know to decide."

For high-stakes documents, read the sections Claude flags. Claude's job is to reduce what you have to read, not eliminate reading entirely. If Claude says "section 4 contains a clause you should review personally," read section 4.

Build a document library in Projects. Claude Projects let you upload reference materials that persist across conversations. If you're regularly working with a particular vendor, client, or internal team, upload the key documents once and reference them in future conversations without re-uploading.

6. How AI Organizes Your Daily Schedule

What to do

Start each morning with a planning prompt that synthesizes your calendar, outstanding email actions, and any tasks you're tracking. The output is a prioritized agenda for the day.

The planning prompt

What Reclaim.ai adds to this

While Claude gives you the plan, Reclaim.ai enforces it on your calendar. It automatically blocks focus time around your meetings, reschedules those blocks when meetings shift, and surfaces time for recurring habits like weekly reviews or project check-ins.

The combination works like this: Claude tells you what your priorities are and how to structure the day. Reclaim makes sure your calendar actually reflects that structure instead of defaulting to back-to-back meetings.

Best practices

Set your weekly priorities on Monday, not daily. Run a slightly longer planning prompt at the start of the week that sets your top three priorities. Then your daily planning prompt can reference those priorities rather than starting from scratch each morning.

Protect at least one two-hour block for deep work. Reclaim can do this automatically. If your calendar doesn't have this built in, AI planning loses much of its value — you'll have clarity on priorities but no time to act on them.

Use the end-of-day check-in. A quick prompt at 4:30pm — "What didn't get done today that I said I would? What needs to move to tomorrow, and what can I drop?" — takes three minutes and keeps your task list from becoming a guilt ledger.

7. A Sample Daily Workflow

Here's what a realistic AI chief of staff routine looks like for a manager:

7:30am — Morning triage (10 minutes) Open Claude, run the email triage prompt. Review the output. Approve any drafted replies. Note the Action Required items.

8:00am — Day planning (5 minutes) Run the planning prompt with your calendar and action items from triage. Reclaim has already blocked your focus time. Claude gives you the order of priorities.

9:45am — Meeting prep (3 minutes per meeting) Before each significant meeting, run the prep prompt. Upload the agenda if there is one.

Throughout the day — Document summaries as needed Any document that lands in your inbox goes to Claude before you read it. You read Claude's summary first, then decide whether to go deeper.

Post-meeting — Follow-up drafts (5 minutes) Paste your notetaker's summary into Claude. Get a draft follow-up email. Edit and send.

4:30pm — End of day check-in (3 minutes) Quick prompt on what's open, what moves to tomorrow, what drops.

Total active AI time: roughly 30 minutes spread across the day, replacing what was previously 2-3 hours of scattered processing.

8. Next Steps: Going Beyond the Basics

Once the core workflow is running, there are a few directions worth exploring.

Save your prompts as reusable templates. Claude Projects let you store instructions and context that persist across sessions. Build a project for your chief of staff workflow, save your triage prompt, planning prompt, and prep prompt there, and reference them daily without retyping.

Build a personal knowledge base. Upload your key reference documents to a Claude Project — org charts, project briefs, client context, strategic plans — so Claude can pull from them in any conversation without you re-explaining the background.

Automate recurring summaries with scheduled tasks. Claude Desktop's scheduled tasks feature lets you set up a recurring prompt that runs automatically. For example: every Monday at 7am, run a weekly review prompt that surfaces your top three priorities and flags any unresolved items from last week.

Add a CRM connector. For roles where client relationships are central, tools like Alfred or Carly extend the chief of staff workflow into CRM data — surfacing contact history, deal status, and follow-up items alongside your calendar and email context.

Start delegating research. Once you're comfortable with the core triage and prep workflows, extend into having Claude do preliminary research before important decisions — competitive landscape summaries, background on a new vendor, synthesis of a topic you need to brief your team on. The same summarization skills apply; it just takes a different kind of prompt.

9. What AI Cannot Do (Yet)

This workflow is practical and available today, but it has real limits worth naming.

It cannot read political context. Claude can tell you what was said in a meeting. It cannot tell you what it meant, who's frustrated, or where the interpersonal tension is. That reading still requires you.

It cannot make judgment calls. AI can flag that an email is urgent. It cannot judge whether the person sending it is reliable, whether the request is reasonable, or how to navigate the relationship. That's yours.

It cannot build trust for you. A follow-up email drafted by AI and sent without thought lands differently than one that reflects genuine engagement. People notice. Use AI to draft, but make sure the final message sounds like you and reflects what you actually think.

Accuracy is not guaranteed. Claude can misread tone, miss context, or categorize something incorrectly. The review step in every workflow above isn't optional — it's the safeguard that keeps this system trustworthy.

10. FAQs

Is it safe to connect Claude to my Gmail and Calendar? Claude's Google Workspace connectors use OAuth, the same authentication standard that other Google-connected apps use. Claude does not store your emails — it reads them within the conversation to respond to your prompt. For sensitive organizational email, check your company's policy on third-party AI tool connections before setting this up.

Do I need a paid Claude subscription for this? The Google Workspace connectors are available to all Claude users, including the free tier. However, the free tier has message limits. For daily professional use, Claude Pro ($20/month) is practical — it gives you priority access and higher limits.

What if I use Outlook instead of Gmail? Claude's Google Workspace connectors work with Gmail and Google Calendar. If you're on Outlook and Microsoft 365, Microsoft Copilot is the equivalent tool — it's natively integrated into Outlook, Teams, and the rest of the Microsoft suite. The workflow principles in this guide apply, but the specific setup differs.

Will AI make mistakes in my emails? Yes, sometimes. Which is why the rule is always draft, never auto-send. Every email Claude drafts should be reviewed before it goes out. Over time, you'll develop a sense for where Claude's drafts need editing (usually: making the tone feel more like you, adjusting anything nuanced) and where they can go out with minimal changes.

How long does it take to set this up? The Claude connector setup takes about five minutes. Writing your first triage and planning prompts takes another 15-20 minutes — most of that is thinking through what you actually want in the output. The workflow becomes habitual within a week.

Can I use this with a team? Yes. The principles work at the individual level, but teams can share prompt templates, run collective meeting prep (each person runs the same prep prompt for a shared agenda), and use Fellow or Read.ai to create shared meeting notes that Claude can then pull from. Claude's Team plan adds admin controls and shared projects.

 🔥🔥🔥 Three Highly Curated AI Stories

1. Starbucks is testing ChatGPT inside its app

Your next latte order might start with a conversation, not a menu. Starbucks is quietly testing a ChatGPT integration that lets you describe what you're in the mood for — "something warm but not too sweet, I didn't sleep well" — and get a drink recommendation back.

A sample prompt response in ChatGPT using Starbucks’ beta app. (Source: Starbucks)

This is interesting because it shifts AI from being a “tool you open” to something embedded inside everyday apps. You’re not going to ChatGPT anymore, it’s coming to you inside products you already use.

To be fair, most people ordering coffee already know what they want. This won't replace the regulars tapping their usual in three seconds. But for the long tail of "I don't know, surprise me" customers — that's a huge slice of revenue most brands currently lose to decision fatigue.

Watch which brand does this next. Once one major chain proves conversational discovery converts, the rest of the category has about eighteen months before it becomes table stakes.

2. Gemini finally gets a desktop app (and it’s actually useful)

 The browser just lost its monopoly on AI. Google just launched a standalone Gemini app for Mac this week, and with ChatGPT and Claude already there, we can officially call it: the AI assistant has moved out of the tab and onto the OS.

Here's why that small UX shift matters more than it sounds: every time you had to copy something out of your work, paste it into a browser, and paste the answer back, you were paying a tax. That tax made AI feel like a detour from your actual job. A desktop assistant with screen context collapses the detour into a keystroke. The work stays where the work lives.

The counterpoint is real — most people don't change their habits just because a keyboard shortcut exists. Plenty of users will keep pasting into chat.anthropic.com or chat.openai.com out of muscle memory. But the ceiling on what's possible just got higher, and power users will notice first.

3. Adobe brings AI assistants into creative workflows

Adobe is introducing a Firefly assistant powered by Anthropic’s Claude, bringing AI directly into its creative ecosystem.

Picture this: you open Premiere Pro for the first time, stare at the twelve-panel interface that has broken the spirit of a generation of creators, and instead of Googling "how to color grade a clip," you just ask. The assistant walks you through it inside the app — not a tutorial video, not a forum thread, an actual guide using your footage. The feature panel stops being a maze.

This isn’t just about generating images. It’s about helping users navigate complex workflows inside Adobe tools. Think assistance layered into the process, not separate from it. Now, the honest part: taste is harder to teach than software, and it's the thing that actually separates good work from generated-looking slop. The tools getting easier doesn't make more people good — it makes more people capable. Those are different.

🔥🔥 Two Things to Do With AI

1. 🧩🧠 Gemini in Gmail (Studio Automations)

If you’re on Google Workspace, there’s a feature hiding in plain sight inside Gmail called “Studio.” It lets you create simple automations or use ready-made templates for common workflows.

You can set up things like email summaries, meeting prep, or even keyword alerts. One useful example is a notifier that pings you whenever a specific word shows up in your inbox, like “feedback” or “urgent.”

It’s simple to set up and surprisingly powerful for something built right into Gmail. If your inbox feels chaotic, this is worth exploring.

2. 🤖 Copilot in Word (smarter document workflows)

Microsoft is expanding Copilot inside Word with new capabilities designed to improve how you work with documents.

Twitter tweet

The focus is on making writing and editing more dynamic. Instead of static documents, you get assistance throughout the process, helping you refine, restructure, and move faster. If you spend a lot of time in Word, this is another step toward AI becoming a built-in collaborator rather than a separate tool.

🔥 1 Pro AI Tip To Try Today

Train Claude to remember how you work (using .md files)

Most people use LLMs like a fresh start every time. New chat, new context, same instructions repeated again and again.

But there’s a better way. You can give LLMs like Claude a memory system using simple markdown files so it actually understands your preferences, your style, and your workflows.

Here’s how to set it up in Claude:

1. Create a file called “CLAUDE.md” and write clear instructions about how you want responses structured.
2. Add supporting files like “about-me.md” or “voice-and-style.md” with more context.
3. Upload or reference these files inside your Claude project.
4. Use them consistently so Claude starts adapting to your patterns.
5. Update the files over time as your needs evolve.

You stop repeating yourself and get more consistent, higher-quality outputs.

We break it down in a detailed guide with steps here.

 🔥🔥🔥 Three Highly Curated AI Stories

1. Audit your entire Meta Ads account in 5 minutes with Claude

A new Claude Cowork skill takes a Meta Ads CSV export and returns a full account audit, scored across six dimensions: creative health, audience efficiency, budget allocation, and three more. Here’s how:

• It calculates wasted spend in real dollars, flagging specific ads with spend and zero purchases

• The output is a fix list ranked by dollar impact, so you know which ad to kill first

• No API connection, no account access, no permissions dance — just a CSV upload and a prompt

Meta's native reporting buries the answer in 27 columns. This skill collapses a multi-hour weekly task into a single upload and beats most agency audit decks people pay four figures for.

2. Claude Code now runs multiple sessions in parallel

The Claude Code desktop app shipped a redesign with a sidebar that lets you run multiple sessions side by side from one window.

Key details:

• One session for debugging, one for planning, one for writing — all live, no context loss

• Each session keeps its own state, so switching tasks doesn't reset what you were doing

• The sidebar handles session management, killing the multi-window juggling act most heavy users were already hacking together

This is the update that turns Claude Code from a chat interface into a workspace. For anyone using it daily, it'll save more hours per week than any model bump on the roadmap.

3. Microsoft launches MAI-Image-2, built for cost-per-image

Microsoft's new image model, MAI-Image-2, is optimized for compute efficiency rather than benchmark wins.

Pay attention to:

• The pitch is comparable output quality at a fraction of the inference cost

• Aimed at production use cases — design ops, ad creative, e-commerce — where teams generate thousands of assets a week

• Slots into Microsoft's broader play to own the enterprise image stack via Copilot and Azure

The frontier image race used to be one game: who makes the prettiest pixel. Now it's two. Microsoft isn't trying to beat Midjourney on aesthetics; it's trying to make image gen cheap enough to embed in every business workflow.

🔥🔥 Two Pro AI Tips

🧩 Google Chrome AI Skills

Chrome now lets you save any Gemini prompt as a reusable Skill, triggered by / or a plus button. "Summarize this doc," "compare products across these tabs," "rewrite this email" — all one click on whatever page you're on. Repetitive prompts become shortcuts.

Thye’re also launching a library of ready-to-use Skills for common tasks and workflows:

☁️ Claude Dispatch

You can now trigger tasks on your computer from your phone via the Claude app. Send a command, your computer runs it, and you get a screenshot back when it's done. If it needs input mid-task, it pings you. Useful for kicking off long-running jobs and walking out the door. See how to set this up.

🔥 Things You Didn’t Know

Google Workspace has 7 features hiding in plain sight

You're already paying for them. You're just not using them.

Most people never discover the shortcuts that turn Docs into decision engines, Sheets into data pipelines, or Gmail into a productivity machine.

The people who do move faster.
They organize better.
They forget less.

These aren't buried in settings or hidden behind paywalls. We put together seven that pay back the most time. Each one takes under a minute to set up. Most people who try the first one end up running through all seven the same evening

TL;DR — Too Long Didn’t Read

• Gemini side panel lives inside Gmail, Docs, Sheets, Drive, and Slides — and it's not just a chatbot. It reads your open file and actually helps with it

• Google Meet's "Take Notes For Me" generates AI meeting notes and saves them to your Drive. You can stop pretending to type and listen to the meeting like a normal person

• Appointment Schedules in Google Calendar is a booking page you don't have to pay Calendly $16/month for

• Gmail's Nudges resurface important emails you forgot to reply to. It's embarrassingly useful

• Smart Chips in Google Docs turn an @ symbol into live-linked people, files, meetings, and dates

• Looker Studio connects directly to your Sheets and turns data into a shareable visual dashboard — for free

• Named Versions in Drive let you label and roll back to specific versions of any Doc or Sheet, so "final_FINAL_v3_USE_THIS" is no longer a file name

1. The Gemini Side Panel

There's a small sparkle icon sitting in the top-right corner of your Gmail, Google Docs, Sheets, Slides, and Drive. That's the Gemini side panel. It's been there for a while. Most people have never clicked it.

Here's why that's a mistake: unlike going to a separate AI tool and copy-pasting your content in, the Gemini side panel already has access to what you're looking at. Open a long email thread → click the panel → ask it to summarize. Open a Google Doc → ask it to rewrite a section. Open Sheets → ask it to explain what's happening in your data.

No switching tabs. No re-explaining context. It's already there and it already knows what's on your screen.

What it can actually do:

• Summarize long email threads in Gmail in one click

• Draft replies based on the conversation so far

• Generate a first draft in Docs from a short prompt — and pull context from files already in your Drive

• Answer questions about data in Sheets without you having to know any formulas

• Suggest changes to slide content in Presentations

Where to find it: Look for the sparkle (✦) icon on the right sidebar in Gmail, Docs, Sheets, Slides, or Drive. If you don't see it, check that Gemini features are enabled for your account under your Workspace admin settings. Business Starter plans and above have had Gemini included since January 2025.

The people who love this feature most are the ones who spend a lot of time in Gmail trying to catch up on long threads. One click to summarize a 22-message chain. That's it. That's the feature.

2. “Take Notes For Me" in Google Meet

We've all been in a meeting, nominally taking notes, but mostly re-reading the same sentence three times while three other people started talking. Google Meet has a feature that removes you from that situation entirely.

"Take Notes For Me" uses Gemini to generate a structured summary of your meeting — key discussion points, decisions, and action items — and saves it automatically to your Google Drive. It creates a Google Doc linked to the calendar event, so you can find it later without a scavenger hunt.

There's also a separate Transcript feature that captures a word-for-word record of the call and saves it to Drive too.

How to enable it:

1. Join a Google Meet call

2. Click the pencil/notes icon in the meeting controls (or the three-dot menu → "Take notes with Gemini")

3. A Doc is automatically created and saved to the meeting organizer's Drive

What plan you need: The AI-generated notes feature ("Take Notes For Me") requires a Google Workspace plan with Gemini included — Business Standard or higher. Transcripts are available on Business Starter and above. Check your admin panel if you're not sure what you have.

One thing worth knowing: The notes aren't perfect. If your call has heavy crosstalk, jargon, or a lot of "wait, can you go back to that slide," the summary will reflect that chaos. But even a 70% accurate automated summary is more useful than a blank Doc you swore you'd fill in after the meeting.

The best use for this: recurring team syncs, client calls where you're doing most of the talking, and any meeting where you already know you won't remember the details by 4pm.

3. Google Calendar Appointment Schedules

Quick question: are you paying for Calendly? Or Acuity? Or any scheduling tool that generates a booking link so people can find time with you?

If yes — and you have a Google Workspace paid plan — you've been double-paying.

Appointment Schedules in Google Calendar generates a booking page with your real availability, lets people self-select a time, and automatically creates calendar events for both parties. It syncs with your calendar in real time, respects your working hours, and sends confirmation emails without you touching anything.

How to set it up:

1. Open Google Calendar

2. Click the + button to create a new event

3. Select Appointment schedule (not "Appointment slots" — that's the older, simpler version)

4. Set your duration, availability windows, and buffer times

5. Share the generated booking link

You get a public URL you can drop in your email signature, LinkedIn, or wherever. People click, they see your open slots, they book. Done.

What makes this different from the old "Appointment Slots": The new Appointment Schedules feature (launched in 2022, still massively underused) creates a dedicated booking page, supports multiple duration options, and works across time zones automatically. The old slots were clunkier and required people to have a Google account to book.

What it can't do that Calendly can: complex routing, round-robin team scheduling, and some of the more advanced CRM integrations. If you're running a sales team at volume, Calendly still has an edge. For individuals and small teams managing their own time? Google's version is more than enough.

4. Gmail Nudges

You know that email you were going to reply to, then got distracted, and now it's buried under 47 newer messages and you've completely forgotten about it? Gmail has been quietly solving this problem for years. Almost no one has the feature turned on.

Nudges are automatic reminders that surface old emails back to the top of your inbox when Gmail detects you haven't replied and probably should have. They look like a small banner at the top of the email: "You may want to follow up on this" or "Sent 3 days ago, no reply yet."

It's not nagging. It's a nudge. The distinction matters when you're managing a full inbox.

How to turn it on:

1. Open Gmail → Settings (the gear icon) → See all settings

2. Under the General tab, scroll to Nudges

3. Check both options: "Suggest emails to reply to" and "Suggest emails to follow up on"

4. Save changes

That's genuinely it. Gmail now watches your inbox behavior and surfaces threads you've been ignoring.

Bonus feature while you're in there: Scheduled Send. Most people know you can schedule emails in Gmail, but few people use it deliberately. The sweet spot: write emails when you're in flow (even if it's 11pm or Sunday morning), schedule them to send at a normal business hour. Nobody needs to know you were drafting at midnight. Your inbox management improves, your response rate improves, and you stop sending emails that start with "sorry for the late reply."

To schedule: compose an email → click the dropdown arrow next to the Send button → Schedule send → pick a time.

5. Smart Chips in Google Docs

Type @ in a Google Doc and you'll see a dropdown appear. Most people type a colleague's name, tag them, and move on. That's the surface level. Here's what's actually there.

Smart Chips let you embed live, interactive references directly in the body of a document:

• @People chips — hover over someone's name and see their job title, email, and location. Click it to start a chat or email without leaving the Doc

• @File chips — reference a Drive file inline. Anyone reading the Doc can see the file name, type, and last modified date without navigating away

• @Date chips — add a formatted date that you can click to see the day of the week, add to Calendar, or use as a countdown

• @Meeting chips — embed a Calendar event so the Doc and the meeting are connected. Useful for meeting prep docs and agendas

• @Calendar draft building block — draft a Calendar invite directly inside a Doc, then push it to Calendar when ready

Why this matters: The default workflow is: write a doc, go find the file you're referencing, copy the link, paste it in. Smart chips collapse that. The document becomes a single source of truth where everything is connected rather than a flat text file with a bunch of links at the bottom.

Where it gets genuinely useful: Project briefs, meeting agendas, proposals, onboarding docs — anything where multiple people, files, and dates need to be connected to the same document. Teams that use this consistently say their docs require significantly less "wait, which version?" back-and-forth.

To use: open any Google Doc → type @ → start typing a person's name, file name, or choose a building block from the dropdown.

6. Looker Studio + Google Sheets

Google Sheets is where data goes to become a problem someone else has to interpret. Looker Studio is where that data goes to become something people can actually read.

Looker Studio (formerly Google Data Studio) is a free visualization tool from Google that connects directly to your Sheets, and turns rows and columns into charts, graphs, and shareable dashboards that update automatically whenever the underlying Sheet changes.

No exporting. No rebuilding charts every Monday. No "I'll just screenshot the table" emails. The dashboard lives online, pulls from your live Sheet, and stays current without you touching it.

How to set it up (this actually takes about 10 minutes):

1. Go to lookerstudio.google.com

2. Click Create → Report

3. Select Google Sheets as your data source

4. Choose the Sheet and tab you want to visualize

5. Drag and drop chart types — bar charts, scorecards, line graphs, pie charts — into the report

Once it's built, you share a link and anyone with access sees a live, interactive dashboard. You can filter by date, click into segments, and see the same data in different ways.

Best use cases:

• Weekly team metrics or KPI reports you currently build manually

• Client-facing reporting that needs to look polished

• Sales or pipeline tracking from a Sheets-based tracker

• Any dashboard your team checks regularly and currently screenshotted from a spreadsheet

What it costs: The base version of Looker Studio is free and connects to Google Sheets, Google Analytics, Google Ads, and more. Looker Studio Pro (for additional collaboration features) is a paid upgrade, but for most small and mid-size teams, the free version is completely sufficient.

If you have a Sheet that someone on your team opens every week to "check the numbers," turning it into a Looker Studio dashboard is a 20-minute project that will save that person from ever doing that manually again.

7. Named Versions in Drive

Google Docs, Sheets, and Slides automatically save a version history of every change you make. That part most people know. What most people don't know: you can name those versions, and you can roll back to any of them with two clicks.

This is the difference between having a vague timeline of edits and having a clear, labeled archive of your document at specific moments — before client feedback, after the first draft, before the big pivot, after legal reviewed it.

How to name a version:

1. Open any Google Doc, Sheet, or Slide

2. Go to File → Version history → Name current version

3. Give it a clear label: "Before client revisions — April 10" or "Board-approved final"

That version now appears as a named milestone in your history, separate from the noise of every minor edit.

How to roll back:

1. File → Version history → See version history

2. The right panel shows your full edit timeline

3. Click any version to preview it

4. Click Restore this version if you want to go back

Why this is quietly one of the most useful things in Workspace: Anyone who has ever:

• Sent the wrong version to a client

• Had someone else edit a document and lost their original work

• Needed to retrieve a section they deleted three weeks ago

...will immediately understand the value. Version history has always existed in Drive. Named versions make it usable.

One extra detail: For Google Workspace paid plans, version history is kept indefinitely (compared to 30 days for Drive files that aren't Docs/Sheets/Slides). Naming versions means you're not just hoping you can find the right timestamp later — you have labeled checkpoints.

TL;DR — Too Long Didn’t Read

• A .md file is just a plain text file with simple formatting symbols — no special software required to create one

• Markdown files matter in Claude because they're the format Claude reads most cleanly when you upload content or write instructions

• There are three places .md files show up in Claude: Project knowledge bases, Project instructions, and CLAUDE.md (for Claude Code users)

• You can create a .md file using Notepad (Windows), TextEdit (Mac), VS Code, or any text editor — just save it with .md at the end

• The biggest mistake people make is writing long, vague instructions — shorter and more specific always wins

• If you use Claude Projects, a well-written .md file in your knowledge base can cut down on context-setting by a significant margin

1. What Is a Markdown (.md) File?

A markdown file is a plain text document that uses simple symbols to add structure and formatting. The .md extension tells whatever is reading the file, "this is markdown — please render it."

Here's what that means in practice. Instead of clicking a Bold button, you wrap text in double asterisks: **like this**. Instead of applying a heading style, you add a # before the line. Instead of a styled bullet list, you just add a - or * before each item.

That's it. There's no proprietary format, no file size bloat, no software lock-in. It's just text — which is exactly why it became popular among developers, writers, and now AI tools.

A quick example of markdown syntax:

# This becomes a heading

**This text is bold**

- This is a bullet point
- So is this

This is just a regular paragraph.

When rendered (like on a website or in Claude), that looks like a proper formatted document. When stored, it's just a lightweight .txt-style file with a .md extension.

Markdown was created in 2004 by John Gruber and Aaron Swartz with one goal: make text that looks good as plain text and can be converted to formatted HTML without friction. It's now used across GitHub, Notion, Reddit, documentation tools, and — relevant to you — Claude.

2. Why .md Files Matter in Claude?

Claude is trained on vast amounts of text written in markdown. Headers, bullet points, bold text, numbered lists — Claude doesn't just display markdown, it thinks in it. When you give Claude structured markdown input, it picks up on the hierarchy and relationships between pieces of information more reliably than with a block of unformatted text.

This matters in two directions:

When Claude reads your files: If you upload a .md file to Claude's knowledge base, Claude parses the structure — headings, subheadings, lists — to understand how your content is organized. That makes it easier for Claude to find and reference the right section when you ask a question.

When Claude writes for you: Claude defaults to markdown in most responses. Those headers and bullet points you see in Claude's answers? That's markdown being rendered by the interface. Understanding this helps you know how to ask Claude to format things differently, and why copying its output into other tools sometimes looks cluttered with symbols.

For anyone using Claude Projects regularly, .md files are the most practical format for uploading reference material, brand guidelines, SOPs, or any document you want Claude to actually use — not just acknowledge.

3. The Three Places Markdown Shows Up in Claude

1. Claude Projects — Knowledge Base

Claude Projects is a feature that lets you create persistent workspaces with their own file storage, instructions, and chat history. Free users can create up to five projects; paid subscribers get more.

When you upload files to a project's knowledge base, Claude reads them at the start of every conversation in that project. This means you don't have to paste your brand guidelines, client context, or research notes every time you start a new chat.

Markdown files work particularly well here because:

• Claude can navigate headers and subheadings to find the right section quickly

• The clean structure means less "noise" for Claude to filter out compared to a PDF with embedded formatting or a Word doc with tracked changes

• You can update the file and re-upload it without reformatting anything

Practical example: A freelance consultant creates a client-brief.md for each client. It contains the company overview, key stakeholders, tone guidelines, and current project goals. Every time they start a Claude chat in that project, Claude already knows the context.

2. Claude Projects — Custom Instructions

Every Claude Project has an instructions field where you set how Claude should behave inside that project. Think of it as a persistent system prompt that applies to every conversation.

You write these instructions in plain text — and markdown formatting makes them more effective. Using headers and bullets helps Claude parse what applies when, especially if your instructions cover multiple scenarios or roles.

Example instructions (written in markdown-style formatting):

You are assisting a marketing team at a B2B SaaS company.

**Tone:** Professional but conversational. No jargon.

**Always do:**
- Reference our target audience (mid-market operations teams)
- Keep subject lines under 50 characters
- Ask clarifying questions before writing first drafts

**Never do:**
- Use superlatives like "best-in-class" or "world-leading"
- Write in first person unless drafting a CEO message

This isn't a file you upload — it's text you type directly into the instructions box. But it works best when written with markdown structure in mind.

3. CLAUDE.md — For Claude Code Users

If you use Claude Code (Anthropic's command-line tool for developers), there's a special file called CLAUDE.md that Claude reads automatically every session. It lives in your project's root folder and acts as a permanent briefing document.

CLAUDE.md is specifically relevant to technical workflows: it holds things like your codebase structure, preferred libraries, testing commands, and coding conventions. Claude Code reads it before every session so you don't have to re-explain how your project is organized.

If you're not using Claude Code, this doesn't apply to your workflow. It's worth knowing exists, but non-developers won't encounter it in day-to-day Claude use.

4. How to Create a .md File?(No Coding Required)

Creating a markdown file takes about 60 seconds. The only technical step is saving the file with .md instead of .txt.

Option 1: Use Notepad (Windows)

1. Open Notepad (search for it in the Start menu)

2. Write your content using simple markdown symbols (see below)

3. Go to File → Save As

4. In the "Save as type" dropdown, select All Files

5. Name your file something like brand-guidelines.md

6. Click Save

Option 2: Use TextEdit (Mac)

1. Open TextEdit

2. Go to Format → Make Plain Text (this is important — TextEdit defaults to rich text)

3. Write your content

4. Go to File → Save

5. Name it with .md at the end, e.g., project-notes.md

Option 3: Use VS Code (Any Operating System)

VS Code is a free text editor that also shows a live preview of your markdown as you write. It's the most comfortable option if you'll be creating .md files regularly.

1. Download VS Code free at code.visualstudio.com

2. Open the app, create a new file, and save it with the .md extension

3. Press Cmd+Shift+V (Mac) or Ctrl+Shift+V (Windows) to open the markdown preview panel

Option 4: Use an Online Markdown Editor

If you don't want to install anything, sites like Dillinger.io or StackEdit.io let you write markdown in a browser and export it as a .md file. Good for occasional use.

Basic Markdown Syntax Reference

You don't need to memorize all of this. For most Claude use cases, headings (#, ##) and bullet points (-) are all you need.

What to Actually Put in Your .md File?

This depends on how you're using Claude. Here are the most common use cases and what to include for each.

If you're uploading to a Claude Project knowledge base:

Brand or style guide:

# Brand Voice

## Tone
Professional, direct, and warm. We avoid corporate speak.

## Words we use
- Clear, practical, human

## Words we avoid
- Synergize, leverage, paradigm shift

Client or project brief:

# [Client Name] Project Brief

## About the company
[2-3 sentence description]

## Project goal
[What we're working toward]

## Key contacts
- [Name, Role]
- [Name, Role]

## Things to keep in mind
- [Constraint or preference]

Research or reference doc:

# Research Notes: [Topic]

## Key findings
- [Finding 1]
- [Finding 2]

## Sources
- [Source name, URL if applicable]

If you're writing Project Instructions:

• Start with who Claude is being in this project (role/context)

• Add your "always do" and "never do" rules

• Include format preferences (length, structure, tone)

• Keep it under 500 words — longer isn't better

5. Best Practices That Make a Difference

• Be specific, not vague. "Write professionally" tells Claude very little. "Use short paragraphs, avoid jargon, and write at a Grade 10 reading level" gives Claude something to work with.

• Use headers to organize sections. Claude navigates your document using its structure. If your 3,000-word brand guide is one continuous block of text, Claude has a harder time finding what's relevant. Break it into labeled sections.

• Keep each file focused. One file per topic works better than one massive file covering everything. A tone-of-voice.md, a product-faq.md, and a target-audience.md are easier to update and easier for Claude to navigate than a single everything.md.

• Update your files when things change. A stale brand guide or an outdated project brief will produce outdated output. Make it a habit to keep your .md files current — they only work as well as the information inside them.

• Don't include sensitive information. This applies especially to CLAUDE.md files and any documents you upload to Claude: avoid API keys, passwords, private client data, or any information you wouldn't want stored in a third-party system.

• Test it after uploading. After uploading a .md file to a Claude Project, start a new chat in that project and ask Claude a question that should be answered by that file. If Claude's response is off, your file may need clearer headings or more specific content.

6. Common Mistakes and How to Fix Them

Mistake: Writing instructions that are too long Claude's context window is finite. A 5,000-word CLAUDE.md or project instructions file doesn't give Claude more guidance — it dilutes it. Stick to what Claude genuinely needs to know for the task at hand.

Fix: Aim for under 500 words in instructions. If you need more, split into separate focused files.

Mistake: Uploading a PDF or Word doc when a .md file would work better PDFs with images, complex formatting, or scanned text are harder for Claude to parse accurately. Claude reads plain text best.

Fix: Convert important reference documents to markdown. Copy the key content into a .md file with proper headings.

Mistake: Using markdown files as a dump for everything People sometimes upload entire websites, extensive research archives, or sprawling notes hoping Claude will just "figure it out." Claude performs better with curated, relevant content.

Fix: Only include what's directly relevant to the work you'll be doing in that project.

Mistake: Forgetting to update files after things change Brand guidelines evolve. Projects shift scope. If your .md file still references last year's strategy, Claude will work from outdated context.

Fix: Review your project files quarterly or whenever there's a meaningful change in scope.

Mistake: Treating markdown syntax as optional An unstructured .md file (just paragraphs with no headers or bullets) loses the primary advantage of markdown — structured navigation.

Fix: Always use at least ## headers to break your document into sections, even for short files.

7. FAQs

Do I need to know how to code to use .md files? No. Markdown is not a programming language. If you can write a bulleted list in an email, you can write a .md file. The only "technical" step is saving the file with a .md extension.

Can I use a Word document instead of a .md file in Claude Projects? Yes, Claude Projects accepts multiple file formats including PDF, DOCX, TXT, CSV, and HTML. However, markdown files tend to be parsed more cleanly because they're plain text with explicit structure. For anything text-heavy and reference-focused, .md is the better choice.

How big can a .md file be before it affects Claude's performance? Claude's context window has limits. Larger files consume more context, leaving less room for the actual conversation. As a practical guideline, keep individual .md files under 10,000 words and use multiple focused files rather than one large one. Paid Claude plans (Pro, Max, Team, Enterprise) have access to extended context through Retrieval Augmented Generation (RAG), which can handle larger knowledge bases more effectively.

Does Claude automatically read .md files I upload, or do I need to reference them? In Claude Projects, files you upload to the knowledge base are available to Claude in every chat within that project — you don't need to reference them explicitly. Claude draws on them as context throughout the conversation.

What's the difference between project instructions and an uploaded .md file? Project instructions are a persistent prompt — they tell Claude how to behave. Uploaded files tell Claude what to know. Think of instructions as the rulebook and uploaded files as the reference library.

Is CLAUDE.md the same as Claude Project files? No. CLAUDE.md is specific to Claude Code (the developer command-line tool) and lives inside your local coding project folder. Claude Project files are uploaded through the Claude web or desktop interface and are separate from any local file system.

Can I share a .md file across multiple Claude Projects? Not natively — you'd need to upload it manually to each project. If you have a central reference document (like company-wide brand guidelines), consider keeping a master version and uploading copies to each relevant project. When you update the master, re-upload to each project.

 🔥🔥🔥 Three Highly Curated AI Stories

1. Copilot is killing the "figure it out" step

Microsoft's Copilot push isn't really about better outputs. It's about deleting the steps you'd normally take to get to them.

You used to open a tool, think through the process, execute. Now you describe the outcome and Copilot handles the in-between.

Sounds incremental. It's not. Because once the steps disappear, the skill shifts — from knowing how to do something to knowing what you want done. That's a very different kind of leverage. And side bet — whoever explains Copilot clearly on short-form first is going to print attention. The gap is wide open.

2. Meta's AI clones are a bet on behavior, not tech

The headline is flashy: AI versions of you that talk, respond, exist online while you don't.

The interesting part isn't the tech. It's the assumption underneath — that people want to be replicated. That outsourcing parts of your presence feels fine.

I'm not convinced. This isn't scheduling posts or auto-replying to DMs. It's identity-adjacent. It changes how people experience you. Power users will lean in (they always do). But for most people, this crosses an invisible line we didn't know we had. Watch the opt-out rate, not the sign-ups.

3. Lovable is attacking the exact spot where people quit

Payments setup is the task everyone postpones. Not because it's hard. Because it's annoying. Lovable is trying to erase that friction.

Twitter tweet

Describe what you want to sell, and it handles Stripe, Paddle, Shopify, taxes, currencies — across 200+ countries.

But the real shift is psychological. You're not "setting up payments." You're continuing the same conversation you started the project in. That's how you get more things across the finish line. Not better tools. Less friction at the part where everyone usually taps out.

🔥🔥 Two Pro AI Tips

1. 🍌 Prompting Gemini like a Director (Nano Banana)

Most image prompts fail for a simple reason. They’re vague. Gemini’s Nano Banana approach fixes that by treating prompts like a creative brief, not a search query. At the core, you define five things clearly:

Subject (what’s in the image), Composition (framing), Action (what’s happening), Location (where it’s set), and Style (the aesthetic).

Even a basic upgrade like:

will give you far more consistent results. This was my first attempt:

From there, you can layer in control:

• Aspect ratio & framing (like 9:16 for reels)
• Lighting & camera details (golden hour, soft shadows)
• Text placement (exact wording + style)
• Direct edits (“remove the car in the background”)
• Factual constraints (for diagrams or historical accuracy)

And if you’re using reference images, assign roles clearly. One for pose, one for style, one for background.

2. ⚡Tab Tab Tab (Google AI Studio)

Start typing something half-formed in AI Studio, hit tab, Gemini completes the thought.

Sounds minor. It's not. Most friction isn't in doing the work — it's in starting it. This kills the pause where you're trying to phrase things perfectly. You just begin, and the system meets you halfway. Try it →

🔥 Things You Didn’t Know You Could Do With AI

Use Claude Code to skip the "learning phase" entirely

There's a version of this where you spend three weeks watching tutorials on how to build a thing. And there's a version where you just... build the thing.

Claude Code is the second one.

Here's the play:

1. Get a paid Claude plan — Claude Code is included.

2. Open the desktop app so it can actually read and write your files.

3. Give it one clear goal. Mine: "Create a clean portfolio website with projects, experience, and testimonials."

4. Let it generate the structure, files, and initial logic. Don't hover.

5. Refine by asking for changes ("swap the font," "add a contact form") instead of starting from scratch.

Before: weeks of setup, syntax, and half-finished side projects rotting in a folder called /new.

After: a working starting point in minutes — something you can actually keep building on tonight.

For a more thorough look at how you can set up Claude Code, read our complete walkthrough here.

TL;DR — Too Long Didn’t Read

• Claude Code is a version of Claude that takes action on your computer — it creates files, runs commands, and builds things, rather than just giving you text suggestions.

• You need a paid Claude subscription (Pro at $20/month works fine to start).

• There are four ways to use it: the Desktop App (easiest), inside VS Code or Cursor (most popular), in the terminal, or directly in a browser with no install at all.

• A file called CLAUDE.md is how you give Claude persistent instructions. Without it, every session starts from scratch.

• Plan Mode, Ask Mode, and Code Mode control how much Claude does on its own versus asking your permission first.

• Git is your undo button. Set it up before you start anything serious.

1. What Is Claude Code?

When you use Claude at claude.ai, it gives you text. Good text, usually. But it doesn't touch your files, doesn't install anything, doesn't create folders, and doesn't run programs. It talks.

Claude Code is different. It acts.

Open it in a folder on your computer and it reads everything in that folder — your files, your structure, your code. Then when you give it an instruction, it writes files, edits them, runs commands, installs packages, and produces working output. Not suggestions. Actual output.

Here's what that looks like in practice: You type "build me a personal portfolio website with a dark theme." Thirty seconds later, there's an HTML file sitting in your folder that you can open in a browser. You didn't write a line of code.

A few other things that make Claude Code meaningfully different from the browser version:

• It understands your whole project. Drop it into any folder and it immediately understands the context of what's there — the file structure, the tech stack, the naming conventions.

• It runs in your environment. When Claude Code executes a command, it runs on your machine. It can install software, use your existing tools, and access local files.

• It connects to your other tools. Through something called MCP (Model Context Protocol), Claude Code can read from Notion, write to Google Sheets, create GitHub pull requests, send Slack messages, and more — from a single prompt.

2. What You Need Before You Start

A paid Claude subscription

Claude Code doesn't work on the free plan. Your options:

• Pro ($20/month): Enough to learn and do real work. Usage limits apply but most beginners won't hit them regularly.

• Max ($100–$200/month): For heavy daily use where you don't want to get cut off mid-project.

• Team/Enterprise: For organizations with multiple people using it.

• API key via Anthropic Console: Pay-per-use pricing, which suits developers who want programmatic access.

Sign up at claude.ai if you don't have an account, then go to Settings → Billing to upgrade.

A compatible computer

• macOS 13.0 or newer (Intel and Apple Silicon)

• Windows 10 or newer (Windows also requires Git for Windows — more on this below)

• Linux — Ubuntu 20.04+, Debian 10+, and most major distribution

3. How to Install Claude Code?

Option 1: The Desktop App (Best Starting Point)

If you've never used a terminal and want to skip all of that, start here. The Desktop App gives you a visual interface — no command lines required.

How to install:

1. Go to claude.com/download and download the installer for your operating system (macOS or Windows).

2. Run the installer like you would any normal application.

3. Open the app and sign in with your Claude account.

4. Click the Code tab at the top of the app.

5. Select a folder on your computer where you want to work, or create a new empty folder.

That's it. You're in.

The Desktop App shows you every file Claude creates or changes, lets you run multiple sessions side by side, includes a visual diff view so you can see exactly what changed, and even supports scheduled tasks that run on a timer.

One honest note from people who've used both: the Desktop App is a great way to start, but if you want to do serious recurring work, learning to use Claude Code inside an editor like VS Code will serve you better over time.

Official guide: code.claude.com/docs/en/desktop-quickstart

Here’s a quick demo:

Option 2: VS Code or Cursor (Most Recommended for Regular Use)

VS Code is a free code editor from Microsoft that has become the standard development environment for millions of people — including many who aren't professional developers. Think of it as a more powerful version of a text editor, where you can see all your files in a sidebar and Claude Code works alongside them.

Cursor is a fork of VS Code with extra AI features built in. Either one works.

Step 1: Download VS Code or Cursor

• VS Code: code.visualstudio.com

• Cursor: cursor.com

Install it like any normal application.

Step 2: Install the Claude Code extension

1. Open VS Code or Cursor.

2. Click the Extensions icon in the left sidebar (four small squares).

3. Search for "Claude Code."

4. Click Install on the official extension by Anthropic.

5. After installing, an orange Claude icon appears in the top-right corner of the editor.

Step 3: Open a project folder

Go to File → Open Folder, navigate to the folder where you want to work (or create a new empty folder), and open it. This tells Claude Code where to operate.

Step 4: Launch Claude Code

Click the orange Claude icon. A panel opens on the side, and you can start typing prompts immediately. Every file Claude creates or changes will be visible in the sidebar on the left.

Why this setup matters: being able to see the files while Claude works gives you visibility and confidence. Even if you never plan to read code yourself, knowing what's being created keeps you in control.

Official guide: code.claude.com/docs/en/vs-code

Option 3: The Terminal (Full Control, More Setup)

The terminal gives you access to every Claude Code feature and is the most powerful way to use it. If this feels unfamiliar, start with Option 1 or 2 and come back.

On Mac or Linux:

Open your Terminal app (on Mac, press Cmd + Space and search "Terminal") and paste this:

curl -fsSL https://claude.ai/install.sh | bash

On Windows (PowerShell):

Open PowerShell and paste:

irm https://claude.ai/install.ps1 | iex

On Windows (Command Prompt):

curl -fsSL https://claude.ai/install.cmd -o install.cmd && install.cmd && del install.cmd

Alternative install methods:

• Homebrew (Mac): brew install --cask claude-code

• WinGet (Windows): winget install Anthropic.ClaudeCode

After installing:

Type claude --version to confirm it worked. Then type claude and press Enter. On first launch, it will ask you to sign in through your browser. Follow the prompts.

If something goes wrong at any point, type claude doctor — this command diagnoses your installation and tells you what needs fixing.

Official setup guide: code.claude.com/docs/en/setup

Option 4: Use It in the Browser (Zero Install)

If you don't want to install anything, go to claude.ai/code and start working. Your code runs on Anthropic's cloud infrastructure, which means you can start long-running tasks and check back later, work from any device including a tablet or phone, and run multiple parallel tasks without using your own computer's resources.

The trade-off: you don't have direct access to files on your local machine. For exploring Claude Code, creating things from scratch, or working on cloud-hosted repositories, it's a solid option. For work that requires access to files already on your computer, you'll want one of the other options.

4. What to Do When You Open Claude Code?

No matter which option you chose, the first thing to do is understand what you're working with.

Trust the folder: When Claude Code opens in a directory for the first time, it asks you to "trust this folder." This is a safety check. Confirm it. Claude Code needs permission to read and write files in that folder.

Start by asking what it can do: Literally type this:

What can you do?

Claude will explain its own capabilities. Get in the habit of asking any AI tool to explain itself at the start — it orients you and sometimes reveals features you didn't know existed.

Understand how permissions work: By default, Claude Code asks your permission before doing anything that changes your system — writing a file, running a command, installing a package. You'll see prompts like "Allow write to index.html?" with Yes/No options. This is Ask Mode. It's slow but safe, and it's where you should start.

As you get more comfortable with a task, you can switch to auto-approval or set specific permissions. More on modes below.

CLAUDE.md: How to Give Claude a Persistent Memory

Here's something that trips up almost every new user: Claude Code starts fresh every single session. It doesn't remember what you told it last week, what rules you set, or what decisions you made. Every session, you're starting over.

The fix is a file called CLAUDE.md. Place it in the root of your project folder, and Claude Code reads it at the start of every session. Think of it as an instruction sheet for a new assistant who shows up every morning.

What to put in your CLAUDE.md

Keep it under 200 lines. Longer files cause Claude to start losing track of earlier instructions. Only include things Claude can't figure out by reading your project files:

• Your tech stack — what languages, frameworks, or tools the project uses

• Code style preferences — naming conventions, formatting rules

• Workflow rules — "always create a Git branch before making changes," "always run tests after editing code"

• Commands Claude needs — how to build, test, or run your project

• Things to avoid — specific patterns, libraries, or approaches you don't want

What not to put in your CLAUDE.md

Skip anything Claude can already figure out from your code. Skip standard conventions it already knows. Skip long explanations or tutorials. If Claude is already doing something correctly without the instruction, don't add it.

A simple CLAUDE.md that actually works:

# Code style
- Use ES modules (import/export) syntax, not CommonJS (require)
- Destructure imports when possible

# Workflow
- Run npm test after making code changes
- Commit with a descriptive message after each working milestone

# Stack
- React frontend, Node.js backend, PostgreSQL database

How to create one

Option 1: Type /init in any Claude Code session. Claude will analyze your project and generate a starter file based on what it finds.

Option 2: Ask Claude to create one: "Analyze this project and create a CLAUDE.md that covers the tech stack, project structure, how to run the project, and any conventions you can detect."

The second approach often produces more thorough results because it's a custom prompt rather than an automated scan.

Check your CLAUDE.md into Git so it travels with the project. Review it when Claude starts making mistakes. Prune it when it gets too long. Treat it like code.

Official memory guide: code.claude.com/docs/en/memory

5. Plan Mode, Ask Mode, and Code Mode: What Each One Does

Claude Code has three working modes. Understanding when to use each one changes how useful it is.

Plan Mode

Claude reads your files and creates a detailed plan — but writes nothing and changes nothing. Use this for anything non-trivial: new features, significant refactors, anything that touches multiple files. It forces Claude to think before acting and gives you a chance to catch problems before they happen.

Toggle Plan Mode with Shift + Tab in the terminal, or use the mode selector in VS Code or the Desktop App.

Here’s how this works:

Ask Mode (Default)

Claude does the work but pauses and asks permission before each change. You approve or reject file writes and command executions one at a time. This is the right mode when you're learning or doing something unfamiliar. It's slower, but you stay in control.

Code Mode (Auto Accept)

Claude writes and executes without stopping to ask. Fastest option, but you need to trust what it's doing. Only use this once you've seen how Claude handles a particular kind of task and you're confident in the direction. Having Git set up is essential here, because mistakes can happen quickly.

Recommended workflow for any non-trivial task: Start in Plan Mode. Review the plan. Switch to Ask Mode (or Code Mode if you trust the task) to execute. Commit the result with Git.

The Commands and Shortcuts That Actually Matter

Which model to use

Claude Code lets you switch models. The three main ones:

• Opus: The most capable. Use it for complex tasks, architecture decisions, anything requiring deep reasoning. Consumes more of your usage.

• Sonnet: The everyday model. Good for most tasks. Start here.

• Haiku: Fast and lightweight. Use it for quick questions and simple tasks.

Type /model to switch. For most people, Sonnet is the right default.

6. How to Connect Claude to Your Other Tools (MCP Servers and Plugins)

MCP (Model Context Protocol) is what allows Claude Code to interact with apps you already use. Without MCP, Claude Code only works with files on your computer. With MCP, you can ask it to check your Notion page, update a Google Sheet, create a GitHub issue, or send a Slack message — all from one prompt.

What you can connect

Some commonly used integrations:

• GitHub — create repositories, manage pull requests, review code

• Notion, Google Sheets, Airtable — read and write to documents and databases

• Slack — send messages and read channels

• Figma — pull design specs into your code

• Playwright — automate browser testing

• Context7 — pull up-to-date documentation so Claude doesn't use outdated code examples

How to add an MCP server from the terminal

claude mcp add <server-name> -- <command>

Add --scope user if you want the server available across all your projects, not just the current one.

You can also install MCP servers through the Desktop App by going to Customize → Browse Plugins.

Plugins: One-Click Bundles

Plugins are pre-packaged bundles of MCP servers, skills, and settings that install in one step. Type /plugin in Claude Code to browse the official marketplace.

Three worth knowing about:

• Playwright — browser automation and testing

• Superpowers — adds brainstorming, parallel agents, and debugging workflows

• Context7 — ensures Claude always references current documentation

Official MCP guide: code.claude.com/docs/en/mcp
Plugin marketplace: claude.com/plugins

7. Git: The Undo Button You Need From Day One

Every tutorial, every guide, every experienced Claude Code user says the same thing: set up Git before you do anything else.

Git is a version control system that takes snapshots of your project. If Claude makes a mistake — and it will sometimes — you can roll back to the last working state instead of trying to manually undo a series of changes you can't fully see.

How to get Git set up

If you don't have Git installed, you can ask Claude Code to install it for you. Type: "Can you install git for me?" and it will figure out the right commands for your system. Or download it directly from git-scm.com.

Connecting to GitHub

GitHub stores your code in the cloud, which means it's backed up and you can access it anywhere.

1. Install the GitHub CLI (gh) by downloading it from cli.github.com, or ask Claude Code to do it.

2. Run gh auth login and follow the browser prompts to authenticate.

3. Claude Code can now create repositories, push code, and manage branches on your behalf.

Commit early and often

Every time Claude produces something that works, commit it. Tell Claude to do this automatically by adding a rule to your CLAUDE.md: "Always commit changes to Git with a descriptive commit message after completing each task."

If something goes wrong after you commit, you can always get back to the last good state.

8. Three Prompts to Try When You First Open Claude Code

These are designed to show you what Claude Code can actually do, not just what it can talk about.

Prompt 1 — Build something immediately:

Create a beautiful personal portfolio website with the following sections: a hero section with my name and a short bio, an About section, a Projects section with 3 placeholder cards, and a Contact section with a simple form. Use modern design with a dark theme. Make it a single HTML file I can open in a browser.

You'll have a working website in about 60 seconds. This works because the request is specific about the output, the structure, and the constraints.

Here’s how we did it:

Prompt 2 — Automate a real task:

Research the latest news about [topic you care about] on the web. Write a LinkedIn post about it in a professional but conversational tone, under 200 words, ending with a thought-provoking question. Save it to a file called linkedin-draft.md.

Replace the topic with anything relevant to you. This shows Claude Code can search the web, write content, and save files — all from one prompt.

Prompt 3 — Plan before you build (switch to Plan Mode first with Shift+Tab):

I want to build a simple task manager web app where I can add tasks, mark them complete, and delete them. Data should be saved locally so it persists when I refresh the page. Ask me questions to clarify the requirements and design preferences before you create a plan.

Claude will interview you before doing anything. This "ask me questions first" technique produces significantly better results than vague starting prompts because it forces alignment before any work happens.

9. Key Resources

Everything you need is in Anthropic's official documentation. These are the pages worth bookmarking:

 🔥🔥🔥 Three Highly Curated AI Stories

1. Claude now works across Word, Excel, and PowerPoint

Claude can now share context across Word, Excel, and PowerPoint in a single conversation. So instead of jumping between files and re-explaining everything, Claude understands your entire working set.

This matters because most real workflows don’t live in one document. You’re analyzing numbers in Excel, writing summaries in Word, and presenting in PowerPoint. Claude now connects that loop.

The immediate win here is speed. The bigger shift is that AI is starting to operate at the “project level,” not just file level.

2. Claude Managed Agents go public beta

Anthropic just made a strong play for anyone building agents.

Claude Managed Agents gives you the infrastructure layer that usually takes months to build. You define tasks, tools, and guardrails. Claude handles the execution and scaling.

What’s interesting isn’t just the launch, it’s what it removes. You don’t need to worry about orchestration, deployment, or performance tuning as much anymore.

This is how we go from “cool agent demo” to actual production systems. Faster.

3. Netflix’s VOID removes objects from video (and fixes physics)

This one is… a little wild.

Netflix dropped VOID, a tool that removes objects or people from videos. But it doesn’t stop there. It adjusts lighting, shadows, and motion so the scene still makes sense.

That second part is the real story. Removing something is one thing. Making the world behave as if it was never there is another.

We’re getting very close to editing reality at a level that was previously locked behind high-end VFX teams.

🔥🔥 Two Pro AI Tips

1. 🧠🧾 Claude Tax Connectors (TurboTax + Aiwyn)

Claude now integrates directly with tax platforms like TurboTax and Aiwyn. Once connected, you can upload your documents and have Claude estimate your taxes, explain forms, and even draft returns.

This is especially useful if you’re behind on filings or just want clarity before talking to an accountant. Early feedback suggests it’s surprisingly reliable for first-pass estimates. Here’s how you can set it up.

2. 🧮 ChatGPT Instruments

ChatGPT recently introduced “Instruments,” which are interactive widgets triggered by your prompts.

Think weather cards, calculators, currency converters. Instead of just text answers, you get structured, visual outputs that update dynamically.

This is a direct move toward search-like experiences. Fewer hallucinations, more grounded answers. And honestly, a clearer signal that AI tools are coming for everyday search workflows.

🔥 Bookmark This

Turn raw footage into clean, edited videos using AI

There’s a guide we just published that shows how to use tools like Claude + Remotion to automate video editing.

You can identify the best cuts in long recordings, generate timestamps automatically, and stitch everything into a polished video without scrubbing timelines manually.

It also walks through how to turn rough clips into structured content like tutorials or social media videos. The kind of stuff that usually takes hours suddenly becomes… a workflow.

If you’re working with video at all, this is worth a look!

 🔥🔥🔥 Three Highly Curated AI Stories

1. Claude's new "advisor strategy" lets small models phone a friend

Anthropic rolled out something that's less of a feature and more of a philosophy shift. They're calling it the advisor strategy.

Here's the idea: you run a smaller, cheaper model like Sonnet or Haiku as your executor. When it hits a tricky decision mid-task, it taps Opus for guidance, gets an answer, and keeps going — all inside the same API call.

This is a big deal for anyone building with AI agents. You’re essentially getting near Opus-level intelligence without running Opus the entire time, which keeps costs lower while still improving decision quality

Right now, this is limited to the API, not the regular Claude chat interface. But if you’ve seen how multi-model orchestration is evolving, it’s easy to imagine this becoming standard soon.

2. Perplexity now connects to your bank accounts (yes, really)

Perplexity just added a Plaid integration that lets it connect to your bank accounts, credit cards, and loans.

Once connected, it can track your spending, build custom budgeting tools, and visualize your net worth alongside your investments in real time. On paper, this is incredibly useful. It turns AI into a personal finance dashboard that actually understands your data.

But here’s where it gets complicated. This launch comes at a time when Perplexity is already facing scrutiny around data privacy. So adding deeper financial access into the mix is… a bold move.

The capability is powerful. But the timing raises questions.

3. Gemini can now build interactive charts and 3D models in chat

Gemini just leveled up how it handles data visualization. You can now generate fully customizable interactive charts and even 3D models directly inside the chat.

This means you’re no longer stuck with static outputs. You can tweak, explore, and interact with the data in real time, which makes analysis far more intuitive.

For anyone working with reports, dashboards, or presentations, this closes a big gap. Instead of exporting data into separate tools, the analysis and visualization can now happen in one place.

It’s a quiet update, but one that changes how people actually use AI for thinking and not just writing.

🔥🔥 Two Pro AI Tips

1. 🧠 Claude power use cases (that most people miss)

Most people still use Claude for basic prompting. That’s barely scratching the surface. Here are three high-impact use cases:

• Reddit research workflow: pull real user discussions to find content ideas that already have demand
• Blader Humanizer skill: rewrite AI content to sound more natural and less robotic
• Frontend slide skills: generate clean, structured presentations directly from prompts

These are practical workflows you can use immediately, whether you’re creating content, researching, or building assets. Here are even more advanced use cases in the full walkthrough.

2. 📄 Adobe Student Spaces

Adobe introduced Student Spaces, which works a bit like NotebookLM but focused on academic workflows.

It helps students organize documents, understand material, and work through content in a more structured way. Think of it as a study assistant that can break down complex information and help you interact with your notes more effectively.

If you’re a student or doing heavy research work, this is worth exploring.

🔥 Bookmark This

The Claude glossary everyone should bookmark

You've been using Claude for months. You still write prompts the same way you did on day one. The fix isn't a fancier model — it's a better vocabulary.

A few terms worth knowing on sight:

• System prompt — the persistent instructions that shape Claude's behavior before your conversation even starts. Think of it as setting the stage, not the scene.

• Context window — the total amount of text Claude can "see" at once, including your prompt, uploaded files, and prior messages. Hit the ceiling and the earliest stuff starts slipping out of view.

• Tool use — when Claude calls external functions (search, code execution, APIs) to do things beyond generating text. It's the difference between an assistant that tells you and one that does.

• Artifacts — the separate window where Claude renders code, documents, or visuals you can edit and iterate on, instead of burying them in chat.

These are four of the big ones. The full glossary has the rest — skills, MCP, extended thinking, and more — and it's the kind of page worth pinning in a tab.

Open the glossary →

TL;DR — Too Long Didn’t Read

• Remotion is a React-based framework that renders video programmatically. You describe what you want, Claude writes the code, and Remotion renders it to MP4.

• The Remotion Agent Skill teaches Claude how to write Remotion code correctly. It's a knowledge file, not a plugin — you install it once and it changes how Claude generates video code.

• There are two distinct workflows: one for adding overlays and captions on top of a clean video, and one for AI-assisted segment cutting using audio transcription.

• Automated "video cleaning" — trimming filler words frame-precisely — is not possible with this workflow. AI processes video as sampled frames, not at millisecond precision. This is a hard technical limitation, not a gap that will be fixed with better prompting.

• What does work reliably: animated word-by-word captions synced to speech, segment cutting from manually-defined timestamps, and motion graphics like lower thirds, logo intros, and animated overlays.

• Your source video must be H.264 MP4. If you're recording on a Mac or iPhone, you're probably shooting HEVC or MOV — you'll need to convert first.

• Vague prompts produce generic results. The more specific you are about timing, font, color, and layout, the better the output.

• Everything renders locally. No cloud service. No per-minute fees.

1. What Is Remotion for AI Video Editing?

Traditional video editors — Premiere, Final Cut, DaVinci — work on a visual timeline. You drag clips, drag effects, and export. It's intuitive but manual. Every project is one-off. You can't version control it, you can't generate 50 variations from a template, and you certainly can't describe what you want in plain English and have it built for you.

Remotion approaches video differently. Instead of a timeline, it's code. Specifically, it's React — the same framework used to build web applications. You write components that describe what should appear on each frame, and Remotion renders those components into an MP4 file. Time is just a variable. Want something to fade in at second 3? You write that in code.

Here’s a brief overview from their own YouTube channel and this video was edited using Claude Code:

For most people, that sounds like a step backward. Learning React to edit a video? That's the wrong way to think about it. Because what it actually means is this: if you can describe a video in English, and Claude can write the React code for you, you've just bypassed the entire learning curve.

That's what the Remotion Agent Skill enables. It's a SKILL.md file — a knowledge package that teaches Claude how Remotion works: its APIs, animation patterns, common pitfalls, and best practices. When you install it, Claude stops guessing at Remotion syntax and starts writing code that actually works.

The Remotion MCP is a separate but related component. Where the agent skill gives Claude knowledge to write Remotion code, the Remotion MCP server gives Claude access to indexed Remotion documentation at runtime — so it can look up specific APIs and check technical details as it builds your composition. In practice, for most users the agent skill handles the majority of the work.

The preview environment runs at localhost:3003 in your browser. You describe a change, Claude updates the code, you reload the preview, and see the result — all without rendering. Rendering to a final MP4 only happens when you're satisfied.

2. What AI Video Editing Can and Cannot Do (With Remotion)

Being clear about this upfront will save you time and frustration.

What it does well

Animated captions synced to speech. Whisper transcribes your audio and Remotion renders word-by-word or line-by-line captions with custom fonts, colors, and animations — frame-perfectly synced. This is the single most polished use case in this entire workflow and it's well-documented and reliable.

Segment cutting from defined timestamps. When you (or an AI analyzing a transcript) provide a list of which segments to keep and which to cut as a JSON file, Remotion renders only those segments seamlessly. The output is a trimmed video with no raw editing software required.

Motion graphics and overlays. Lower thirds, animated titles, logo intros, progress bars, branded overlays — this is Remotion's strongest native capability. These are composited on top of your source video with full control over timing, style, and animation.

What it cannot do

Automatic filler word removal. This is the most common misconception about AI video editing, and it's worth explaining the exact reason. AI models process video by sampling frames at intervals — typically one frame per second. If your ideal cut point is at 00:02:15 but the AI only sampled the frame at 00:02:00 and 00:03:00, it cannot identify that precise moment. Frame-perfect cleaning requires external tooling that can operate at the actual frame level, not sampled approximations. This is not a limitation of Claude's intelligence — it's a fundamental constraint of how AI tokenization handles video.

Direct editing of raw, uncut footage. Remotion is an editing and animation layer that works on top of a clean source video. It doesn't analyze messy footage and produce a highlight reel automatically. Your source needs to be pre-cut before Remotion adds to it.

We tested this out for ourselves:

And here’s the finished output:

When we tested this, we noticed that AI isn’t always perfect at identifying exact cut points. That’s because it doesn’t process every single frame of a video — it looks at sampled moments — so you’ll often get close suggestions, but not frame-perfect edits.

3. AI Video Editing Setup (Two Workflows Explained)

There are two distinct workflows, and they have different requirements. Know which one you're doing before you start installing things.

Scenario A — Overlay & Animation: You have a clean, pre-cut video and you want to add captions, motion graphics, or effects on top of it.

Scenario B — AI-Assisted Cutting: You have raw or semi-raw footage and you want to use Whisper to transcribe the audio, then use Claude to identify which segments to keep, and render a trimmed output.

If you're not sure which to start with, start with Scenario A. It has fewer dependencies, faster iteration, and the results are immediately visible.

4. How to Add AI Captions and Motion Graphics with Remotion

Here’s a quick overview before the detailed step-by-step breakdown:

Step 1: Install Node.js and npm

If you don't have Node.js installed, download it from nodejs.org. The LTS version is fine. npm is included.

Verify in your terminal:

bash

node --version
npm --version

Step 2: Create a new Remotion project

Navigate to the folder where you want to work, then run:

bash

npx create-video@latest

This scaffolds a Remotion project with the starter template. When prompted, give the project a name. Don't worry about which template you choose — Claude is going to rewrite most of it anyway.

Step 3: Install the Remotion Agent Skill

The skill can be installed at the project level (only applies to this project) or globally (applies to all Claude Code sessions).

Project-level install (recommended to start):

bash

npx skills add remotion-dev/skills

This downloads the skill files into .claude/skills/ in your project directory. You should see a SKILL.md file and subdirectory files like rules/animations.md and rules/audio.md. These are the knowledge files Claude will reference.

Global install (if you plan to use Remotion across multiple projects):

bash

npx skills add remotion-dev/skills --global

To keep the skill current:

bash

npx remotion skills update

Step 4: Prepare your source video

This is where most first attempts fail. Remotion's video seeking depends on H.264 encoding. If your source file is HEVC, MOV, or anything else, convert it first:

bash

npx remotion ffmpeg -i source.mov -c:v libx264 -crf 23 output.mp4

Put the converted file in your Remotion project's public/ folder. This makes it available to Remotion as a static asset.

Step 5: Open Claude Code and describe what you want

Open your terminal in the Remotion project directory and start a Claude Code session. Reference the skill explicitly to make sure Claude activates it:

Using the remotion-best-practices skill, take the video in public/output.mp4 and add animated word-by-word captions in white Inter font, 36px, centered at the bottom of the frame, with the active word highlighted in yellow. Match the timing to the audio.

Claude will write or update your Remotion composition code. You'll see it modify files in src/.

Step 6: Preview in Remotion Studio

Start the preview server:

bash

npx remotion studio

Open your browser at localhost:3000 (Remotion Studio). You'll see your video composition and can scrub through it, test animations, and check caption sync before committing to a render.

Make adjustments by going back to Claude Code and describing what to change. This loop — describe, preview, adjust — is the core of the workflow. You're not rendering between every iteration.

Step 7: Render the final video

When you're satisfied with the preview:

bash

npx remotion render

A 1080p video typically takes 5–15 minutes to render locally. The output lands in the out/ directory.

5. How to Use AI for Video Cutting with Whisper + Claude

This scenario uses Whisper to transcribe your footage and produce a timestamped transcript. Claude then analyzes that transcript to identify which segments are worth keeping, and outputs a JSON file of keep/cut markers. Remotion renders only the kept segments into a final video.

Step 1: Complete all Scenario A setup first

Scenario B builds on everything in Scenario A. Make sure your project is created, the skill is installed, and your video is in H.264 MP4.

Step 2: Install ffmpeg

If ffmpeg isn't already on your machine:

bash

npx remotion ffmpeg

Remotion bundles a version of ffmpeg through its CLI. This is the safest way to get a version that's compatible with Remotion's pipeline.

Step 3: Install whisper.cpp and a language model

whisper.cpp is a C/C++ port of OpenAI's Whisper transcription model that runs entirely locally — no API key, no cloud, no usage fees.

Install via Remotion's package:

bash

npm install @remotion/install-whisper-cpp
npx remotion whisper install

This will download whisper.cpp binaries and prompt you to download a language model. For most content, the medium.en model balances speed and accuracy well. The large model is more accurate but slower.

Step 4: Transcribe your audio

Remotion provides a transcription command that handles the audio extraction and whisper.cpp processing:

bash

npx remotion whisper transcribe public/output.mp4

This outputs a JSON file with word-level timestamps in the format:

json

{ "text": "So today we're going to", "startMs": 1200, "endMs": 2800 }

Step 5: Have Claude analyze the transcript and generate cut markers

This is where your Claude API key is required. In Claude Code, provide the transcript JSON and ask Claude to identify the segments worth keeping. Be specific about your criteria:

Here is the transcript JSON from my video. Identify the segments where I'm speaking clearly about the main topic and mark them as "keep". Mark all pauses longer than 2 seconds, filler sections, and off-topic tangents as "cut". Output a segments JSON with startMs, endMs, and action for each segment.

Claude will output something like:

json

[
  { "startMs": 0, "endMs": 4500, "action": "keep" },
  { "startMs": 4500, "endMs": 7200, "action": "cut" },
  { "startMs": 7200, "endMs": 18400, "action": "keep" }
]

Important: Review this output before using it. Claude's AI-analyzed cut points are a strong starting point but are not frame-perfect. You may want to manually adjust timestamps for critical transitions before feeding the JSON to Remotion.

Step 6: Render the cut video

Pass the segments JSON to your Remotion composition. Claude Code can wire this up for you — describe what you want:

Using the remotion-best-practices skill, update the composition to read segments from public/segments.json and render only the "keep" segments from public/output.mp4 in sequence, with a 3-frame crossfade between each segment.

Preview, adjust, then render.

6. How to Prompt Claude for Better Video Results

The quality of your output is almost entirely determined by the specificity of your prompt. Remotion has a lot of knobs — and if you don't specify them, Claude makes assumptions that may not match what you had in mind.

Be explicit about every visual dimension

Weak prompt: "Add captions to the video."

Strong prompt: "Add word-by-word animated captions to the video. Font: Inter Bold, 40px, white. Position: horizontally centered, 80px from the bottom. The active word should scale up to 44px and turn #FFD700 yellow. Inactive words stay white and 40px. Animate in from opacity 0 over 4 frames."

The difference in output quality is significant. Every detail you leave unspecified is a detail Claude will decide for you.

Specify timing in frames, not vague language

Remotion works in frames. A 30fps video has 30 frames per second. When you say "fade in quickly," Claude will guess what that means. When you say "fade in over 8 frames," Claude has an exact number to work with.

Reference remotion.dev/prompts before writing your first prompt

The Remotion prompts library is a public gallery of community-created prompts with preview videos. Before prompting Claude for something you haven't tried before, find a similar example in the library. It will show you what's achievable and give you the right vocabulary to describe the result you want.

Ask for one thing at a time

Claude can handle multi-part requests, but video compositions are complex. If you ask for captions, a lower third, a progress bar, and an intro animation all in one prompt, the chance of something breaking is high. Build incrementally — get captions working, then add the lower third, then add the progress bar.

When something looks wrong, describe the visual state — not the fix

Instead of: "The animation is broken, fix it."

Say: "The caption text is appearing all at once instead of word by word. Each word should appear as it's spoken in the audio."

Claude can reason from a description of the visual problem to the code fix. It cannot reason from "broken" alone.

7. Common AI Video Editing Problems (And Fixes)

The video won't seek / captions are out of sync Almost always a codec issue. Your source video is probably not H.264. Run the ffmpeg conversion command from Step 4 of Scenario A and replace your source file.

Claude Code generates code that throws errors in Remotion Studio The skill may be out of date. Run npx remotion skills update to get the latest rules and patterns, then ask Claude to regenerate the component.

The preview server won't start Check that you're in the right directory (the Remotion project root, where package.json is). Then check that Node.js is installed correctly with node --version. If you see a version number, run npm install inside the project directory to ensure all dependencies are installed.

Whisper transcription is slow The large model is accurate but can take 2–3x real-time on a standard laptop. Switch to medium.en for faster transcription if perfect accuracy isn't required. base.en is the fastest but less accurate for technical or accented speech.

AI-generated cut points don't match the actual content This is expected. Claude's transcript analysis identifies likely cut points but cannot see the video — it only sees the words and timestamps. Review the segments JSON manually and adjust any timestamps that feel off. Remotion's frame-level control means you can be precise once you're in the JSON.

Render fails partway through This usually means Remotion ran out of memory on a long video or complex composition. Try rendering in segments (--start and --end flags on npx remotion render) and concatenating the outputs with ffmpeg.

Captions overlap with existing on-screen text You need to specify the position more carefully. Give Claude explicit coordinates: "Position the caption block at Y: 820px from the top, horizontally centered, in a 1920×1080 frame." If the video has a speaker title lower third, note that in your prompt so Claude positions captions to avoid overlap.

8. FAQs

Do I need to know React or JavaScript to use this workflow? No. Claude writes all the Remotion code for you. You describe in English, Claude generates the code, you preview and iterate. That said, being able to read code helps when things go wrong — you can describe the error to Claude more precisely.

Does this replace video editing software like Premiere or DaVinci? For motion graphics, caption overlays, and programmatic edits — it can replace those workflows. For anything involving complex multicam editing, color grading, or audio mixing, it doesn't. Think of it as a complement: edit and clean your footage in your usual tool, then use Remotion for the programmatic layers on top.

How much does this cost? Remotion itself is free and open source for personal and non-commercial use. Rendering is local — no per-minute cloud fees. You'll need a Claude Code subscription (currently $100/month) and, for Scenario B, the whisper.cpp model runs locally with no API cost. The only additional cost is if you use paid Remotion components like Animated Captions.

What video formats can Remotion output? By default, Remotion renders to H.264 MP4. It also supports WebM, HEVC, ProRes, and GIF. The format is set with the --codec flag when rendering.

Can I use this for videos longer than a few minutes? Yes, but render times scale with video length and composition complexity. A 10-minute video with animated captions can take 30–60 minutes to render locally. For long-form content, consider rendering in segments.

Can the Remotion workflow generate AI voiceover? Not natively, but it integrates well with tools like ElevenLabs. You generate the audio externally, export it as an MP3 or WAV, place it in the public/ folder, and reference it in your Remotion composition. Claude can wire the audio source into the composition for you.

What's the difference between the Remotion Agent Skill and the Remotion MCP? The agent skill is a knowledge file that teaches Claude how to write Remotion code. The Remotion MCP is a separate server that gives Claude runtime access to indexed Remotion documentation via a vector database. In practice, the agent skill handles most use cases. The MCP is more useful if you're building complex custom compositions and need Claude to look up specific API details.

My source video is a screen recording from Loom or Zoom — will this work? These files are often exported as MP4 but encoded with H.265 or with variable frame rates, which can cause seeking issues in Remotion. Always run the ffmpeg conversion step regardless of the file extension to ensure you have a reliable H.264 source.

 🔥🔥🔥 Three Highly Curated AI Stories

1. Meta Superintelligence Labs introduces Muse Spark

Meta's new lab just shipped its first model, and it's not a quiet release. Muse Spark is a natively multimodal reasoning model with tool-use, visual chain-of-thought, and multi-agent orchestration baked in from day one.

Translation: it can look at things, think through them step by step while showing its work, call tools, and coordinate with other agents on longer tasks. That's the full stack most labs are still stitching together.

It's live now on meta.ai and the Meta AI app, with a private API preview for select partners. Meta is also hinting at open-sourcing future versions, which would be a big deal for anyone building on top.

2. HeyGen says it solved avatar consistency with Avatar V

Character consistency has been the uncanny valley of AI video. Your avatar looks like you in one clip and like your cousin in the next. HeyGen is claiming they've fixed it.

Avatar V captures you in 15 seconds and then holds your identity across every video you generate. Swap outfits, change the setting, rewrite the script, and the person on screen still looks like you.

💡We're going to put this head-to-head with Gemini's Veo and Copilot's video tools on our channels over the next week and report back with the honest differences. If you've tested it already, reply and tell us what broke.

3. Gemini finally adds notebooks

Gemini now has project organization. You can spin up a notebook, pull in past chats, attach relevant files as sources, and keep everything for one project in one place instead of scrolling through an endless sidebar.

Hit "New notebook" in the side panel to start one. It's the kind of feature that sounds boring until you've lost a conversation thread for the third time that week. For anyone juggling multiple clients or research threads, this is the quiet upgrade that actually changes your daily flow.

🔥🔥 Two Pro AI Tips

1. 🤖 Subagents in Claude Code

Claude Code now lets you spin up subagents: smaller, task-specific assistants that handle a piece of a larger job while the main agent keeps the plot. Think of it as delegating. One subagent reviews code, another writes tests, another handles docs, and none of them clutter the main context window.

Best for developers running bigger projects where context bloat has been slowing Claude down.

2. 💬 Claude's reply-to-selection feature

Finally in Claude chat: highlight any chunk of text in Claude's response and you get a "Reply" button that lets you respond just to that piece. The selected text gets quoted in your next message, so Claude knows exactly what you're reacting to.

Best for long responses where you want to push back on one point without re-explaining the whole thread. Small feature, surprisingly good upgrade.

🔥 Things You Should Know About AI

Actually understanding MCP (Model Context Protocol)

You've seen "MCP" in every other AI post this month. Most people nod along. Almost nobody can explain what it actually does.

MCP is the open standard that lets AI models plug into your tools, data, and apps in a consistent way. Instead of every app inventing its own integration, MCP gives models one common language to talk to Gmail, Notion, GitHub, your database, whatever. It's why Claude can suddenly "use" your stuff without custom code per tool.

Here's a simple way to start using it today:

1. Pick one tool you use daily (Notion, Gmail, GitHub, a local folder).

2. Install the matching MCP server from the public directory.

3. Connect it to Claude Desktop or another MCP-compatible client.

4. Ask the model to fetch, summarize, or act on something from that tool in plain English.

5. Chain a second MCP server in and let the model move data between them.

We broke down the full protocol, how it actually works, and the workflows that matter. If you want the deep version with the working examples, check this out.

🧠 Key Concepts

A reference for the terms that come up consistently when using Claude written for people who don't want to wade through specification documents.

Token — The smallest unit Claude reads and writes, roughly ¾ of a word. "Hello world" is about 3 tokens.

Context window — Claude's working memory. Everything it can hold in mind at once: your conversation, files, and instructions.

Model — The specific version of Claude you're talking to. Haiku is fast and light, Sonnet is balanced, Opus is the most capable.

Extended thinking — Claude reasons step-by-step internally before giving an answer. Useful for complex or multi-part questions.

Hallucination — When Claude confidently states something false. It sounds certain, but the information isn't accurate.

Prompt — The message or instruction you give Claude to get a response.

System prompt — Hidden instructions set by a developer before your conversation starts. They shape Claude's tone and behavior without you seeing them.

Temperature — A setting that controls how creative vs. predictable Claude's responses are. Low = consistent and precise. High = more varied and surprising.

⚙️ Core Features

Artifact — A file or mini-app Claude creates in a side panel — a document, code snippet, chart, or interactive widget.

Project — A dedicated workspace with its own files, instructions, and chat history. Useful for ongoing work.

Memory — Claude remembers facts about you across separate conversations, so you don't have to re-explain yourself each time.

Deep research — Claude searches the web, reads multiple sources, and writes a comprehensive report — like having a research assistant.

Web search — Claude looks things up on the internet in real time to answer questions with current information.

Style — Claude learns your writing voice from samples and matches it when drafting content for you.

User preferences — Settings that tell Claude your preferred tone and format, applied automatically across conversations.

Compaction — When a conversation gets very long, Claude quietly summarizes older messages to keep things running smoothly without losing the thread.

🖥️ Products

Claude.ai — The main website and app where most people chat with Claude. Works in any browser or on mobile.

Claude Desktop — The downloadable app that unlocks deeper integrations like Cowork and Claude Code.

Cowork — Claude as a desktop agent. It can open apps, manage files, and take actions directly on your computer.

Claude Code — Claude for developers, living in your terminal and IDE. It reads your codebase, writes code, and runs commands.

Claude in Chrome — A browser extension that lets Claude see and act on the web pages you're visiting.

Claude in Excel — Claude embedded inside Excel. It can read, analyze, and build out your spreadsheets.

Claude in PowerPoint — Claude inside PowerPoint, helping you build and edit slides without leaving the app.

Claude Marketplace — A hub to discover and install connectors, plugins, and skills that extend what Claude can do.

 🤖 Agents & Automation

Connector — A live link between Claude and an external app like Gmail, Slack, or Notion, so Claude can read and act there.

MCP (Model Context Protocol) — The open standard that powers all connectors. Think of it as a universal plug for AI integrations.

Skill — A file that teaches Claude a specific workflow or best practice, so it handles a task the same way every time.

Plugin — A bundled package of skills, connectors, and commands built for a whole job role — like a Sales or HR plugin.

Sub-agent — A helper Claude that handles a parallel subtask inside a bigger job. Like delegating to a team member.

Slash command — A shortcut like /sales:call-prep that instantly triggers a preset workflow without you having to explain it each time.

Scheduled task — A recurring job Claude runs automatically on a timer, like a daily summary or weekly report.

Dispatch — Remote control for Cowork from your phone. Kick off desktop tasks via the Claude mobile app without being at your desk.

Computer use — Claude controlling your mouse and keyboard to take actions directly on your screen.

Context file — A background document Cowork reads at the start of every session — your standing instructions, preferences, or project details — so you never have to re-explain.

TL;DR — Too Long Didn’t Read

• MCP stands for Model Context Protocol. It's an open standard created by Anthropic in November 2024 that defines how AI models connect to external tools, data, and services.

• Before MCP, every AI-to-tool connection had to be built from scratch. MCP replaces that chaos with a single, universal standard.

• Think of it like USB-C for AI: instead of a different cable for every device, you get one connector that works across everything.

• OpenAI, Google, Microsoft, Cursor, and Windsurf have all adopted it. It's no longer just an Anthropic thing — it's becoming industry infrastructure.

• MCP doesn't just let AI read your data. It lets AI act on it — search files, send messages, query databases, run code.

• There are over 5,800 MCP servers available today, covering everything from Google Drive to Slack to Stripe to GitHub.

• Security matters here. Poorly configured MCP setups can be exploited. There are real, documented attack vectors to know about.

• You can start using MCP today through Claude Desktop without writing a single line of code.

1. Why MCP Exists (The AI Integration Problem)

AI models are powerful but isolated. When you ask Claude or ChatGPT a question, it draws on everything it learned during training — and nothing else. It doesn't know what's in your Google Drive, it can't check your Slack messages, and it has no idea what's in your company's database.

To fix that, developers started building custom connectors — essentially bridges between an AI model and a specific tool. Want your AI to read from Notion? Build a connector. Want it to pull from Postgres? Build another connector. Want it to access GitHub? Yet another connector.

This created what engineers call an N×M integration problem. You have N AI models and M tools, and you need a separate integration for every combination. Each one has to be built, maintained, and updated individually. Every time OpenAI releases a new model or Slack changes its API, something breaks.

The result was fragmented, fragile, and impossible to scale. The AI ecosystem was drowning in one-off integrations.

MCP was the answer to that.

2. What Is Model Context Protocol (MCP)?

Model Context Protocol (MCP) is an open standard that defines a common language for AI models to communicate with external tools and data sources.

Instead of every developer building their own custom bridge, MCP gives everyone the same blueprint. If your tool supports MCP, any AI that supports MCP can connect to it — without any additional custom work.

Anthropic announced MCP in November 2024 and immediately open-sourced it. The reasoning was deliberate: a standard only works if everyone uses it. By making it open, Anthropic invited the entire industry to build on it — and they did.

The analogy that gets used most often is USB-C. Before USB-C, every device had a different charger. Now one cable works across laptops, phones, and accessories from different manufacturers. MCP is doing the same thing for AI tool connections.

What makes MCP different from a typical API is intent. Traditional APIs are built for developers to connect software to software. MCP is built for AI agents — systems that are autonomous, can make decisions, and need to operate across many tools in sequence. It assumes the caller is intelligent but untrusted, which shapes how it handles security and permissions differently from standard API calls..

3. How MCP Works: The Architecture

 🔥🔥🔥 Three Highly Curated AI Stories

1. Anthropic launches Claude Mythos Preview and Project Glasswing

Anthropic's newest frontier model is here, and it comes with a mission. Claude Mythos Preview is the engine behind Project Glasswing, an urgent initiative focused on securing the world's most critical software.

What makes Mythos different is where it's pointed. The model is specifically built to identify software vulnerabilities, and Anthropic says it performs better than all but the most skilled human security researchers. That's not a casual benchmark — most enterprise security teams would pay serious money for that capability.

Project Glasswing is the broader wrapper around this: using AI to get ahead of attackers rather than just reacting to them. It's Anthropic planting a flag in one of the highest-stakes corners of the AI space.

2. Claude Code's /autofix-pr now runs straight from the command line

Claude Code users, this one's for you. The /autofix-pr command can now be triggered directly from your terminal, skipping the context-switch back to the UI.

Finish a PR, run /autofix-pr, and it sends your full session to the cloud. The autofixer gets complete context to address CI failures and reviewer comments without you having to explain anything.

It's a small surface change with a real productivity payoff — exactly the kind of thing you don't notice until it's gone.

3. HappyHorse-1.0 drops at number one on the video leaderboard

Twitter tweet

A new video model called HappyHorse-1.0 entered the scene and immediately claimed the top spot, leading in both text-to-video and image-to-video. The name is a lot. The performance apparently backs it up.

Early testing shows it handles multi-shot videos and detailed prompt instructions better than most models in the space — both of which have historically been weak spots for AI video generation.

No public release timeline yet, but it's already one to watch.

🔥🔥 Two AI Tools To Try Today

1. 📡 Exa Monitors

Exa is a search engine built from scratch specifically for AI agents — not the kind designed for humans typing into a search bar. Today they launched Monitors, which gives your agent a live radar for the web. You define what to find and how often, and Monitors surfaces new matching results automatically via webhook.

If you're building agents that need to track competitor moves, job postings, research updates, or any kind of real-time signal, this closes a gap that most search APIs still leave wide open.

P.S. We wrote a full guide on using Claude and Exa together for LinkedIn lead generation. If you want to see what this combo can actually do in practice, worth bookmarking

2. Clicky

Clicky is an AI teacher that lives next to your cursor. It can see your screen, speak to you in real time, and literally point at elements on the page — like a real tutor sitting beside you while you work.

The creator built it while learning Davinci Resolve and gave it a 10/10 rating after a few days of use. If you're picking up any new tool or software, this is a genuinely different approach to learning than watching a YouTube tutorial on mute.

🔥 Things You Should Know About AI

Turn Claude and Exa into a LinkedIn lead generation machine

Most people use Claude to write things. Fewer people know you can wire it to a web search engine built for agents and automate a big chunk of your lead gen workflow. Here's how:

1. Sign up for Exa and grab your API key from the dashboard.

2. Write a clear lead profile in Claude: target role, industry, company size, and any keywords that define your ideal contact.

3. Feed those criteria into Exa's search to pull a list of matching LinkedIn profiles.

4. Ask Claude to draft a personalized outreach message for each lead using the profile details Exa returns.

5. Review the messages, make small tweaks, and send.

Full step-by-step walkthrough here.

 🔥🔥🔥 Three Highly Curated AI Stories

1. Suno turned your voice into an instrument

Suno's v5.5 update is a hard pivot from generic AI music to something that actually sounds like you made it.

The headline feature is Voices (Beta) — upload or record your own voice and drop it straight into AI-generated tracks. There's verification and privacy controls baked in, so your voice stays locked to your account (a rare moment of "they thought about this first").

The bigger play is Custom Models. Train Suno on your own tracks and it learns your style, not just your sound. Pair that with My Taste, which adapts to your creative preferences over time, and you've basically got a personal music engine.

Why it matters: AI tools are starting to learn your creative identity, not just respond to prompts. That's a meaningful line to cross.

2. Sync-3 is dragging lipsync out of the uncanny valley

Sync Labs just dropped Sync-3, and judging by the demos, it's a real jump.

Their pitch: every model they've built has been pointing here. The focus is video that feels natural, expressive, and actually adaptable across different content types.

Lipsync is one of those things — when it's off, it's immediately off, and the whole illusion collapses. Nail it, and AI video suddenly becomes usable for storytelling, content, and business stuff that used to feel a bridge too far.

We're getting close to AI video that doesn't broadcast "AI video."

3. Pika gave AI agents a face (and a personality)

Pika rolled out a beta that lets any AI agent jump into a real-time video chat, powered by their new PikaStream 1.0 model.

This is not a talking head. The agent holds memory, adapts mid-conversation, and — if it's hooked up to tools — can execute tasks live while you're talking to it. Take a look here:

The interesting shift isn't the tech, it's the interface. We're moving from chat windows to face-to-face. That changes how people trust these things, how they use them, how they feel about them.

Which raises a question worth sitting with:

🔥🔥 Two AI Tools To Try Today

1. Copilot's "Model Council" mode

Buried inside Microsoft Copilot's Researcher tool is a setting called Model Council that almost no one is using.

Flip it on and your query gets run across multiple models (ChatGPT included), each generates its own report, and Copilot stitches them into a final answer that flags where they agree and where they don't. Built-in cross-verification — a real upgrade if accuracy matters.

2. Practicaly AI Course Platform

We’ve officially launched the Practicaly AI course platform — this is something we’ve been quietly building, and it’s finally ready.

This isn’t another “here are 50 AI tools” library. It’s designed to help you go from experimenting with AI to actually using it in your day-to-day work — with structured workflows, real use cases, and step-by-step breakdowns.

Think: how to turn one prompt into hours saved, how to actually apply AI in marketing, ops, research, or content, and how to build systems that compound over time.

It’s still early, and we’re building this in public — so if you try it, I’d genuinely love to know what works, what doesn’t, and what you want to see next.

🔥 Things You Should Know About AI

Turn Google Analytics into your personal AI analyst

You know the routine. Traffic dips, someone Slacks "what happened?", and you're suddenly four dashboards deep trying to look like you have answers.

You can now hook Google Analytics directly into LLMs like Gemini using MCP (Model Context Protocol) — and just ask.

Here's the setup:

1. Spin up the MCP server for Google Analytics using the official guide

2. Connect your GA property so the model can actually see your data

3. Open Gemini (or another supported LLM) and link it to the MCP server

4. Ask things like "What caused the traffic drop last week?" or "Which channels are driving conversions?"

5. Let the model surface the insights while you do literally anything else

 🔥🔥🔥 Three Highly Curated AI Stories

1. Nous Research drops Hermes Agent, and people are paying attention

Nous Research just released Hermes Agent, an open-source coding and reasoning agent that's being talked about as a serious alternative to closed tools like OpenClaw. The short version: it's built to run locally, hooks into your own models, and gives you full control over the agent loop instead of locking you into someone's cloud.

Where OpenClaw leans on proprietary infra and a polished UX, Hermes goes the other direction. Bring your own model, bring your own tools, own the whole stack. For devs who've been waiting for something genuinely open to build on, this is the one.

Community reaction has been loud and mostly thrilled. The "finally, something we can actually fork" energy is strong, and early testers are already sharing custom setups on GitHub.

2. Karpathy's LLM Wiki idea is quietly brilliant

Andrej Karpathy posted about how he's been using LLMs less for code and more for building personal knowledge bases, and the tweet went flying. The workflow: dump raw sources (papers, articles, images, repos) into a folder, let an LLM "compile" it into a structured markdown wiki inside Obsidian, then query the wiki like it's your own private research assistant.

The interesting bit is what he did next. Instead of shipping an app, he shared the whole thing as an "idea file," a gist you hand to your own agent so it can build you a custom version. It's a small preview of how software might get distributed in an agent-first world. Less "download my app," more "here's the idea, go build yours."

If you've ever felt like your notes app is a graveyard, this is worth a weekend experiment.

3. HeyGen unlocks Seedance 2.0 for real human faces

Video just got a big upgrade. HeyGen announced it's the only platform where Seedance 2.0 works with real human faces, supporting multi-avatar scenes with up to three Digital Twins in a single shot. That means cinematic, multi-character video using your actual likeness, not a generic avatar.

The reason HeyGen pulled this off while others can't comes down to their identity verification layer, which they built from day one to protect likenesses from being misused. It's the rare case where the safety guardrail is also the feature unlock.

Expect your LinkedIn feed to get very interesting very fast. Also expect the deepfake conversation to get a lot louder.

🔥🔥 Two AI Tools To Try Today

1. 📱 Google Edge Gallery + Gemma 4

Google quietly made it possible to run a capable AI model directly on your phone, no internet required. Here’s how:

1. Download Google Edge Gallery from your app store.

2. Get the free Gemma 4 model, and you've got a private assistant that can chat, analyze images, and transcribe audio entirely offline.

3. It's small enough to fit on your phone and, by most accounts, smart enough for daily tasks.

Best for anyone who cares about privacy, travels a lot, or just likes the idea of AI that doesn't phone home.

2. 🤫 5 unofficial Claude shortcuts

These are community-discovered prompt prefixes that change how Claude responds. Drop /god mode before your prompt for a more aggressive, comprehensive answer. Add L99 at the end for expert-level depth. Use OODA to run a military decision framework on your problem. Tack on artifacts to get a working app or dashboard built live in chat. And /ghost gives you a humanized answer that reads less like AI. Best for Claude power users who want to squeeze more out of every prompt.

🔥 Things You Should Know About AI

Run B2B sales prospecting without hiring a sales team

You don't need a five-person SDR team to fill your pipeline. Here's the quick version:

1. Define your ideal customer profile in plain English and feed it to an AI research agent.

2. Have the agent pull a list of matching companies and decision-makers from public sources.

3. Ask it to research each prospect and surface a personalized hook (recent news, a hire, a product launch).

4. Generate a first-touch email for each lead using that specific hook, not a template.

5. Review, tweak, send, and let the agent track replies so you can focus only on the warm ones.

Here's the full walkthrough.

TL;DR — Too Long Didn’t Read

Skip the guide and do this:

• Exa.ai is a semantic search engine built for AI — it understands the meaning behind a query, not just keywords, which makes it far more effective than LinkedIn's native search for ICP-based prospecting.

• Connecting Exa to Claude via MCP (Model Context Protocol) removes the copy-paste step. Search and analysis happen in the same session — you go from ICP to outreach-ready pipeline without switching tools.

• The anchor-based approach works better than broad industry searches: give Exa 3–5 existing customers and ask it to find 50 similar companies. It matches on structural patterns, not just keywords.

• The MCP session can run parallel searches — company discovery, decision-maker lookup, and buying signal detection simultaneously.

• Roles outperform names for durability: titles like "Head of CX" or "VP of AI" stay accurate longer than specific individuals whose LinkedIn data may be months out of date.

• Claude handles ranking and outreach angles in the same session — 50 raw companies becomes a top-10 prioritized list with one outreach hook per account.

• Installation takes one command for Claude Code users. Claude Desktop users need a config file edit and a restart.

1. What Is Exa.ai for LinkedIn Prospecting?

Most search tools return links. Exa returns structured intelligence.

Exa.ai is a search engine built specifically for AI applications — not for human browsing. It uses neural embeddings, the same underlying technology that powers large language models, to understand the meaning and intent behind a query rather than matching keywords. For LinkedIn lead generation, this distinction matters a lot: you can describe your ICP in plain English and get back companies, roles, and decision-makers that actually match the pattern — not just companies that mention certain words in their LinkedIn descriptions.

Traditional keyword search returns results based on what appears in text. Exa's semantic search returns results based on what something means. Ask for "companies similar to Walmart and Verizon" and it understands industry profile, size, complexity, and structural characteristics — not just names that appear alongside those keywords.

What this means for prospecting:

Exa raised $85M in Series B funding at a $700M valuation in September 2025, led by Benchmark with participation from Lightspeed and NVentures (NVIDIA's venture fund). It powers AI-native applications across thousands of companies, including Cursor and Notion AI.

2. Exa.ai vs Claude MCP: What’s the Difference?

This distinction is important because the same underlying search capability produces very different output depending on where you access it.

Exa's web interface (exa.ai and websets.exa.ai) is genuinely useful for prospecting. You can run natural language ICP queries and get real results. The bottleneck is what comes after: the responses are large and raw. You get a wall of text, then you have to copy it into another AI tool to clean it up, analyze it, build a table, or turn it into ranked pipeline. That extra step breaks the workflow and adds significant manual time.

Exa via MCP inside Claude removes that step entirely. When Exa is installed as an MCP connector inside Claude, retrieval and analysis happen in the same session. Claude gets the search results directly and immediately processes them — ranking accounts, identifying decision-makers, flagging buying signals, and building outreach hooks — without you doing anything in between.

The retrieval is identical. What changes is everything that comes after it.

There is also a practical cost advantage. Enrichment, structured output, deep research, and table-building are paid features inside Exa's web UI. When Exa is connected via MCP inside Claude, Claude handles the analysis layer — you get equivalent output without paying for Exa's premium tier on top of your Claude subscription.

3. How to Install Exa MCP in Claude (Step-by-Step)

Installation is simpler than most MCP setups. Exa provides a hosted MCP server, so you don't need to run anything locally.

You need:

• A Claude Desktop or Claude Code installation

• An Exa API key (available at exa.ai — new accounts get $10 in free credits, no credit card required)

Installing in Claude Code (one command)

Run this in your terminal to enable the full lead generation tool set:

claude mcp add --transport http exa \
"https://mcp.exa.ai/mcp?tools=web_search_exa,linkedin_search_exa,\
company_research_exa,deep_researcher_start,crawling_exa"

This connects Claude Code to Exa's hosted MCP server with all five lead generation tools enabled.

Here’s a quick overview of what that looks like:

Installing in Claude Desktop

Exa's documentation at docs.exa.ai/reference/exa-mcp generates a tailored install prompt based on the specific tool you're using. The general process:

1. Go to docs.exa.ai/reference/exa-mcp and select Claude Desktop

2. Copy the generated configuration JSON

3. Paste it into ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or the equivalent path on Windows

4. Add your Exa API key where indicated

5. Restart Claude Desktop

The MCP endpoint is: https://mcp.exa.ai/mcp

By default, only web_search_exa and get_code_context_exa are enabled. For the full LinkedIn lead generation workflow, you need to explicitly enable the additional tools in your configuration (see the tool list in the next section).