https://youtu.be/32aIxFroG0Y

I watched Anthropic's AI Fluency course so you don't have to.

It's a free official course called AI Fluency: Framework & Foundations. Their own view of what good AI use actually looks like in 2026. About 90 minutes of video. I sat down with a coffee, took notes, ran a highlighter through the slides, and condensed the whole thing into the deck below… plus the cheat-sheet PDF at the bottom of this email.

I want to extract the main frameworks for you. Give you the gold nuggets from the course. Absolutely take the course yourself (there’s certification too so that’s nice) but for those with less time use this guide to get the core value.

Slide deck is below. Today's PDF giveaway is the AI Fluency Cheat Sheet - the whole framework on ten pages, the 4Ds, the 3 modes, the loop, all of it.

The Big Takeaway

Most people think AI skill = better prompts.

Anthropic says AI skill = four competencies. The 4Ds:

• Delegation - what should AI even do?

• Description - how should I brief it? (THIS is the prompting part)

• Discernment - can I trust this output?

• Diligence - did I use it responsibly?

Three Ways To Use AI

Anthropic splits AI use into three modes. This is the first framework they give us in the course:

Automation = "do this". You give AI a repetitive task. It does it. You walk away. Best for transcription, formatting, summarising long docs, repetitive boilerplate.

Augmentation = "work on this with me". You and AI are sat next to each other (metaphorically) iterating together. Best for drafting, brainstorming, problem-solving, anything where your judgment is constantly in the loop.

Agency = "go do this and come back". You delegate a multi-step outcome. Agent goes off, plans, executes, returns. Best for research, complex tasks, anything you'd otherwise hand to a junior employee.

Most people are stuck in Automation mode (do my task) and don't realise Augmentation is where 80% of the leverage lives. Agency is where the next leap will come from, and we're already seeing it with Claude Code, Codex, Hermes, OpenClaw. This is the more “modern” way of using AI in 2026 that in just 2025 was a bit…wobbly.

Quick gut check: when you opened ChatGPT or Claude this morning, which mode did you use? Probably Automation. Probably could've used Augmentation or maybe even Agency? Worth thinking about.

The 4Ds: The Framework Worth Stealing

This is the most valuable part of the course.

Delegation - What Should AI Do?

The first question here is not "what's a good prompt?". The real first question is "what part of this job should the AI even touch".

Anthropic uses a venn-style picture: ME, SHARED, AI. Your judgment, context, and relationships sit with you. Scale, speed, pattern recognition sits with AI. The middle is where you collaborate.

Examples:

• First draft of an email? AI. (Yes.)

• Final opinion on whether to send it? You. (Probably.)

• A comparison table of three suppliers? AI. (Yes.)

• Blindly trusting the AI in strategic decisions? No. (Never.)

Most bad AI use is not a prompt problem. It's a work allocation problem.

Previously I outlined my content production system in this rather, ahem, comprehensive diagram:

Blue are the AI processes. Orange are the human processes.

You need to decide yourself for your AI usage and workflows what to offload and what to retain. Delegation is the first job.

Description - How Should I Brief It?

This is the slot where prompting actually lives. Anthropic breaks it into Product (what do you want made), Process (how should it approach the task), Performance (how should it behave while collaborating).

My god they like their frameworks to all use the same letter huh?…

When it comes to what you actually include in your prompt they break it out into the usual prompting tactics. If you want a clean structure for this, the RISEN framework works here.

Discernment - Can I Trust This Output?

The missing half of how most people use AI. Description asks well. Discernment judges well. You (the human!) critically evaluate three things: the output, the thinking, the behaviour.

Remember that AI fails smoothly. Polished does not mean right. Confidence does not mean competence - shit, AI tends to be more confident when it is hallucinating!

You need to use both Describe (Input) and Discern (Output) and check the difference. There’s a loop here:

This is the part most people mess up on honestly. They create something with AI. It looks pretty good. Cool, ship it.

Don’t do that.

READ it. Check it. Tighten up anything that is incorrect. Repeat this until the quality is where it needs to be!

Diligence - Did I Use It Responsibly?

This last step is very Anthropic…a bit dull, not at all sexy, but important.

This is the ethical layer. You need to make sure that you’ve been responsible in your usage of the AI. Did you use the right tool, right data? Transparency? Verification?

Here’s a quick checklist to make this actionable:

Anthropic succinctly put this as “If this is wrong tomorrow, could I explain what AI did and what I checked?”. Did you at least put in the due diligence here or were you lazy? Can you cover your arse?

If the answer is no… you skipped Diligence. Don't ship it. Do the work properly. Lazy-bones.

What Anthropic Got Right

Three things:

• Judgment not hacks. The whole course teaches you how to think, not what to prompt.

• Collaboration not magic. AI is positioned as a partner, not a replacement. Augment yourself with AI rather than purely automating - use AI in different modes as needed.

• Human expertise still matters. Your experience is the edge. You need this for the discernment step. Otherwise your output will be crap and you won’t know it is.

It's a more grown-up version of AI literacy than the pure prompt-engineering content dominating LinkedIn and the like. Here’s the full course from Anthropic: https://anthropic.skilljar.com/ai-fluency-framework-foundations

Want to get the presentation + summary cheat sheet? Available here https://aiwithkyle.com/resources/ai-fluency

Kyle

https://youtu.be/c_Bm6Mx-214

Marc Andreessen posted his ChatGPT custom prompt on X this week. Two million views. Most of them being mean to him…

Twitter tweet

Marc isn't stupid. He co-founded Netscape, runs a16z, worth a cool couple of billion. The man knows tech. The prompt though… is a museum piece. It's the kind of thing people were posting on Reddit in early 2023, reposted unironically in 2026 by someone who really should have updated his playbook…

Slide deck is below and resources below. Today's PDF giveaway is the System Prompt Builder Kit - what to put in, what to leave out, and an interview prompt that turns ChatGPT or Claude into a system-prompt architect that asks YOU the questions and writes a clean custom prompt for you.

The Prompt Doing The Rounds

Looking at Marc’s system prompt reveals some best practices.

It looks powerful because it's long and intense. That does NOT make it good. Mostly performative theatre. Let me break down what's actually wrong, because the lessons here apply to whatever you've got pasted into your ChatGPT custom instructions right now.

System Prompt vs Conversational Prompt

Quick definition first because half the confusion in prompt-engineering content comes from people mixing these up.

A conversational prompt is what you type in the box every day. "Rewrite this email." "Summarise this PDF." "What's the cheapest flight to Athens?" Low-stakes. Single-use. Doesn't need “engineering”. This is what most of us do most days.

A system prompt is the persistent instruction that sits behind every conversation. You write it once. It loads every time. In ChatGPT it lives in Settings -> Personalisation -> Custom Instructions. In Claude it's Settings -> General -> Instructions. In both you can also stick a system prompt at the project level.

OK caveat over!

Role Helps. Pretend Genius Doesn't.

The biggest tell that Marc's prompt is from the early-2023 era: "You are a world-class expert in all domains."

Marc. Comeon man.

You cannot make the model smarter by telling it it's smart. This has been tested by Ethan Mollick and the Wharton GAIL group. Telling the AI "you are a great physicist" does not make it noticeably better at physics. In fact telling it "you are a lawyer" doesn't make it worse at physics either.

So is role useless? No. But think of it as a steering wheel, not an engine upgrade.

Saying "you are a world-class expert in all domains" is generalising it further, which is the opposite of what role is for. Role works when you constrain the model to a specific lens.

Bad: "You are a world-class expert in all domains."
Better: "You are a senior B2B marketing strategist helping a founder turn rough ideas into clear LinkedIn posts."

The second one shapes tone, audience, and lens. It changes how the AI works with us - which is still super useful!

Adjectives Are Not Instructions

Marc tells the AI to be incisive, erudite, provocative, aggressive, argumentative. That's a vibes list.

Adjectives not verbs.

The AI doesn't know what erudite means in actionable terms. Shit, I don’t know what that means in actionable terms. It'll generate longer sentences, throw in some Latinate vocabulary, sound impressive, and that's about it.

Translate every adjective into a verb. Tell it what to do, not how impressive to sound.

Bad: "Be incisive, erudite, provocative, aggressive, argumentative." - does nothing.
Better: "Challenge weak assumptions. Separate facts from opinions. Give the strongest counterargument. Recommend a practical next step."

Steps Are Useful. Forced Thinking Isn't.

Marc has "process information and explain your answers step by step". This was actually excellent advice in 2023.

Not it ain’t.

In 2026, every modern reasoning model is already thinking step by step internally. GPT-5, Claude Opus 4.7, Gemini 2.5 Pro all run extended chains of thought before they answer. You can literally click and watch them reason for 10 minutes if you want. "Think step by step" is now redundant.

What you actually want is a useful answer structure. Not "think step by step" but "answer in this order".

Bad: "Explain your answers step by step."
Better: "Start by giving me your with the answer. Then a short reasoning summary. Then trade-offs and risks. Then next actions."

The “thinking” part will be done behind the scenes. You don’t need to force it.

You Cannot Request Your Way Out Of Hallucinations

Marc has "never hallucinate or make anything up". (I covered hallucinations in detail two days ago…) The short version: the AI doesn't know when it's hallucinating. There's no internal True/False switch you're flipping. There’s no database of true and false information,.

So telling it "don't hallucinate" does sweet FA. If anything it makes the model more likely to confidently claim it didn't hallucinate, which is worse.

What we can do is give some rules to deal with uncertainty.

Bad: "Never hallucinate or make anything up."
Better:

• "If a claim depends on current facts, check a source."

• "If something is uncertain, label it."

• "If information is missing, say what's missing."

• "Do not invent names, dates, studies, prices or citations."

Rules give the model an action it can execute. Otherwise we are just giving it hopes and dreams.

More Detail Is Not More Value

The single worst line in Marc's prompt: "Make your answers as long and detailed as you possibly can."

Wow.

The AI can generate infinitely. Tell it to write as much as possible and it will. You'll get bloated, padded responses where any actual signal is buried under verbose nonsense. AIs are terrible for this already!

Bad: "Make your answers as long and detailed as possible."
Better: "Give enough detail for a founder to make a decision. Prioritise practical trade-offs over exhaustive background."

The goal isn't more words. That’s like a school kid trying to hit a minimum word count for an essay. Leads to padding and bloviating.

RISEN: The Framework I Use

I trademarked a framework called RISEN™ a few years back when my first viral TikTok went to 3.5 million views. (Half the comments were people thanking me. The other half were people complimenting my eyebrows. Bizarre weekend.) I trademarked it out of interest honestly! To see how that worked!

I check it against the latest research every six months or so. So far it's held up. Yay!

R - Role. Who should the assistant be? Specific lens, not "genius".
I - Instructions. What should it actually do? Verbs, not adjectives.
S - Steps. What workflow or answer order should it follow?
E - End goal. What should you be able to do after the answer?
N - Narrow. What constraints, formats, sources or boundaries matter?

Let’s take Marc’s system prompt and make it better with RISEN:

Quarter the length. Triple the usefulness. Stays inside the global custom-instructions character limit. And every line is doing actual work.

Do You Always Have To Prompt Like This?

No.

Don't engineer every conversation. "Rewrite this email" doesn't need a five-part framework. It needs the email.

Conversational prompting is for: quick edits, simple summaries, brainstorms, low-stakes one-off questions. Just ask.

Engineered/RISEN™ prompting is for: custom instructions, repeatable workflows, team prompts, research tasks, anything you'll do more than three times, anything where the output goes somewhere public.

Prompt casually when the task is casual. Use RISEN when the output needs to be reliable. This ain’t rocket science really!

Today's Giveaway

Couple of goodies for you:

1. The walkthrough of what each RISEN element actually does and where to put your custom instructions in ChatGPT and Claude.

2. The Interview Prompt. Paste it into Claude or ChatGPT, the model interviews YOU one question at a time about your work, your style, your goals, your constraints, then writes you a polished RISEN-shaped system prompt you can drop straight into custom instructions. This is the bit most people will actually use. Skip everything else if you must, but use this.

3. The full livesteam Slide Deck.

It’s all available here: https://aiwithkyle.com/resources/system-prompt-builder

Kyle

https://youtu.be/GunzJavhSMk

OpenAI and Anthropic both moved into consulting on Monday. Same day. By accident? Something in the air!

Anthropic announced a $1.5B joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman. OpenAI announced something they're literally calling The Deployment Company… ~$10B valuation, 19 investors, TPG, Bain, Brookfield, BCG, McKinsey, Accenture, Capgemini, the lot.

Bizarrely both announced on the same day. Same target. They are not selling you software anymore. They are moving in.

Slide deck is below. Today's PDF giveaway is the 2-3 Year Window Playbook - the action steps for doing what these labs are doing, on a smaller scale, in your industry, before they get there.

But first…wtf are they doing here?

What They're Actually Doing

Both ventures send engineers inside businesses. Not "here's the API key, good luck". Engineers walk into real companies, sit with real workers, map real workflows, and rebuild how the business runs around AI.

First argets: healthcare, manufacturing, finance, retail, real estate. The Fortune 500. The PE-owned. Basically BIG BOYS. (That’s where the money is so makes perfect sense).

Funnily enough this is the Palantir playbook. Sit inside. Map everything. Become infrastructure. Once you're that deep, you cannot be ripped out without basically dismantling the company. Palantir is now worth $300B+. Built entirely on this move. It’s a smart one honestly…

The era of "we'll buy some ChatGPT licences and figure this shit out ourselves" is officially over.

If this follows Vampire rules they’ve just been invited in.

Still one of the creepiest scenes ever

The Jagged Frontier Just Moved

Ethan Mollick has been writing about The Jagged Frontier for years now. Hell, I’ve been writing about him writing about it. It’s important though! The idea is simple - AI is brilliant at some tasks, mediocre at others, and the boundary between the two is jagged, not smooth.

Ethan Mollick’s Jagged Frontier

For two years that frontier was coding. The reason being… coders built the AI. They knew coding. They optimised for coding. And once you have AI that's really good at coding, you can use it to build better AI. Snowball effect or, if you fancy, “recursive self-improvement”.

Then the frontier moved into writing and marketing - LinkedIn is now a slop avalanche, no offence. Then analysis. And now…

Consulting. Knowledge work. White collar.

Like…all of it.

This is a Trojan horse.

That’s what these new ventures from Anthropic and OpenAI are for. I can’t make this much clearer than this:

• To get into businesses.

• Work out what they do.

• Get AI to do the work.

• Remove the humans.

The Banks Don't Know They're On The List

Here's the funny bit.

Have a look at the list of investors backing this push. Blackstone ($1.3T AUM). Goldman Sachs ($3.7T AUM). TPG. Bain. Apollo. SoftBank. Their logic: AI automates our portfolio and client companies, those companies become more profitable, we win. Bonuses all around.

Cool. But….

Financial analysis. Due diligence. Deal modelling. Research. Reading a thousand-page document and producing a confident verdict. All of that lands directly inside what AI handles well right now.

Translation: the banks think they're running the AI play. They are also on the play. The list. OpenAI and Anthropic are going to come and absolutely each their (rich boy) lunch. Delicious.

Remember: every industry thinks it's the exception. None of them are.

Where We Are

Mental map. Five stages. We're in stage three.

Stage 1 - Coding / software - Donezo
Stage 2 - Content, writing, marketing - Underway
Stage 3 - Consulting / professional services - Now
Stage 4 - Finance, law, accounting - 2-3 years
Stage 5 - Broader white collar - Coming soon.

If your job is sitting in front of a computer moving information around - reports, decks, emails, models, briefs - it's on the list. All of those roles go to zero on a long enough timeline. Five years for most. Ten if you're a bit fortunate. Maybe to retirement if you're really fortunate.

The companies, when they can replace you, will. They don't owe us anything.

Even worse, the ones that don't replace you (out of the goodness of their hearts) will be undercut by the ones that do, by 30-40%, and lose the market regardless. That's the dynamic. Sorry. That's the bit I refuse to sugarcoat!

But.

There's a window before this hits at scale. And that window is where you (personally) live.

What This Means For You

The labs are going after Fortune 500 and PE-owned first. That's where the engineers scale. That's where the money is.

Mid-market is next. Small business after that. There's a 2-3 year window before this reaches your industry at scale. Right now, in the small-and-mid-market layer, demand is huge and supply is basically zero. That gap is yours.

Three moves.

Advisory. Businesses in your sector need someone who actually understands AI and understands the industry. That person barely exists yet. Charge accordingly. Retainer if you can. One-off if not.

Implementation. Don't just tell them what to do. Go in and do it with them. One business transformed properly = referrals, case studies, a track record. Project fees, or a monthly retainer when you build them an internal tool.

Education and content. Become the AI person in your niche before anyone else does. Build the audience. Own the vertical. One of my students, Katie, started doing this for veterinarians. She talks to vets about what they need to know about AI. Vets don't need the Sora updates - who cares. They need to know how to verify AI on diagnoses, or indeed if they should even be using AI at all for that. Helping them with these topics is valuable.

The labs are charging corporations millions for this. You can do it in your industry, at a smaller scale, and still do very very well.

Where To Start

Workshops.

Boring answer. Best answer. Sorry, not sorry!

Get a room of business owners in your sector. Teach them what AI means for their industry and what to do first. Pitch it as basic - because for them, they’ll need basic. If you are reading this now you're light-years ahead, even if it doesn't feel like that (it won't, that's just impostor syndrome talking, ignore it).

Shit just take all my stuff and feed it into ChatGPT to convert it to your niche and go teach it!

My workshop floor is $2,000/hour. Higher rate is $4-5,000/hour. I used to make that in a month - now it’s an hourly rate.

Sounds wild until you do the maths - 400 people in the room for $4000 is $10/person/hour. Cheap as chips for what they get.

If you want to explore this as an option have a look at https://aiwithkyle.com/ai-workshop-series - it’s a free mini-course about how to create your first workshop and market it.

Kyle

https://youtu.be/UToP5wfXXjs

The Chicago Sun-Times printed a 15-book summer reading list last year. Ten of them did not exist.

Fake titles. Fake plots. Andy Weir's The Last Algorithm… Andy Weir didn't write The Last Algorithm. Isabel Allende, Min Jin Lee, Maggie O'Farrell, Percival Everett - all credited with books that have never existed and never will. Ironically the model included a real Ray Bradbury one. Bradbury famously hated computers and the internet - he would have HATED generative AI! Fitting.

The journalist did not check. Nobody at the Sun-Times checked. It went to print.

Embarrassing.

And it's not just them. Deloitte (in)famously shipped two AI-hallucinated reports last year. One to the Australian government. One to a provincial government in Canada. Fake footnotes, fake citations, the lot. Sold to actual governments. By a Big Four firm. They got bullied for it, rightly…and a fair amount of that was me. Sorry!

Grab the full slide deck + resources here https://aiwithkyle.com/resources/ai-hallucinations

The Numbers Are Mad

According to an Exploding Topics poll 92% of people never bother to verify what their AI tells them.

Out of every 100 people prompting ChatGPT, 92 just take the answer. Paste it into the doc. Send it off to the client. Print it in the paper. Just ship it out the door. Whatever.

About 45% of AI responses have significant problems (BBC/EBU tested 3,000 of them). MIT measured something even nastier - the model is 34% more confident when it's wrong than when it's right. So when it sounds the most certain… that's exactly when you should be the most suspicious.

And there are 486+ court cases globally that involve AI-fabricated citations. That we've caught. The actual number is far higher. Lawyers submitting briefs with cases the AI invented.

This is the world we're in. Nobody checks. The AI is most confident when it's wrong. Hallucinations are seeping into the published record.

Hallucinations Aren't A Bug

Go and read Andrej Karpathy’s “dream” tweet as its important:

Twitter tweet

Here’s the most important part:

Hallucinations is all that Large Language Models (LLMs) do.

Remember that.

The model is not looking anything up. There's no database. No filing cabinet of facts with True/False stickers next to each one. It does not know things.

When you type a prompt, it predicts the next most probable token, then the next, then the next. It's been trained on roughly the entire internet - all the books it could scrape, every YouTube transcript, the lot - and from that mountain of words it learned what confident, correct-looking answers tend to look like.

That's all. Tbf that’s still SUPER useful and we can do amazing things with it. But that’s the underlying mechanism.

So technically… every output is a hallucination. We just call it intelligent when the output lands on something true and call it a hallucination when it doesn't. But it’s the same mechanism either way. We're the ones adding the value judgement, after the fact.

This is why the "AI lied to me" thing is a category error. A liar knows the truth and hides it. The AI has no concept of truth. It cannot lie.

Type 1 vs Type 2

There are two flavours of hallucination and only one of them is easy to spot.

Type 1 - Invents something that does not exist. Fake book. Fake court case. Fake statistic. The Sun-Times list is the textbook case. These are the obvious ones - go to fact-check, the thing isn't there, five seconds on Google and you're sorted.

Type 2 - Drifts from a document you actually gave it. This one's nasty. You hand the model a real PDF, a real report, a real source. It then misquotes it. Changes a number. Reorders a paragraph. Summarises a thing that isn't quite what the document said. This is verisimilitude (what a Wow word!) - the semblance of truth.

Type 1 you can catch.

Type 2 is the insidious one that’ll catch you unaware.

Will This Get Fixed? OpenAI Says Nope.

September 2025, OpenAI published a paper called Why Language Models Hallucinate. The headline finding: hallucinations are mathematically inevitable with current training methods.

Why? Because models are trained to answer. Abstaining gets penalised in training. So when the model doesn't know, it guesses… confidently. That's the optimal strategy under the current reward structure.

It’s kinda like if you take an exam. If you write NOTHING you will get 0 points. If you write something resembling an answer (even if you have NO idea what you are talking about) you’ll get more than 0. So: write something. That’s basically what the models are trained to do.

What Actually Helps (In Order Of Effort)

So: hallucinations happen. And will continue to happen. What can we do?

Here are 4 easy modifications.

Turn on web search. This is the single biggest free win. ChatGPT, Claude, Gemini all have a toggle. On the free plans it's sometimes off by default. Flip it on. GPT-5 with web search makes around 45% fewer factual errors. Easy win.

Use the stronger model for high-stakes work. Claude Opus 4.6 or 4.7. GPT-5 Pro with thinking on. Gemini 2.5 Pro. They're slower. They cost more - or have limited usage on free plan (this is one of the biggest reasons to oay). But when the output is going somewhere public… use them! I built today's slide deck on Opus and the research alone was 20 minutes of thinking time before it even started drawing. That's the price of getting it right. Probably the best $20/month most people can spend, tbf.

Deep research / extended thinking. Toggle it on for anything where the answer needs to be correct, not just fast. The model goes off, checks citations, pulls real sources, comes back with something defensible. Every chatbot has its own name for this and they all do roughly the same thing. Use it.

NotebookLM for grounded learning. This is the bonus one. NotebookLM gives you a personal RAG system - you upload PDFs, transcripts, books, and the model is forced to answer from those sources only. Citation hallucination drops dramatically when the model is grounded in your documents. Great for studying. Great for working with research papers. Great when you need this answer from this document… not the model's vibey “memory” of it.

These are more about the harness and tools around the model than the model itself. This is where most improvements in accuracy occur now.

Remember it’s not Google

The reason all this is hard is that we've been trained for 25 years to think "computer + question = answer". You typed it into Google. Google found the page. The page had the answer. Find the thing, show you the thing.

That mental model does not apply here. It’s over.

AI is a writing tool. A drafting tool. A thinking tool. It's excellent at working with information you give it. It's unreliable at sourcing information from its own memory - because it doesn’t have an indexed database like Google does. Treat it accordingly.

The people getting the most out of AI right now aren't treating it as a smarter Google. They realise it’s not just Google 2.0. It’s a tool for different types of work. Use AI to work with information. Don't trust it to source information from memory.

Resources: you can get the full slide deck from the Livestream and a PDF quickstart guide here https://aiwithkyle.com/resources/ai-hallucinations

Kyle

https://youtu.be/4gI09RlCUi0

OpenAI banned goblins.

Not joking. Buried in the Codex system prompt is a literal rule: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query." Someone spotted it last week.

Bizarre? Yes. Useful? Surprisingly, also yes. Because the goblins are a super clean case study we've had all year of how AI models develop verbal habits that nobody asked for. And once you understand how the goblins got there, you understand every other AI tell - delve, em dashes, "it's not X, it's Y", the lot. So that's what we're doing today.

Slide deck is below. Today's three free resources are at the bottom: an AI Tells cheat sheet (PDF), an AI Signs skill file (drop into Claude Projects), and a how-to guide for using them.

Why Goblins

OpenAI actually wrote a paper about it. A whole research post called "Where the goblins came from". I love that this is a real thing scientists at OpenAI had to write. Here’s an example from their Chief Scientist:

The story... is actually quite interesting. ChatGPT has a few personality presets - Default, Cynic, Robot, Listener, Nerdy. The Nerdy preset got rewarded during reinforcement learning for being "playful" and using "creature-heavy metaphors". Once that reward kicked in, the model started producing more goblin-flavoured rollouts. Those rollouts were then used to fine-tune the next version. And so on.

3,881% increase in Goblins

Nerdy mode was only 2.5% of all responses. But it produced 66.7% of all goblin mentions. By the time GPT-5.4 shipped, goblin mentions in Nerdy mode were up 3,881% vs GPT-5.2. The behaviour also bled into the base model via fine-tuning data. Goblins were leaking everywhere.

Their fix? A hard ban in the system prompt. "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures." Crude. Effective. And pretty funny.

So why am I covering this?

This isn't really about goblins. It's about the mechanism. Models develop verbal habits, those habits get reinforced, and eventually they leak everywhere. Goblins is the silly (and most up to date) version. Delve was a previously famous version. Em dashes are the version we're still living through. They all came from the same feedback loop.

What An AI “Tell” Actually Is

A pattern that appears so consistently in AI output that it signals - consciously or not - that a machine wrote it.

Three flavours:

Vocabulary - certain words AI just loves. Delve. Tapestry. Pivotal. Multifaceted. Robust. Vibrant. You know them when you see them.

Structure - "It's not X, it's Y." Bullet points for two-sentence answers. Markdown headers on a quick reply. Rule of three for everything. Curly quotes. Em dashes at every clause boundary - like this - constantly.

Tone - "Great question!" "Absolutely!" "I'd be happy to help." "It's important to note." "Certainly!" That sycophantic eager-intern energy.

These exist because every frontier model is trained on the same data, rewarded for the same signals, optimised against the same human preferences. Predictable input gives predictable output. The mechanism's working fine, it's just working in overdrive.

The Wall Of Shame

Here’s a wall of shame of some of the most famous tells. Which is your fave? 😅 

Delve went viral in April 2024 when Paul Graham posted "Someone sent me a cold email. I noticed it used the word 'delve'. It's a sign that text was written by ChatGPT." Once the lid came off, delve usage plummeted. OpenAI quietly trained it down. It's now a fraction of what it was.

The em dash crisis peaked in February 2025. Everyone had ChatGPT-shaped em dashes in their LinkedIn posts. OpenAI announced a fix in November. Did the fix work? Sorta. They still leak in. Annoying for those of us who actually like using em-dashes!

The structural ones are stickier. "It's not X, it's Y" (Wikipedia calls this "negative parallelism" - which is a brilliant fancy-schmancy academic term for an annoying habit). You're not a writer, you're a storyteller. Nope. It's not a tool, it's a thought partner. Stahp!

And then the openers. "Great question!" "I'd be happy to help with that!" "Certainly!" If a human opens an email with "Certainly!" they're in a P.G. Wodehouse novel. Otherwise, it's a bot.

(I worry even including these examples in my newsletter will get it flagged as AI!)

Oh and if you want to train your eye for these, just open LinkedIn. It's an AI petri dish. The posts are AI. The comments are AI. The replies to the comments are AI. It's bots all the way down.

Wikipedia's Living List (Bookmark It)

The single best free resource on this topic isn't from OpenAI or Anthropic. It's from Wikipedia.

Wikipedia:Signs of AI writing is roughly 15,000 words of hyper-forensic detail on how to spot AI-generated text. It's actually not a Wikipedia article, it's an internal guide for editors. And it's maintained by people whose actual role is detecting AI - because Wikipedia banned AI-generated article content in March 2026 and editors are now actively cleaning up the slop!

Sections cover early-AI-era tells (GPT-3.5 and below), GPT-4 era tells, later-model tells (GPT-4.5, Claude 3.x, Gemini 1.5+), structural tells, edge cases, and false positives. It's organised by era, which is useful, because...

...it's a moving target. Tells change. Delve was the king two years ago and barely shows up now. The em dash discourse peaked last year. Goblins is this week’s tell. Whatever's flagged today might be clean in a year. Whatever's clean today might be the next delve.

The point is, this list is a living document. Bookmark it. Check it monthly. Or…we’re going to be sneaky: use it as your AI input.

Belt And Braces: How To Actually Use This

Here's the practical bit. How can we make our AI writing more “human”?

I use a belt and braces 2 layer system.

Layer 1 - Before you write. You compress the Wikipedia list into a short markdown file (or download mine - link below). You paste it into your Claude Project instructions or system prompt. Format: "Avoid the following words and structures: [list]. Write in this voice: [your voice description]." Now every draft you generate dodges most of the common tells from the get-go. You do this ONCE and it’s locked in.

Layer 2 - After you write. You build a dedicated AI Signs skill / custom prompt. You run it on every draft as a second pass. "Check this for AI tells using the following list. Flag anything and suggest a human alternative." It catches what slipped through Layer 1.

The system prompt is the belt. The skill is the braces. Both together = your trousers stay up. Either alone = there's a chance something slips through.

Now do you need to always use these two layers? Nope - absolutely not. Most of the time it doesn’t matter if something sounds AI generated. This is primarily for public facing writing.

AI Detectors Are Basically Useless

Quick aside, since this comes up every single week in my comments.

Use Turnitin! Use Copyleaks! Use GPTZero! No. They don't work. OpenAI launched their own AI Text Classifier in January 2023 and quietly killed it six months later "due to its low rate of accuracy." If the people building the models can't reliably detect their own outputs, no third-party tool can either - this fact should have demolished the AI detection industry but somehow it persists!

So. Don't trust the detectors. Don't pay for the detectors. If your school or your client is using them, push back. They're snake oil with a UI that sells false certainty.

Free Resources

Today's four giveaways:

1. AI Tells Cheat Sheet (PDF) - the compressed Wikipedia list, paste-ready, print-ready

2. AI Signs Skill File (.md) - drop into Claude Projects or your system prompt as a second-pass checker

3. Belt and Braces How-To (PDF) - the actual workflow, step by step

4. The livestream slide deck

All available at aiwithkyle.com/resources/goblins

Goodbye Goblins!

Kyle

https://youtu.be/hzLjVBSohcY

A reader posted in the WhatsApp group yesterday: "Have you done a video on setting up GitHub? With people in this group spending lots of money on code I'm guessing it becomes important. Any tutorials on linking GitHub to Claude, Manus etc would be amazing."

Yep. Great idea! And well aligned with my new push to bring you more 101 guides. So here we are.

GitHub.

The word that has scared off half my non-technical audience for the last two years. Sounds technical (ok and a little bit silly). Certainly looks technical. The website itself is a wall of jargon that makes you want to close the tab and go and do something easier. Like learn Greek (which I am, badly).

But it isn't really. Not the bit you actually need.

We’re not even going to memorise any commands… we’ll keep it simpler than even this!

Today's the gentle 1-0-1 walkthrough. By the end you'll have a free account, an app on your laptop, and one repo to your name. Total time… about 15 minutes. Tea included.

What GitHub Actually Is

Forget code for a second.

GitHub is a shared folder. That's it. That's kinda the whole thing. It sits in the cloud, all your files live in it, and every tool on every device you own can connect to that folder and stay in sync.

Same idea as Google Drive. Same idea as Dropbox. The only reason it looks scary is that the developer crowd built it for themselves and the website is decorated in their lingo (issues, pull requests, actions, whatever). You can ignore 95% of that. What you actually need is the folder.

Because it’s devs they’ve even given it a fancy term: repository, or repo. But to keep things simple a repo is a folder. End of.

Why You Should Care

If you're vibe coding without one, here's what your project life looks like.

Maybe start a website in Lovable. Then you decide to keep working on it in Claude Code on your desktop. Then you switch to Cursor on your laptop. Now you've got three versions of your project in three different places, none of them in sync, and you can't remember which is the latest.

This is the reportDraft-final-v2-(4)-FINAL-FINAL-actuallyfinal.docx problem… for code.

It's a nightmare. People lose hours to it. People lose actual paying work to it. And the fix is just… one repo. One source of truth that every tool talks to. You commit changes, every device pulls them, you keep working from wherever you are.

Done. Cool?

The Terminology That Sounds Scary But Isn't

You'll hear five words thrown around. Once you know what they actually mean they stop being intimidating.

Repo. Your project folder. That's it.

Commit. Saving. You've made some changes, you give them a name ("added homepage hero"), you commit. New save point.

Push. Sending your saved changes from your computer up to the cloud repo.

Pull. Grabbing the latest version of the cloud repo down to your computer. Opposite of push funnily enough.

Branch. A copy of your project to mess about with experimental stuff in, without breaking the main version. When the experiment works you merge it back in. When it doesn't, you bin the branch.

That's basically the lot. Yes there's more. No you don't need to know any of it on day one. There's a brilliant zine by Julia Evans called How Git Works if you ever do want to get under the hood. Not sponsored or affiliated. Just cool.

https://wizardzines.com/zines/git/

The 5-Minute Setup

Right, the nuts and bolts.

1. Go to github.com and create an account. Free. Use your Gmail. Done.

2. Download GitHub Desktop. GitHub Desktop is a normal Mac/Windows app that means you never have to touch the terminal. Push, pull, commit… all buttons. No command line. None.

3. Open GitHub Desktop, log in with the account you just made, and click "create new repository." Give it a name. Set it to private.

4. Commit your new repo and it’ll be sent to Github.

That's it. You have a repo.

I know…what was all the fuss about eh?

You now have a folder on your computer that's connected to the cloud. Every save you make from now on can live in this repo and be available on any device you log into.

Connecting It To Your AI Tools

OK a cloud folder for your project is neat and all. But how do we use it. Here's where it gets useful.

Every modern AI coding tool plugs straight into GitHub. Lovable has a little Octocat icon in the top right - click it, connect, sync. Claude Code, Codex, Cursor - all of them open a GitHub repo as a project. The repo becomes the single source of truth and the tools become interchangeable. Start in Lovable, finish in Claude Code, pop into Cursor for a tweak. Same project.

So: connect them all up to your repo. Any tool you use look for how to connect it to Github. Once you know about this you’ll start seeing connection points EVERYWHERE.

It's Not Just For Code

What if you aren’t a coder. You’ve probably left by now. Hmm. Bummer. But if you haven’t this part will be handy!

GitHub is fantastic for non-code projects too. A book draft. A content brain. A personal knowledge base. Anything where you've got a bunch of files you want synced across your devices and connected to your AI tools.

You've probably seen the viral "second brain" posts on Twitter - someone showing off their gorgeous Obsidian setup with the spider-web graph view. Pretty. Mostly useless tbh… but pretty!! Underneath that pretty interface is a GitHub repo. That's kinda all Obsidian is - a viewer for a folder of markdown files that lives in (you guessed it) a GitHub repo.

ooo pretty

I have one of these. Mine's called ai-brain (creative, I know!!). It holds my book outline, my livestream transcripts, my newsletter drafts, my content scripts, all my playbooks. Every morning Claude Code reads from it and writes to it. This newsletter you're reading right now lives in that repo, first as a draft and then the final human edited version (and yes the AI learns by looking at the difference between the two…more on that in another issue!)

Your One Job Today

Stop reading. Stop watching tutorials. Don't read all the GitHub docs (please don't, you'll never come back…and they are SO boring).

Go to github.com. Make an account. Install GitHub Desktop. Make one repo. Call it whatever. Set it to private. Commit something.

Then for bonus points, install GitHub Desktop on a second device - phone, laptop, whatever - and pull the same repo. The first time you make a change on one device and watch it appear on the other one is the moment this clicks.

See! Github isn’t that scary!

Kyle

https://youtu.be/rUKDzvlHvFI

The best prompt engineers aren't typing. They're talking.

Power users figured this out early: speaking a prompt gives you 10x more context in half the time. You include the edge cases, the examples, the tone you want — because talking is fast enough that you don't skip them.

Wispr Flow captures everything you say and turns it into clean, structured text for any AI tool. Speak messy. Get polished input. Paste into ChatGPT, Claude, Cursor, or wherever you work.

89% of messages sent with zero edits. 4x faster than typing. Works system-wide on Mac, Windows, and iPhone.

Start flowing free

You can run AI on your laptop. Right now. Today. Without the terminal, without writing code, without being technical. Five or ten minutes to get it set up.

I know it sounds intimidating. It isn't. Local LLMs (Large Language Models) are now a download-and-run job. Click an app, pick a model, type a prompt. And then you basically have a local ChatGPT, except it runs on your hardware, costs zero per token, and your data never leaves your machine.

This guide is the gentle introduction for those who think this sort of thing is only for techies. It’s not. So… no jargon, no scary command lines, no sending you out to buy a $10,000 computer. If you have a laptop made in the last three years with at least 8GB of RAM, you can do this today.

What Even Is A Local LLM?

First up what’s an LLM? It’s a Large Language Model. GPT5.5, Opus 4.7 and Gemini 3.1 are LLMs. ChatGPT, Claude and Gemini are the chatbot applications built on top of the LLMs.

When you use ChatGPT or Claude, you type a prompt, it leaves your computers, travels to a data centre on the other side of the world, the model thinks about it on someone else's hardware (run by the AI company), and the answer comes back. That's cloud AI. It’s like renting.

Why You Should Care (Four Reasons)

Why bother with local AI? Sounds like a lot of faff right? Well…here are a couple of compelling reasons.

Privacy. Your prompts and data never leave your machine. That client document you'd never paste into ChatGPT? Paste it into a local model. The medical letter, the legal contract, the salary spreadsheet, the personal journal entry. All fine. Nothing logged, nothing trained on, nothing leaked.

If you are running a model for your business this becomes VERY important.

Cost. £0 per token. The model is free to download. After that, the only cost is electricity. Run it 1,000 times a day, run it 10,000 times. Same bill.

Offline. On a plane, on a train, in a coffee shop with rubbish wifi, in a power cut with a charged laptop. Works. Cloud AI doesn't - you’re donezo.

Learning. If none of the above float your boat honestly it’s worth doing this just to build your confidence. Local AI seems like one of those complicated advanced topics. It’s worth deploying a local model just to see how easy it is.

Can Your Device Actually Run This?

Probably yes. The honest answer depends on your RAM. And some other factors but this is a good rule of thumb to start.

If you have an Apple Silicon Mac (M1 or later), you're in better shape than equivalent Windows machines because the unified memory architecture is brilliant for local models. An M2 MacBook Air with 16GB RAM runs an 8-billion-parameter model very comfortably.

"I don't have the right hardware" is the most common excuse for not trying this. Almost always it's wrong. If you bought your laptop in the last three years and you didn't go for the cheapest possible spec, you can run something useful.

Pick Your Tool: LM Studio

There are about a dozen apps for running local LLMs. I’d recommend LM Studio to start. Free, polished, works on Mac and Windows. Nice and easy to use.

Why LM Studio over the alternatives:

• Free, no signup, no credit card

• Click an app, browse models, click download, run

• Works exactly like the ChatGPT interface you already know

• No terminal, no command line, no "config.yaml"

• Built-in model recommendations based on your hardware

The other tools are fine but they're for people who already know what they're doing. LM Studio is for you, today.

Literally here is your playbook:

Download from lmstudio.ai.

Pick Your First Model

Hugging Face (a repository of downloadable models) has 2M+ models. Yikes.

Which to use??

This is a moving target. Whatever I tell you now will be out of date next month. So we’ll just cover the basics of how to choose one.

LM Studio has a browser with all the models. And it tells you which models will work on your current system. Choose something and download it. Don’t overcomplicate this.

If really in doubt find a Google model. Right now (April 2026) that’s Gemma 4. Gemma is the micro Gemini and works well on local devices. Even iPhones!

Don't try to download the biggest model that just-about fits. Technically bigger models will work but they’ll be slow as hell.

The model that fits comfortably and runs fast will get actually get used. The one that takes 90 seconds per response won’t. Speed matters more than benchmark scores if you actually want to build this into your day to day.

What Local LLMs Are Not (Yet)

Honesty section. Local LLMs are not a Claude or ChatGPT replacement, and pretending otherwise will set you up for disappointment.

A 7B local model ain’t going to write Python as well as Claude Code. It won't do agent tasks like Codex. It won't stay coherent over a 100,000-token conversation. It will hallucinate more on factual questions. It will sometimes lose the thread mid-paragraph.

It will not be as good. Simple as.

But…sometimes your AI doesn’t need to be bleeding edge. We don’t need to bring an ICMB to a rock fight.

What it WILL do: chat, summarise, draft, translate, brainstorm, restructure, classify, extract, format. About 80% of what most people use AI for. All while being free and private. It’s a tradeoff.

I personally treat my LLM as a workhorse for routine tasks and my cloud subscriptions as the frontier model I reach for when the task actually needs it. Different tools for different jobs.

Your next step? If you've never run a local model: Download LM Studio. Pick a model that fits your hardware. Run one prompt. That's it. That’s the homework!

Kyle

https://youtu.be/GVWauQYAbsQ

For most of the last fifty plus years, the deal was simple. Get a job. Get safety. Trade upside for stability. If you wanted the upside, you had to take the risk. Entrepreneurship was the risky choice - sometime it paid off big-time but more often it did not.

That deal just broke. Quietly, but completely.

The 9-to-5 has become MORE dangerous. Without any increased reward.

And entrepreneurship - the thing your parents told you was reckless and a terrible idea - is now increasingly a sensible move.

The Frame Your Parents Taught You Just Broke

Common advice for most kids entering the adult world: get a job, work hard, don't take stupid risks. That advice was correct in 1995. It was correct in 2005. It was correct in 2015. It is no longer correct in 2026. Something shifted.

The reason it worked was simple maths. Jobs paid less but they paid forever. Businesses paid more but they could fail. You traded ceiling for floor.

But the floor is crumbling.

Something has gone horribly horribly wrong.

The Risk Flipped

This is the framework. If you only take one thing from today, take this.

In the OLD WORLD, the deal was clear. A 9-to-5 was small risk, small reward. Entrepreneurship was big risk, big reward. The risk-reward graph was symmetrical and the safe option was obvious. Most people picked the small risk. That was a rational choice for most - especially if you are risk averse.

In the NEW WORLD, the maths flipped. A 9-to-5 is now big risk, small reward. Entrepreneurship is still big risk, big reward.

The job didn't get less risky. It got more risky. AI is shrinking teams. Layoff cycles compressed from years to quarters. Skills that paid £80k five years ago pay £45k now or zero. Your industry can disappear in a fortnight - just ask anyone who was a stock photographer in 2023 or anyone who used to take Chegg gigs. Gonezo.

 Business still carries risk, but business gives you upside. That's the entire point. That part of the equation hasn’t changed. It’s just now a LOT more sensible a risk to take than it was before.

Now I'm not telling you to quit your job tomorrow. In fact: keep it. Jobs are going to get rarer. KEEP your 9-5 for as long as humanly possible.

But you DO need to be thinking about what this means moving forward. What comes next.

Gen Z Is Already Doing It

The Guardian published a great pice on Saturday. Here’s the link: https://www.theguardian.com/technology/ng-interactive/2026/apr/25/gen-z-entrepreneurs-business-ai

https://www.theguardian.com/technology/ng-interactive/2026/apr/25/gen-z-entrepreneurs-business-ai

Gen Z are getting hit hardest by shifts in the job markets. With insane joblessness levels and the painful task of sending out hundreds of applications for potential gigs only to be turned down (and ground down) by months of rejection.

The article focuses on the increasing number of Gen Z who realise its hopeless and have decided to go their own way.

When the traditional route is blocked it’s time to forge your own path. And that’s what they are doing.

These and more stories are in the article - highly recommend reading the original.

Importantly most of these young people didn’t PLAN to go this route. They all tried the traditional route first, hit a wall, and pivoted. This isn’t just Gen Z aren't being entrepreneurial because they're brave. They're being entrepreneurial because the alternative - waiting for a graduate scheme or finding a job that no longer exists - is worse. It’s necessity, not idealism.

Realise though that this is not just Gen Z. It’s just that Gen Z are the first to be affected. The canaries in the coalmine - sorry Gen Z. Because entry-level jobs are the first to fall to AI.

But … AI will roll up the chain. Unless you are within spitting distance of retirement you need to be ready for it to catch up.

I’ve been posting a lot recently about the need for everyone to wake up and realise what’s happening. It’s caused some consternation and (as is to be expected) argument. That’s good! We need to engage. Here are a couple of the objections.

Objection 1: "Not Everyone Wants To Own A Business"

I get this every time I post on this topic. "Not everyone wants to be an entrepreneur. Some of us just want a job. Why does everything have to be a hustle?"

Fair. I hear you. But you're answering the wrong question.

It's not about wanting. It's about the safety net being gone. Wanting a job is fine. Wanting a job in a market where the job is being deprecated is a problem. What you want at that point doesn’t really matter.

The question isn't "do I want to be an entrepreneur or start a business" The question is "what do I do when the thing I wanted to do isn't available."

Also, ownership is a spectrum, not a binary. Freelance is ownership. A side hustle is ownership. Consulting is ownership. Selling something you make is ownership. Multiple income streams is ownership. You don't need to "be an entrepreneur" with a capital E.

I’m not talking about becoming the next Steve Jobs, raising VC money and moving the San Francisco. I’m talking about carving out your own income to support yourself and your family.

67% of Gen Z say they want multiple income streams. Not because they all want to be founders. Because they've worked out, much earlier than my generation did, that single-income employment is structurally fragile. They know how fragile the system is and (correctly) want protection.

Objection 2: "AI Isn't Really Creative"

The other pushback I get: "Sure AI can do some stuff but it can't be truly creative. My job needs human judgement."

Two responses.

One: most jobs are not 100% creative or requiring judgement. Even very creative jobs are probably 60-70% routine. Drafting briefs, formatting decks, scheduling, processing email, writing the boring sections of a report.

Most jobs are full of boring BS.

AI does not need to be perfect to shrink a team. 70% is enough. If AI takes the routine 70% off ten people's plates, you don't need ten people anymore. You need three. Maybe. The other seven are unemployed despite being "creative."

Two: even the genuinely creative work is being commoditised. Look at what Claude Design and ChatGPT Images 2.0 did to "good enough" graphic design last week. Average is no longer a differentiator. Only the top 5% of taste-makers survive that compression. Are you in the top 5%? Be honest. If you are: great! You are genuinely going to be fine - congrats. But for the remaining 95% that’s no comfort.

I'm not saying creativity is dead. I'm saying "I'm safe because my job is creative or requires judgement" is the cope a lot of knowledge workers told themselves in 2023, and the data from 2025-26 has not been kind to that argument. If your defence against AI is that AI can't quite do what you do, that defence has a short half-life.

Objection 3: "I'm Out. Going Homesteading. "

The other end of the response curve: people who do believe that AI is capable. And for that reason are opting out. This comes in a few forms like "I'm out. I'll grow vegetables and raise chickens." Or a milder alternative “I’m going to work with my hands” / “learn a trade”.

Awesome!

I have so much time for this! A homestead is ownership. If you've already pivoted to a trade or running a smallholding, you're not the audience for this newsletter - you're already where I'm trying to get people. You’ve done it! You’ve taken control and made a decision.

The version I have less time for is doomerism as a strategy. "We're all going to be unemployed and money won't matter" is not a plan. It's a way of avoiding having to do anything. Even if you genuinely believe AI will collapse the labour market in five years, are you just going to sit and wait? What if you are wrong?

Agency beats passivity in any future.

You don't have to believe my version of the future to take action - you don’t have to start a business. But you have to do SOMETHING. Just waiting and hoping for the best (or the worst!) isn’t really a strategy!

What To Do This Week

Here are a few options. Start with whatever seems easiest - don’t think about what’s best. Best gets in the way of starting.

1. Sell your time. Freelance. Consult. Take on a service contract. Generally the easiest path to your first non-employer income. Start here to PROVE to yourself it’s possible.

2. Sell your knowledge. Coaching. An info product. A workshop. Take what you already know how to do and turn it into something other people pay you for once.

3. Sell something you make. A digital product. A physical product. A piece of software. A template, a notion doc, a course. Something you can ship that has your name on it.

4. Build an audience and monetise. Pick a platform. Show up daily. Two years in you've got something nobody can take away from you.

5. Build a tool with AI. This used to be the top of the ladder for technical people only. Now it's available to everybody. Vibe code a product, ship it, charge money.

Pick the smallest version. Start this week.

And here's the picture I keep coming back to. There's a guy in Lisbon with a folding table, a tablecloth, and some booze in a doorway. That's his entire “business”. He didn't wait for the perfect setup. You don't need the perfect setup either. You need the smallest possible version to start.

The single biggest blocker I see in entrepreneurs trying to start is "I'll launch when X is ready." The Portuguese booze seller has a folding table. You don't need a website. You don't need a logo. You don't need an LLC. You need to do the smallest version of the thing and find out if anyone wants it. Everything else is procrastination dressed up as preparation. It’s roleplaying business.

The New Safe Is Ownership

Doing the “safe” thing is now risky.

Experimenting with entrepreneurship and starting something is now the play.

Not entrepreneurship-with-a-capital-E. Not "quit your job and become a founder."

Just: own something, anything, even small. Even 5% of your income from something you own changes your psychology, your risk tolerance, and your options. The first £100 you earn outside your salary unlocks a different version of you.

Oh and if you want help working your precise next steps I built FreedomOS for exactly this. Diagnostic, framework, action plan. Sign up at Freedom OS.

Kyle

https://youtu.be/OQeWAfjRBWo

OpenAI dropped GPT 5.5 last week. Opus 4.7 landed a couple of weeks ago. Are they good? Yeah, for the most part. Does it really matter? Not as much as it used to.

The model wars stopped mattering around Christmas last year, we just didn't notice.

What genuinely matters now is the harness - the tools, memory, and workflow wrapped around the model.

That’s the new battleground.

And the story on that front this week is wild: Anthropic just admitted they quietly made Claude Code dumber for the whole of March. Sam Altman fired direct shots at them in the 5.5 launch.

GPT 5.5 Is Out. Don't Focus On The Model.

Discussed at [00:02:23]

The new model dropped last week. Their subtitle: "A new class of intelligence for real work." Note they didn't say coding. They said work. That’s a very specific and important choice.

The press release spells it out: researching online, analysing data, creating documents and spreadsheets, operating software, moving across tools until a task is finished. That's knowledge work. That's what most of you reading this actually do for a living…

I put out a, ahem, somewhat challenging video about this specifically that is upsetting a fair few people on social media:

@iamkylebalmer

ChatGPT 5.5 is specifically coming for white collar knowledge workers. They straight up are telling us this and people are ignoring what’s... See more

How do you actually find GPT5.5? Because it’s not immediately obvious if you are using it or not!

Inside ChatGPT: click the dropdown, you get Instant / Thinking / Pro, but the actual model numbers are hidden. Thinking and Pro are 5.5, Instant is still 5.3. Go to Configure to see them. Or grab it inside Codex, their agent harness (free on the $0 plan, which Claude Code isn't). Or via the API if you're a developer.

OpenAI are tired of people obsessing over model numbers. They hid them. The direction is clear: stop thinking about "which version am I on" and start thinking about "what job do I give it." The model is the engine. The engine used to be the bottleneck. It isn't anymore. I agree with them here. Let’s talk harnesses.

The Model Is Not The Thing. The Harness Is.

Discussed at [00:26:52]

Every week we get a new model. 4.5, 4.6, 4.7 Opus. 5.3, 5.4, 5.5 GPT. Blah blah blah.

The benchmarks go up a few percent. Twitter argues about whether the new one is actually better. A week later another one drops. It's exhausting and it's mostly noise.

Every benchmark result

What matters increasingly is the harness. The harness is everything wrapped around the model: the ability to use tools (Google Drive, Notion, your CRM), memory across tasks, the ability to iterate and check its own work, permissions and authorisations, multi-agent coordination. The model sits in the middle of this. The harness is everything else.

Good model + weak harness = useless. Google Gemini Pro on the API is brilliant on benchmarks. BUT nobody uses it for coding because the harness is weak.

Good model + strong harness = ChatGPT Codex

Amazing model + amazing harness = Claude Code.

The next twelve months of AI are not about which model is smartest. They're about which harness lets you do the most work. Claude Code, Codex, and Open Claw are the three that matter right now. Everything else is chat, and chat by itself is a solved problem.

Sam Altman Just Fired Direct Shots

Discussed at [00:17:52]

Read paragraph two of Sam's 5.5 launch tweet. It’s the important one here:

https://x.com/sama/status/2047379615589777666?s=12

"We believe in democratisation. We want people to be able to use lots of AI. We aim to have the most efficient models, the most efficient inference stacks, and the most compute. We want our users to have access to the best technology and for everyone to have equal opportunity."

That is not generic fluff. That is Sam firing directly at Anthropic, the makers of Claude. The same week Anthropic tested removing Claude Code from the $20 plan. The same week they got caught gaslighting about Claude quality - and finally admitted it. The same month they gated their next model - called Mythos - behind US government partnerships and a select list of enterprise customers.

OpenAI are positioning as the AI of the people.

And saying that Anthropic are only for the elites.

The comparison lands because it's true. Anthropic make 80% of their revenue from enterprise and API. They don't want $20 prosumer customers. They certainly don’t want free customers.

OpenAI have nearly a billion users, most of them free, and they're keeping it that way.

So…I’m with Sam here. His words here are backed by their actions - mass access to AI.

I loved the third paragraph less. "We love you" from a multi-billion-dollar CEO is giving weird energy. Unfortunately that’s also the one that everyone focused on instead of the genuinely fascinating second paragraph. Why you gotta be weird Sam? 😅 

What This Means If You Don't Code

Discussed at [00:00:34]

Lots of you will read this and think “well, I don’t code. I don’t use Claude Code, Codex or any of that sort of stuff”. This isn’t important for me. My job and livelihood are safe.

Anything that gets good at coding will be good at other multi-step knowledge work. The same skills - planning, using tools, checking its own output, iterating - transfer directly from "build this feature" to "restructure this sales page" or "turn this 64-page PDF of feedback into a completed revision by morning."

That last one isn't hypothetical. I gave Claude Code a 64-page feedback PDF from my business partner last night, told it to plan, implement, and verify the changes, then check its own work until done. It worked for three hours while I slept. Delivered a completed revision and a status report before I had coffee.

Previously that would have been my time. Or I would have hired someone. Now AI carries the task.

You can do this too. You don't need to be technical. Codex installs in ten minutes. It uses your existing ChatGPT subscription. You give it tasks in plain English. It goes and does them.

The people who've clocked this are quietly pulling ahead. EVEN IF you are not a coder this applies!

Kyle

https://youtu.be/2X6lV-b9LeA

OpenAI dropped ChatGPT Images 2.0 on Monday. Nano Banana Pro has been the uncontested image generation champion for about 10 months. I ran both through five real business tasks on the Livestream. Same prompts. Zero edits. What came out genuinely surprised me.

ChatGPT Images 2.0 won five out of five. Not by a little - by a lot on some of them. There's a catch at the end about price, which is why Nano Banana is still in the stack. But if you make marketing assets for a business, today is the day your workflow changes.

I was genuinely NOT expecting to be so impressed by Image 2.0. But hey, here we are!

The Setup

Discussed at [00:03:23]

I'm not covering this as an AI art tool. I don't care about photo-realism, I don't care about animation, I don't care about music generation. This newsletter is about using AI to run businesses. So I gave both models the kind of jobs a real business actually needs:

• A YouTube thumbnail

• A Facebook ad for a free course

• A four-panel explainer comic

• An event poster with a working QR code

• A full LinkedIn carousel from an existing article

Both models got the exact same prompt. Zero edits. Then one follow-up tweak on each to test sequential prompting.

That’s the setup. Let’s see how they did!

Test 1: YouTube Thumbnail

Discussed at [00:04:17]

I gave a prompt for a YouTube thumbnail, 1280x720, left two-thirds a shocked guy in a hoodie pointing right, right side a vertical split showing ChatGPT vs Gemini generating a menu, clickbait text "ChatGPT just killed Canva."

ChatGPT nailed it first time. Mr Beast shocked-pointing pose, clean mockups of the two interfaces on the right, text legible.

Strong for a first pass

Nano Banana also got it - pretty similar result actually. I preferred the ChatGPT version. Close fight on the initial render.

Perfectly serviceable

Then the real test: I added one follow-up prompt to both - "put him in RayBan sunglasses."

ChatGPT added the sunglasses. Everything else stayed exactly the same. Perfect sequential prompting. This is what we want.

Cool guy!

Nano Banana added the sunglasses and completely reformatted the image into a 9:16 portrait. It forgot we were making a YouTube thumbnail. Completely lost the original task.

Nanobanana flubbed it

KSequential prompting is where most image workflows live or die. You get something 90% right, try to tweak it, and the model rebuilds everything from scratch. ChatGPT Images 2.0 holds the context. That alone changes how usable these tools are for real work. Nanobanana crashed and burned on this one!

Test 2: The Facebook Ad

Discussed at [00:08:41]

Prompt: 1080x1080 square Facebook ad. Flat overhead shot on warm linen. Silver laptop showing a purple-and-white dashboard with a "Day 1 of 5" progress bar. Mug with steam. Tortoiseshell glasses. Leather notebook. Olive plant. Soft morning light. Bold navy headline "Get AI Ready in 5 Days." Pill-shaped CTA. Apple meets Morning Brew aesthetic.

ChatGPT followed everything. Olive plant - check. Linen pattern - check. Steam - check. Text legible, exactly as written. Composition felt like a premium ad.

Image 2.0 version - clean

Nano Banana got it technically right. Text correct. All elements present. But chose a high overhead angle that flattens the composition. Also snuck in a purple banner that looks a lot like a Google Form header, because it's Google and it defaulted to Google interface design.

Nanobanana - strange angle decision

Both are usable. If I'm running a real ad campaign I'd test both in paid traffic and let CTR decide. But as a first-pass output, the ChatGPT version looked more like a real ad I'd spend money running. The NB2 made a strange decision on the angle - but is otherwise mainly serviceable.

Test 3: The Four-Panel Explainer Comic

Discussed at [00:14:59]

Prompt was for a four-panel comic. Frazzled entrepreneur surrounded by ChatGPT, Claude, Gemini, Midjourney logos. Then panels resolve into them finding their AI stack. Specific wording in each panel.

ChatGPT got every logo right. Spelled "entrepreneur" correctly, which AI famously gets wrong (tbf so do I…). Clean layout. Text legible.

Clean

Nano Banana got most of it but put a weird line break through the Claude logo. Small AI artefact, but it’s the kind of thing viewers will clock. Also notice the blur on the guy’s head in 2 and 4.

Follow-up prompt: "put this in Marvel style." Let’s push it and see it content consistency can be maintained whilst changing style.

Not really “Marvel” but content consistency is solid

ChatGPT's reasoning was visible: "I can't use specific Marvel characters or logos since they're copyrighted. I'll apply half-tone shading, dynamic inking and energetic poses." Kept the exact same layout, same logos, same text. Applied stylistic overlay only. Massive engineering feat.

Nano Banana’s attempt duplicated the ChatGPT logo and the Claude logo without being asked (panel 1 and 3). Two extra logos appeared that weren't there before. And the text in 2 and 4 gets a little garbled.

Character and asset consistency across panels is the single hardest thing for image models. ChatGPT Images 2.0 seems to treat the original image as a genuine reference. That last 5-10% of consistency is what separates "ship it" from "spend another hour fixing it." You’ll often spend much more time here than the actual generations!

Test 4: Event Poster With Working QR Code

Discussed at [00:21:16]

This is where it got interesting.

Prompt was for a vertical event poster. Mediterranean sunrise over a rooftop, silhouette with laptop at the edge. Bottom right: functional QR code encoding aiwithkyle.com/join.

With a functioning QR! Very cool

ChatGPT's thinking chain actually showed me what was happening: "I'll create a QR code with Python, leveraging the library like qrcode. Afterwards I'll pass it along as a reference." It built the actual QR code first using a code interpreter, then embedded the finished code into the poster.

I scanned it with my phone. It worked. Opened my site.

Nano Banana looked at the same prompt, thought "they want a QR code, here's a QR code" - and hallucinated black-and-white squares that look like a QR code but encode nothing. I scanned it. Dead.

This is the paradigm shift, buried in one feature. Nano Banana treats an image as pixels to guess. ChatGPT Images 2.0 treats parts of an image as maths to solve. A QR code is maths, not pixels. It reasoned before it drew. That same capability will spread to charts, diagrams, tables - anywhere an image contains actual information. This is bigger than it looks. I’m not just getting excited about a QR code I swear!

Test 5: Full LinkedIn Carousel From An Article

Discussed at [00:26:36]

This is the one that genuinely blew me away. Here’s the carousel on Instagram so you can see it in action:

Instagram post

I gave it the URL to Monday's newsletter (the Opus 4.7 one, full article, images, everything). Single prompt: "convert this into a LinkedIn carousel, AI with Kyle branding." That's it.

ChatGPT went and read the article itself. Worked out what should be on each of 10 slides. Asked permission to generate the images. Then generated all 10 in one go.

The result was a complete, publishable carousel. Each slide had enough context to stand alone. The narrative flowed: problem, cynical theory, the real tokeniser issue, the auto-router problem, conclusion. The copy was short, visuals cohesive, and it pulled key numbers out of the article (the 35% tokeniser jump, the auto-router behaviour) and visualised them.

Nano Banana made 10 slides too. They looked nice. They said things like "Opus 4.7?" and "Sad robot" and "4.6 → 4.7" without any context. Cryptic quiz style. No standalone information.

OK but what does this MEAN???

Here is the actually comprehensible ChatGPT version:

Actual context included

LinkedIn carousels and Instagram carousels are now a solved problem. Completely. This is the single biggest business use case change in image generation this year. If you're running content for yourself or clients, you can generate a week's worth of publishable carousels in an afternoon. I'm building a LinkedIn carousel automation today. Because of course I am!

What This Actually Means For Your Business

If you were paying a freelance designer £200 to knock out Facebook ad creatives, that expense just evaporated. If you were spending hours inside Canva building YouTube thumbnails, that expense just evaporated. If you were outsourcing LinkedIn carousel design, that expense just evaporated.

Localised ads, hook-driven thumbnails, multi-panel carousels, event flyers with working QR codes - all of these just became a prompt.

BUT the tech is available to everyone. So the edge isn't the tool. The edge is the brand context you feed into the tool. Make sure you have at least a one-page AI Brand Guide - hex colours, typography rules, mood, what to avoid, text treatment - pasted into every prompt. To separate "AI slop" from "on-brand, usable output."

AI design just became a commodity. Hell, it has been for a while. But what you do with it isn't. Non-technical entrepreneurs who build the best prompting systems right now have a months-long head start on their entire market. Get your brand document in place. Build an automation to generate assets automatically.

Most people will read this, think "neat," and do nothing. Don’t be that person!

Kyle

https://youtu.be/k-r8addAtDI

Anthropic blinked. For about four hours yesterday they quietly pulled Claude Code out of the $20 Pro plan. George Pu spotted it, 2.3 million people saw his tweet, OpenAI's Codex team took a public swing, and Anthropic reverted the page worldwide.

That reversal is not a victory. At most it’s a reprieve. Most likely it’s a preview.

The flat-fee era of AI is ending, and your $20-a-month subscription is living on borrowed time. Sorry.

Anthropic Just Blinked

Discussed at [00:00:43]

Here’s the time line.

George Pu spotted Claude Code removed from the $20 Pro plan. No announcement. No email. Just gone from the pricing page. Only available on the $100 and $200 plans. The tweet hit 2.3 million views overnight on the tweet.

Twitter tweet

Amol Avasare, Anthropic's Head of Growth, posted a reply: "For clarity, we are running a small test on ~2% of new prosumer signups. Existing Pro and Max subscribers are not affected."

Twitter tweet

They were probably hoping 2M+ people didn’t see their “test”. OOPS.

Tibo Sottiaux from OpenAI's Codex team swept in and took the shot: "I don't know what they're doing over there. Codex will continue to be available on the free and plus plans. Transparency and trust are two principles we will not break even if it means momentarily earning less."

Twitter tweet

The next day Anthropic reverted the pricing page worldwide. The 2% test had leaked to everyone. Very publicaly.

Drama!

Kyle's take: The test itself kinda doesn't matter. What matters is that Anthropic ran it at all. They looked at their unit economics on $20 Pro users burning hour-long Claude Code sessions and decided it wasn't sustainable. More on this below.

They tried to do something about it. They got caught. They backed off. But the maths that made them try has not changed. They will 100% be pulling Code from $20 plans this year - I’m willing to bet on that one.

Codex meanwhile has been absolutely on fire the last month while Claude has been messy. I wouldn't be surprised if we're about to see a public sentiment swing back to ChatGPT.

What It Means For Your Plan

Discussed at [00:04:14]

Right now: nothing. It’s business as usual. BUT Anthropic will likely try to push this later when the dust settles.

If you're on $20 Pro right now, you're safe. Your plan still includes Claude Code. The access is relatively limited on that tier but it's there. I've got a private WhatsApp group where people who started on $20 a month quickly upgrade to $100 once they see how powerful it is.

If you signed up in the last day or two and something looks off, go check your plan. The 2% test caught some accounts mid-flight. The page was reverted but it's worth a look just to see if you got caught in the test. Unlikely but make sure to check!

If you rely on Claude Code for real work, you're probably already on $100 or $200 a month. Plan for that to stay your floor. Like Pro will go to $100 and Max will be $200. Everything shifted one to the right.

The $20 on-ramp is on borrowed time. Honestly, $100 a month for Claude Code is the best subscription I spend money on. I'm struggling to think of a better one…

Kyle's take: If you're running a business on Claude Code right now, use the next six months to hedge. Test Codex in parallel this week. Set up Ollama on a spare machine. The people who get burned will be those who bet their entire pipeline on one subscription and never hedged. There are PLENTY of warning signs right now.

"Intelligence Too Cheap To Meter" Was A Story

Discussed at [00:05:41]

Sam Altman said this around the release of 4o-mini, a couple of years ago. He said it at a Fed speech. He repeated it at the BlackRock Infrastructure Summit in March. Every AI keynote recycled the line for a while.

Twitter tweet

Here's what's true: per-token prices have collapsed. Roughly 280x drop in AI token costs over two years. That is a massive drop. Intelligence on a per-token basis genuinely got cheaper.

https://www.wisdomtree.com/investments/blog/2025/05/19/280x-cheaper-the-real-ai-revolution-is-accessibility

Here's what's not: your per-user bill is going up. Anthropic just tried to push Pro users from $20 to $100. That's a 5x jump. The reason is simple: two years ago we were chatting to AI. Back and forth, a few hundred tokens per session. Now we're running Claude Code, Codex, Claude Design, agents that chat to themselves for 20, 30, 40 minutes at a stretch, burning through tokens in a much greedier way than a human ever could. I've got three agents running in the background right now while doing this stream.

This is Jevon’s Paradox. When something gets cheaper, people use dramatically more of it. Cheaper electricity meant more electricity consumption, not less. Cheaper tokens means more tokens burned, not less. Despite the efficiencies MORE is used. Both things are true at the same time.

anations.com/jevons-paradox

The "too cheap to meter" line was marketing, not maths. It was always going to be marketing. Your 2026 reality is that per-user bills are going up because you're using these tools 1000x more than you were two years ago.

Altman recently (March 2026) adjusted the narrative to this too - AI as a utility with “pay as you go” pricing.

Twitter tweet

Two Costs. Only One Is Falling.

Discussed at [00:09:35]

Every AI company has two cost lines and you need to understand both.

Training is the cost to build the model. Think setting up the kitchen, hiring the staff, building the restaurant. You pay it once per model, then amortise across every future query. Frontier models now cost $100 million plus. Next-gen is estimated in the billions. Direction is up as models scale. But they are one time costs.

Inference is the cost to run the model. Think ordering from the menu - food going back and forth every single time. Inference scales with usage. As more people use AI more intensively, the inference bill balloons. It's now 80 to 90 percent of all AI compute load as of 2026. Agents running for hours multiply this by 1000x per user.

In the recent fracas Amol Avasare actually said the quiet part out loud: "Our current plans weren't built for this." He is straight up saying that this is unsustainable. This feels like CEO level communication rather than something to drop in a random tweet but hey at least we got the truth!

Basically nobody priced a $20 subscription assuming one user would run an agent for eight hours a day. That's what's happening now.

Twitter tweet

This is why flat-fee pricing will die. It's not corporate greed. That’s too simplistic. Instead it's maths. When your top users consume 100x more inference than your median user, a single subscription price stops working. Light users subsidise heavy users until the company notices, runs the numbers, and starts testing what a realistic price looks like. That's exactly what Anthropic just did - and got caught doing! BUT expect every major AI tool to add usage caps or tiered metered billing within six months.

Anthropic and OpenAI originally priced AI like Netflix. $20 a month, all-you-can-eat, feels about right. That was a marketing decision, not a pricing one. And they probably just plucked $20/month from the air because that felt right. Now the industry is maturing and realising light users can't subsidise heavy users forever.

So the pitch is changing. They're reframing AI from "SaaS subscription" to "employee cost". Claude Design isn't $1000 a month compared to Photoshop at $20. Claude Design is $1000 a month compared to a $5000-a-month designer. Totally different pricing anchors and market. Suddenly the price looks reasonable. THAT’s the transition these companies are trying to make.

So What Do You Actually Do?

Discussed at [00:14:05]

OK that’s all well and go. We understand WHY. But what do we actually DO?

Right now: try ChatGPT Codex. It's good. Genuinely competitive with Claude Code. The Codex team has been absolutely on fire for the last month while Claude has been messy. OpenAI just publicly committed to Free and $20 tiers with advance notice of changes. They have the compute. The limits are generous - and they keep capitalising on Anthropic’s missteps by resetting Codex limits…

It's become a meme. There's literally an image of him resetting the limits…

Twitter tweet

6 to 12 months: go local. Go open source. Qwen Coder, GLM, MiniMax and Kimi are closing the coding gap. Run them via Ollama or LM Studio. LM Studio is free, dead simple - click, pick a model, it downloads and runs. It’s not technical - it’s all plug and play nowadays.

Always: make do with less. Token hygiene wins. Default to smaller models like Haiku or GPT-5 mini for routine work. You don't need Opus 4.7 for a tweet draft. Turn on prompt caching so your system prompt isn't re-billed every turn. Feed the model the specific file, not the entire repo. Just because you have a 1 million token context window does not mean you should fill it. Write specific prompts - for examples 4.7 needs a better prompt than 4.6. Build a library of saved Skills.

Kyle

https://youtu.be/uqq4yaunOoE

Anthropic dropped Claude Design on Friday. Figma's stock fell 7% the same day. The "Figma killer" headlines started writing themselves. Your social media is probably FULL of hyperbole about Claude Design killing web designers. RIP.

As always…the truth is a little more complex. And boring. Ha!

I’ve burned my entire weekly $200 Max usage in four prompt. Got kicked out till Sunday. And every single output looked like it was made by the same intern at Anthropic…

So…some issues.

RIP Figma?

Discussed at [00:01:45]

Claude Design showed up as a shadow drop on Friday afternoon. New icon in the sidebar, new URL at claude.ai/design, powered by Opus 4.7, Canva integration built in from day one. Available across Pro, Max, Team and Enterprise.

Twitter tweet

Figma's stock dropped 7% the exact same day. Worth noting: Anthropic had a rep on Figma's advisory board until a few days ago. They pulled out right before this announcement because it’s a competing project. This is genuinely a head to head.

The market panicked. AI-Twitter called it the Figma killer. "RIP designers" posts started trending. Even your mate who's never opened Figma had opinions.

Chill. Absolutely chill. This is the usual overblown AI launch reaction. The people calling this a Figma killer don't use Figma.

If you're a real designer, you're fine. Well…no, your in trouble. But not just because of this launch!

But if you're an entrepreneur (like we are!) who's avoided design tools your whole life, this is actually interesting. More on that in a minute.

What It Actually Is

Discussed at [00:03:45]

This is not Lovable. It's not Bolt. It's not v0 or Replit. Claude Design is not a full-stack builder. You can't deploy an app from it. There's no backend logic.

It's also not Figma. No vector tooling. No complex component library workflows. A real designer would use it for about four minutes and then close the tab.

What it actually is: a prototyping tool for decks, one-pagers, landing pages, and marketing graphics. Think Canva, but with a brain that reads your codebase. That last bit is the interesting part.

The main setup interface

The category matters here. Once you stop comparing it to Figma and start comparing it to Canva or Squarespace, it suddenly makes sense. This is a founder and solopreneur tool. Not a designer replacement. Anyone telling you otherwise is selling something.

It's Not a Designer Replacement.

Discussed at [00:19:54]

Before web builders like Squarespace, having a decent website cost you £10k or three months of your life fighting WordPress plugins. After Squarespace, every solopreneur had a reasonable-looking site by Friday afternoon. The floor went up.

Claude Design is that moment, but for prototypes. If you're trying to validate an idea, pitch a product, or mock up an app for your first ten users, you can ship something that looks fine in an afternoon.

Too many entrepreneurs get stuck on "I need a nice-looking website" and waste six months and £5k before they've tested whether anyone wants the product. Tools like Claude Design let you skip over this. Use it to validate. Then when you actually have revenue and the idea deserves real investment, hire a human designer. The decision tree is simple: v1 prototype? Claude Design is perfect. Real brand that needs to stand out? Human.

Three Ways to Steer (and the Token Bill)

Discussed at [00:06:47]

Inside the canvas you've got a number of ways to move the AI. Chat for broad conceptual shifts like "make it darker" - much like how we prompt any AI.

Inline comments where you click directly on an element and tell it what's wrong, exactly like working with a real designer.

And a drawing tool to allow you to directly “mark up” your page or app and give thoughts.

The comment system is probably the most powerful. You invite teammates, they click on specific bits of your design and leave notes, and it all flows back into the chat. Very similar workflow to working with a human designer honestly. But infinitely more patient.

Now the bad news. This thing burns tokens like it's personally angry at your wallet. I'm on the $200-a-month Max plan. I prompted it once this morning to build an app mockup. Got told I'd hit both my daily and weekly limit. Come back Sunday - sowwy.

The only saving grace is that Design limits are separate from Code limits. If they shared a pool, I'd tell you not to touch this thing as it’ll wreck your week!

As it stands, play around, but don't plan any real work on it until the limits relax. Which they will. Probably. Please?

Reworking projects

Discussed at [00:07:57]

The thing that actually makes Claude Design worth using: you can point it at your GitHub repo. I gave it the repo for aiwithkyle.com yesterday. It read every HTML file, every CSS rule, worked out my colour palette, typography, and spacing, and built me a full design system automatically.

Once it has that context, redesigns become genuinely useful. It gave me three redesigns of my homepage: one called "Practical Skills Without the Hype", one called "Playbook First", and one called "Live and Loud" that foregrounded the daily livestream. That last one I actually liked structurally.

Foregrounded Livestream hero

Without the codebase link, it's just generic. With it, it becomes something closer to an actual design collaborator. Big difference - so I definitely recommend hooking it up to your existing work.

Give it your existing brand. Otherwise you're getting the same generic output as every other person who typed "make me a landing page" into it today. Which brings us to the elephant in the room.

Purple Gradient 2.0

Here’s the worst part.

Every single output from Claude Design has the same aesthetic. Off-white cream background. Terracotta or burnt orange accents. Same card layouts. It is visually impossible to tell two Claude-designed sites apart.

Why? Go look at Anthropic's own website. That cream. That terracotta. That's their house style. They trained Claude heavily on their own design language, and now it bleeds into everything you make. Your brand becomes Anthropic's brand whether you asked for it or not.

Now: I like their style. It’s great for them. Good job.

Do I want every one of my projects to look like an Anthropic site? No.

I told it explicitly "do not use this off-white cream, I hate it." It ignored me. Next iteration: cream. "No, really, different colour please." Slightly different cream. It kept defaulting back every single round.

Here you go idiot, hope you like cream.

This is going to be the defining visual tell of 2026. Three years ago every AI-generated image had a purple gradient. Two years ago it was glassmorphism everywhere. Now it's cream-and-terracotta editorial slop. If your landing page has it, people are going to clock it the same way they clock ChatGPT writing that uses "delve" or em-dashes.

How to Push Past the Default

Discussed at [00:12:05]

Three things make the difference.

First, stop saying "I don't like that blue". That's a child complaining. Say "drop the saturation on the primary, swap the accent from terracotta to deep navy, tighten the line height on the h1." Talk to it like you'd talk to a real designer. Specificity is the whole game. Don’t know how? Take your design out of Claude Design and chat to an AI about how to provide feedback.

Second, reference like a pro. Give it URLs of sites you love. Julian’s site for whitespace. Linear's for type hierarchy. The top menu of Barstool sports. Whatever matches the vibe you want. Screenshots work, or just give it the links.

Third, use Rohit's skill pack. There's a DESIGN.md setup on GitHub (github.com/rohitg00 and github.com/VoltAgent) that injects nine different design systems directly into Claude's context. Editorial minimalism. Terminal core. Warm editorial. Anything to get you away from the one default cream palette. At least this will give you 10 new defaults! A little better.

None of this is new advice. It's how you've been working with AI for years with writing and code. The same rules apply to design. Specificity beats description. References beat adjectives - don’t just tell it “more blue”.

Does It “Kill Designers”?

For designers who already use Figma, no. Claude Design doesn't do vector work, doesn't do complex component libraries, doesn't play nicely with Figma's Make feature or Code to Community. The Venn diagram of "Figma users" and "Claude Design users" barely overlaps. Most of the people screaming RIP Figma haven’t used it.

The real answer though is uglier. 'Good enough' design is now a free commodity. Well…when they sort their limits lol. Average is no longer a differentiator. The drag-and-drop sticker-sheet designers who were coasting on churning out passable Canva templates for £500 a pop? They're in trouble. Hell, they were already. This just accelerates their decline.

But the designers with actual taste, restraint, and specificity? More valuable than ever. The AI cannot do restraint and taste.

Every AI tool like this follows the same pattern. It raises the floor and the ceiling at the same time. The middle gets squeezed. If you're a designer who's been charging for execution alone, start adding taste and judgement to the pitch. Execution is now a few prompts. If you're a founder hiring designers, keep hiring them for the work that needs to stand out. Use Claude for the v1 mockup you're going to throw away anyway.

What This Means For You

Discussed at [00:18:14]

Pretty simple decision tree here. Are you building a new idea? Do you need a v1 to validate fast? Use Claude. Ship a deck or prototype in an afternoon and get moving. Does the idea need to stand out and scale? Hire a human designer.

Use Claude to validate. Use a human when it deserves real investment. That's kinda the whole framework - nice and simple.

Remember you don't need a perfect logo before you have customers. You don't need a beautiful website before you've tested your offer. Claude Design is a tool for getting out of your own damn way. Use it for that. Get moving.

Kyle

https://youtu.be/8ckFZo9seNg

This is a tricky one.

I want to start by saying I love Claude. I use it every day. I pay Anthropic $200 a month and it's the best business expense I have.

I’m a fanboi. And have been for years.

That's the context for what I'm about to say.

I think Opus 4.7 is a stitch-up. This is a bad launch. And worrying for the future of AI as a whole.

Again, love you guys but…yeesh.

The bait-and-switch theory(?)

Over the last month, something weird has been happening with Opus 4.6. I noticed it. Power users on Hacker News and Reddit noticed it. The model seemed to get dumber. Responses used more tokens to say less. Limits that used to be generous started biting within hours.

Twitter tweet

Normally when people complain about this sort of thing it feels like noise. We’re on a hedonic treadmill and constantly expect more for less. I’m one of the first to dismiss it as noise to get clicks.

But I felt it this time too..which makes it hard to ignore. As have many smart people I’ve talked to.

I'm personally on the $200 a month plan. I've never hit limits on Claude before. This month I started slapping into them on Opus 4.6 for the first time. I wasn't doing anything new.

Then last week Anthropic drops Opus 4.7. The community's read on what actually shipped? This is basically the original 4.6. The one from a couple of months back that felt like a step towards AGI. But rebranded so that we accept the tigther limits.

I can't prove any of this. Nobody can. But it no longer feels like that wild of a speculation.

If the cynical reading is right, Anthropic degraded a product users were paying for, then basically re-released it as a new product. But one that uses even more tokens…

The new tokenizer (up to 35% more input tokens)

OK this one isn't a theory. It's in the official migration guide.

Instagram post

Anthropic says Opus 4.7 uses an updated tokenizer. The trade-off, in their own words: "the same input can map to more tokens, roughly 1 to 1.35 times."

From the official launch blog

Read that again.

The headline pricing per token hasn't changed. They repeated that loudly. But if the same chunk of text now counts as 35% more tokens, you're paying 35% more for the same input. Same price per token is meaningless when the token count goes up.

Plus it’s sneakily dropped in as a lower line item. Below the headline pricing per token.

Now, the charitable read is “ah well but the model will be able to solve problems for you much more effectively so the increased tokens doesn’t matter”. Maybe. That doesn’t seem to be the consensus so far - but we are early.

Adaptive reasoning (the auto-router problem)

Opus 4.7 got rid of extended thinking. In its place: adaptive reasoning.

Same thing right? Nope!

Adaptive reasoning is an auto-router. You toggle it on and the model decides for itself how much effort to apply to your question. Light thinking, medium, hard. You don't get to pick.

We've seen this movie before. Six seven months ago, GPT-5 launched with exactly this design. They were trying to make ONE unified model. Users hated it. The auto-router made stupid calls, and because there was no manual override, power users were stuck with whatever the router decided. Within a week, OpenAI backtracked and bolted the model picker back on.

Anthropic has now copied the mistake.

Ethan Mollick said it best on X - adaptive reasoning is bad in the ways all AI effort routers are bad, magnified by the fact there's no manual override.

Twitter tweet

Expect a U-turn within a couple of weeks.

MRCR and long context

This is feeling like a pile on but I’ve started so I’ll finish.

MRCR stands for Multi-Round Co-reference Resolution. It's a benchmark that tests how well a model can find specific information inside a massive context window. Context window is basically (but not really) the memory the model has to work with - the larger the context window the more “stuff” you can throw at it. Useful if you care about the 1 million token window Opus keeps advertising.

Opus 4.6 scored 92% on MRCR. Better than Gemini 3.1 Pro. Better than GPT-5.4 High. Genuinely best in class.

Opus 4.7 scored 59.2%. That's a 30-point drop.

Normally number go up. Why no number up?

MRCR wasn't on the main announcement page. It got buried in the systems card. When someone called Anthropic out on it, the head of Claude Code basically said the benchmark isn't something they're focused on any more.

That's one of three things. Either they trained against the benchmark before and don't want to now. Or the benchmark doesn't matter. Or 4.7 is genuinely worse at long context than 4.6.

Any of those options should worry you if you work with large codebases, long documents, or 1M-token projects.

The big context window was the reason to use Opus. If 4.7 can't actually use that context well, the pitch falls apart. And quietly removing the benchmark you were winning on, right before dropping the new model, is not a good look.

Now there may be a genuine reason for this. As always: USE the model and see what you personally think. Does it work for you? What results do you get.

Let’s get real: you're not the customer

I’ve said this before. And I’ll say it again.

Anthropic isn't optimising for people like us. We pay $20, $100, maybe $200 a month. That's peanuts compared to enterprise contracts and API spend.

About 80% of Anthropic's revenue comes from enterprise and API. Not consumer subscriptions. They're heading into an IPO. They need to show profitability. Individual subscribers are cost centres, not profit drivers.

They don’t want us stinky consumer peasants!

So when you complain that your limits tightened, that the tokenizer changed, that the auto-router took your control away, it doesn't really matter. Fortune 500 companies running API workloads don't care. That's who Anthropic is designing for.

OpenAI has the opposite problem. Almost a billion users, most of them free, bleeding cash. They're trying to pivot to enterprise too, which is why Fidji Simo has been killing side projects like Sora.

This is what Anthropic foresaw and avoided. Anthropic's already there. That's why their run rate is closing in on $30 billion while OpenAI hovers around $25 billion.

Twitter tweet

Now this is not malicious. It's a business. But if you're paying $20 or $100 a month and wondering why things feel worse, this is the reason. Anthropic doesn't hate you. You're just not their customer. Soz.

How to survive the squeeze

Intelligence was supposed to get too “cheap to meter”. It still will, eventually. Hopefully…

But not this year, and probably not next. GPU shortages, RAM prices, data centre bottlenecks, investors wanting returns before IPOs - the short-term direction is more expensive, not less.

Here's what you can actually do to keep working with AI:

1. Use Sonnet as the default. Opus is for planning, architecture, and hard problems. Sonnet is for implementation. If you're on the $20 plan, you probably shouldn't be using Opus at all beyond a few turns a day.

2. Architect in Opus, build in Sonnet. Spend your Opus turns on the plan. Let Opus write detailed, step-by-step instructions. Then hand the plan to Sonnet for the actual work. You'll get 90% of the quality at a fraction of the cost.

3. The caveman method. Stick this in your system prompt: "Respond like a caveman. No superfluous words. No long explanations. Just the output." Sounds daft. Works. I've seen people cut their token usage by 30-40% with one line of config. Or use this skill: https://github.com/JuliusBrussee/caveman/tree/main

4. Codex is genuinely catching up. ChatGPT Codex has more generous limits than Claude Code right now. If you keep hitting ceilings in Claude, use Codex for implementation and keep Opus for planning. They also just shipped a Cowork-style tool where Codex takes over your computer.

5. Local models for implementation work. MiniMax and Qwen 3 run cheaper than the Claude API and are good enough for most coding tasks. The trick is using the state-of-the-art model to write the plan, then letting the cheaper model execute.

6. Cut the Ralph loops, the Obsidian second-brain setups, the recursive self-improving agent loops, the auto-looping workflows on Twitter - they burn tokens like nothing else. Lovely to read about. Expensive to run. Skip them for now unless you have cash to burn.

The days of paying $20 a month for unlimited access to the best AI are gone. Not coming back for a couple of years. That's the reality. So either find workarounds or pay up. If you're running a business on this stuff, $200 a month is nothing. But if you're a hobbyist, Sonnet plus caveman mode is your friend.

Also…can’t overstate this enough: Build now. While the models are still good, while the limits are temporarily loose for the launch window, while the window to use this stuff at scale is still open. The cost isn't going down this year. It’ll get worse.

Member Q&A

"Can I use 4.7 for oversight and 4.6 for context management?" - Yes, and that's actually a clever setup. Let 4.7 supervise the high-level decisions and 4.6 handle the long-context work it's still better at. Worth testing.

"Are local models still worth it?" - Yes if you have the hardware. No (maybe) if you're buying from scratch. RAM prices and GPU costs are climbing fast. If you already own the machine, run MiniMax or Qwen. If you don't, rent GPU space from a provider.

"Is Codex a solid alternative?" - It is now. I switch between Codex and Claude Code depending on the task. Codex has more generous limits and their new Cowork-equivalent looks promising.

"What about Manus?" - I check in occasionally. I don't see the point when Claude Cowork and Codex do the same things natively. Some people love it for research. Your mileage may vary.

Kyle

https://youtu.be/RG04qMlZz_s

You've seen the viral posts. Someone shares their Claude Code setup - custom skills, Claude.MD, MCP servers, a second brain in Obsidian, automated workflows running overnight. And you think: "What on earth are they talking about??? I'm already behind."

You're not behind. Most people aren’t even seeing this sort of stuff.

You just haven't started yet. It’s all good. Starting just got stupidly easy.

Anthropic released a redesigned desktop app for Claude. Not the web version you use in your browser. Not the mobile app. A proper desktop application for Mac and Windows that puts Chat, Cowork, and Code all in one place. They did a big overhaul.

Why should you care?

Claude Code now has a built-in preview pane. You can see what you're building - live - right next to the chat. If you've used Lovable, Replit, or Bolt, you know the experience. Type what you want, watch it appear. This makes Claude Code a lot more accessible. And if you’ve been scared away by the technical complexity this is the time to hop in.

Previously, you had to set up a local development server manually. That meant opening the terminal, running commands, dealing with error messages that look like a foreign language. For anyone who doesn't code, that was the wall. It looked terrifying. Most people bounced off it.

That wall is gone now. Good riddance I say.

The "I'm not technical enough" problem

Here's what I think is actually happening. You hear people talking about their CLAUDE.md files, their MCP servers, their custom Skills installs, their Obsidian second brains synced via GitHub. And it sounds like you need a computer science degree just to get started.

You don't.

All of that stuff is advanced. It's what people build up to over weeks and months. It's like watching someone parallel park a lorry and deciding you can't learn to drive a car. As an afficionado of Euro Truck Simulator (yes, it’s a thing…) I can assure you parallel parking a 16 wheeler is difficult.

The people sharing those setups started exactly where you are - with an empty folder and a vague idea.

The actual starting point is embarrassingly simple.

1. Download the Claude desktop app

2. Click "Code" in the left sidebar

   

3. Click "New Session"

4. Choose "Local"

   

5. Pick a folder (make a new one on your desktop - it doesn't matter where)

   

6. Tell Claude what you want to build

That’s it!

I covered this one the livestream yesterday if you need an exact blow by blow. See the video up top.

But that's it. Six steps. No terminal. No GitHub. No configuration files. No skills. No MCPs.

Can you add that stuff later? YES! Should you? Absolutely!

But first keep it super simple.

The three Claudes (and why it's confusing)

The left sidebar now has three icons. They're small and aesthetic and honestly a bit hard to spot. Here's what they are:

• Chat - This is normal Claude. The chatbot you know. You talk, it responds, you copy-paste what you need. It's for thinking, planning, and discussing.

• Cowork - This is Claude doing work FOR you. Scheduled tasks, computer control, automations. Think of it as an employee you can give recurring jobs to.

• Code - This is for building. Software, websites, apps, tools. It reads your files, writes code, runs commands, and now shows you a live preview of what it's creating.

Chat is for talking. Cowork is for working. Code is for building.

Claude Chat, Claude Cowork, Claude Code is a naming disaster. Everything starts with C.

They are even working with Microsoft right now to release Copilot Cowork!!

Heretics…

But once you use them, the separation makes sense. Most people should start in Chat to plan what they want, then move to Code to actually build it.

The preview pane

This is new. And a BIG shift for beginners.

When you're in Claude Code, there's a small dropdown in the top right corner. Click it, hit "Preview", and a panel opens on the right side of your screen.

Preview. AT LAST

It shows you - live - what Claude is building. Every time it makes a change to your code, the preview updates. You can click on elements in the preview to tell Claude what to fix. You can give feedback by voice or text and watch it adjust in real time.

Key features of the preview:

• No terminal setup - Claude handles the server for you

• Click-to-select - Point at something in the preview and say "change this"

• Live updates - See changes as Claude writes them

• Responsive testing - Toggle between mobile and desktop views

This is what tools like Lovable have been doing from day one. Immediately showing you what you are building and changing the preview live as you work with your AI.

The difference is Claude Code gives you the actual files on your computer. You own what you build. Nothing's locked behind a platform. Oh, and platforms like Lovable use Claude behind the scenes anyway…and add a markup. So….probably makes sense to go to source.

Permission modes (or: how much do you trust the AI?)

Just below the chat pane, bottom left you’ll see a selector for Permission:

When Claude Code runs, it asks permission before doing anything new. There are four modes:

• Ask Permissions - It asks before every single action. Safe, but you'll be clicking "yes" constantly. Good for anything high-risk. But otherwise very tedious.

• Accept Edits - It goes ahead with file changes but asks before new actions (like installing software). This is the sweet spot for most people starting out.

• Plan Mode - It doesn't make any changes at all. Just asks questions and outlines what it would do. Brilliant for the planning phase before you start building.

• Bypass Permissions - The old "YOLO mode". It just goes for it. Risky on anything important, but fine for personal projects on your own computer. Depending on how risky you want to get obviously!

Start with Accept Edits. Move to Plan Mode when you're setting up a new project. Use Bypass only once you're comfortable and the stakes are low.

I know people will yell at me for this, but on personal projects - a portfolio site, a side project, a tool just for you - Bypass is fine. You're not going to break anything that matters. As you get more comfortable, you'll naturally figure out when to be more careful. The bigger risk is being so cautious you never actually build anything.

I personally YOLO Claude every evening when I go to bed.

NOT on critical projects! And nothing public! But otherwise? Yeah, have at it! Claude is a better coder than me so who am I to clip his wings.

What about pricing?

Claude Code is not cheap.

Real talk. The free or $20/month plan won't cut it for Claude Code. You'll run out of tokens too fast to build anything meaningful.

• $20/month - You get access but you'll hit limits quickly on any real project

• $100/month (Pro) - Solid for regular use

• $200/month (Max) - What I use. Generally unlimited enough for daily building

If you're building a business, the $100 or $200 plan pays for itself almost immediately. I struggle to think of a better investment. I've built tools in an afternoon that would have cost thousands to outsource.

But start with $20 to test the waters. See if you actually use it. Upgrade when you hit the ceiling, not before.

Beyond code - this is bigger than building websites and apps

Claude Code has Code right in the name. It must be for coders right? Nope!

The pattern it uses - give it context, let the agent work, get output - that applies to everything.

Content creation. Business operations. Research and analysis. Script writing. Any knowledge work where you'd normally sit at a computer and produce something, Claude Code can help with.

I use it for my entire business. Not just building my website. Drafting newsletters, generating scripts, creating PDFs, managing my project files. Brainstorming and deciding pricing strategies. Bookkeeping, customer support. Everything. The "Code" name is misleading. It's really an AI workspace that happens to be really good at code.

This is where it gets exciting for entrepreneurs. This is not about coding. Sure, you’ll be using code. But you're learning to work with an AI that can code, write, research, and automate - all from one interface. The code itself is sorta inconsequential honestly. It’s a means to an end.

Member Q&A

"Does it work in the terminal?" - Yes, 100%. Terminal gives you more power and flexibility. But if you've never used the terminal, the desktop app is the way to start. You can always graduate to the terminal later.

"What's the difference between Claude Code and Claude Cowork?" - Code is for building things (software, websites, tools). Cowork is for automating things (scheduled tasks, computer control, recurring processes). Code creates. Cowork operates.

"Best practices with large code bases?" - Claude Code now has a 1 million token context window, which is pretty huge. But you still don't want it looking at your entire codebase at once. Work module by module. If you're changing the homepage, you don't need it reading the database code. Focus it on what's relevant. This is good practice regardless of whether we use AI or not.

Kyle

https://youtu.be/KQrZWI8K934

For the last two years, most of us have been using AI the same way: open a chat, type a question, get an answer, copy-paste it somewhere.

That's fine. But it's level one.

Today I walked through what level three looks like - and Anthropic just made it a LOT more accessible than it used to be.

The word is orchestration. And if you're running a business or trying to build one, this is the skill that turns you from a one-person operation into a one-person operation with a team.

Twitter tweet

From Prompting to Orchestration

Discussed at 00:00

Here's the progression:

Prompting - You talk to AI. It answers. You copy-paste the output. This is where 99% of people still are. It works. But you're doing all the driving.

Context Engineering - You get smarter about feeding AI the right files, the right memory, the right project context. MCP servers, skills, custom instructions. All that good stuff. The AI gets better because you give it better inputs to work with.

Orchestration - You design a system of specialist agents, manager agents, and review gates - all working in parallel. You're not doing the work. You're not even directing the work in real time. You're building the system and hiring the staff. Then you walk away.

This is new.

Lots of people talked about this in 2025. But it’s finally a reality.

The Restaurant Analogy

Discussed at 03:33

The difference between using AI normally and orchestrating agents is the difference between being a customer in a restaurant and running the kitchen.

Most people walk in, sit down, get served, and leave. That's fine.

But if you want the next level, you need to get into the kitchen. Design the recipes. Manage the chefs. Define the tools. Set the rules. Let the infrastructure run while you sleep.

You don't cook every dish. You’re too busy for that. Instead you build the system that produces great dishes consistently. That's basically orchestration.

Cowork vs OpenClaw vs Managed Agents

Discussed at 05:22

This is where people get confused. There are now several ways to run AI agents. They're not the same. And honestly this is all a moving target. There is a lot of overlap. A lot of cribbing each others’ notes. A lot of change.

So I’ll focus on the big “names” for now to help you orientate.

Claude Cowork - An agent on your desktop. It controls your computer, opens windows, fills in forms. Good for sequential tasks. Falls apart if you try to do more than one thing at a time because the agents trip over each other. And it ties up your machine while it's running.

OpenClaw - An agent in your chat. Runs on a dedicated machine (Mac Mini or VPS). Can reach out to you proactively via WhatsApp or Slack. Works well for persistent tasks. But scaling it up or giving other people access gets fiddly fast.

Anthropic Managed Agents (new) - Agents in the cloud. Always on, even when your devices are off. Can run hundreds in parallel. Built for multi-user access. This is the business-grade option - and it just got a lot easier to set up.

Each has its place. Cowork is great for quick desktop automation. OpenClaw is brilliant for personal always-on assistants. But if you're building something that needs to scale, run 24/7, or serve multiple users - managed agents are where you should be looking.

I use them all. Different tools for different tasks.

The Quick Start Nobody Knows About

Discussed at 07:28

Here's the thing that tripped me up when I first looked at Anthropic's agent platform. If you Google "Claude agents", you get the developer docs. Terminal commands, SDK setup, config files.

This is a screenshot from the “quick start” page:

If you're a developer, that's fine. If you're not, you're immediately thinking "this is not for me." And you’ll bounce off.

But Anthropic quietly released a graphical interface inside the managed agents platform. It's a chat window. You just type what you want your agent to do and it builds the config for you.

It looks like this:

Yay! A chat window!

Go to platform.claude.com. Find Managed Agents in the left sidebar. Click Quick Start.

That's it. You describe your agent in plain English. It generates the setup. You click Create. You've got a managed agent running in the cloud.

I built a news scraper agent on stream in about two minutes. It wasn't perfect - Twitter scraping is always a pain - but the point is how accessible this has become. There's also a Guided Edit feature that lets you iterate on your agent through conversation.

This is a signposting problem from Anthropic. The Quick Start GUI is genuinely easy to use but it's buried behind developer docs that scare off non-technical users. If you've been avoiding the agent platform because it looked too complex - go look again. It's changed!

The Anatomy of an Agent

Discussed at 20:10

What you're actually building:

The Brain - Claude (Sonnet, Opus, or Haiku depending on your task and budget). This is your AI model sitting in the middle, doing the thinking.

The Sandbox - A cloud environment on Anthropic's servers. A little walled-off computer dedicated to your agents. It runs when you're sleeping. You control what goes in and out - so your agent can't go rogue and start posting to your social media.

Tools, MCP, and Skills - Everything you've learned in the Anthropic ecosystem applies here. MCP servers for connecting to external services. Skills for specialised capabilities. Files for context.

The sandbox is the bit that makes this different from just chatting to Claude. Your agent has a persistent home. It can store files, remember context, and work on schedules. It stays “on” even when you aren’t around.

Multi-Agent Orchestration

Discussed at 21:55

This is where it gets interesting for business.

You build specialist agents for individual tasks. Then you build a manager agent that oversees them all. The manager checks the work, sends it back if it's not good enough, and only passes the final result to you when it meets the standard.

Think of it as literally replicating a human team structure. You've got specialists doing the detail work and a manager making quality decisions. The loop runs until the orchestrator decides the output is good enough. Then it bumps the output over to you the human for final sign-off.

If you've ever managed people, those skills transfer directly. What does this person need to know? What's the quality bar? When do I escalate? Same questions, different team members. It’s just when you provide instructions and training you are talking to an AI not a human.

Someone asked me how to get better at thinking in systems. Some people naturally do it. Others are more detail-oriented.

My honest answer: play Factorio.

Look at this beautiful mess!

I know that sounds bizarre. But Factorio is a factory-building game that teaches you supply chains, logistics, input-output thinking, and how modules connect - in an incredibly intuitive way. You'll probably learn more about systems thinking in 20 hours of Factorio than in most MBA courses…

For a more traditional (less fun!) route - "Thinking in Systems" by Donella Meadows is the book on this topic. Short, clear, and widely considered the primer.

Orchestration is a different (complimentary) skill to prompting. Prompting is about crafting the right questions. Orchestration is about designing a system that doesn't need you to ask questions. If you naturally think in systems, you'll take to this fast. If you don't - start building that muscle now. It's going to matter.

The Entrepreneur Opportunity

Discussed at 27:04

Two ways to use this if you're building a business:

Build agents into your products. If you're selling a service or SaaS, managed agents can be the brain. I gave the example of company formation on stream - you could build an agent system that collects documents, checks for missing info, compiles everything, and only puts a human at the final review stage. That's a process someone is paying thousands for today.

Some examples that would apply to my business

Use agents for your operations. Marketing, bookkeeping, customer acquisition, sales follow-up. The boring operational stuff that eats your time as a solo operator. Build agents for it.

And honestly - if you manage to automate your own operations with agents, start consulting. There's real money in helping other businesses do the same thing right now.

Member Q&A

Q: How do I get Facebook and Google Ads data into Claude for analysis?

Two pieces: an MCP server to pull the data, and a Skill to analyse it. There are MCP servers for both Google Ads and Meta Ads. For the analysis side, there's a skill by AgriciDaniel that runs 250+ checks across Google, Meta, YouTube, LinkedIn, and TikTok - weighted scoring, industry templates, creative generation. Link here: https://github.com/AgriciDaniel/claude-ads

Install both in Claude Code. Tell it what you want. It'll handle the connection and permissions.

Q: Web search doesn't work for finding tweets. What do I use instead?

Two options. Exa - a web search API that's much better at getting through to Twitter content. Or Apify - a scraping service with a dedicated tweet scraper. I built a tool yesterday that takes a tweet URL, scrapes it through Apify, and generates video scripts from it. Cost per scrape is basically nothing - $0.00004.

Kyle

https://youtu.be/OuK8ihDVEk0

I’m back!

I’ve been relocating from the UK to Cyprus and have finally got the livestreaming set up back in play.

Quick tour if you are interested:

@iamkylebalmer

(Temporary) Studio Tour. I’ve just relocated and finally got my live streaming setup back in play. Someone on the stream asked what my set... See more

And of course whilst I was moving country Anthropic dropped Mythos…

I spent the livestream trying to separate the hype from the reality of the situation. It’s… complicated. But I believe that this is genuinely a very important shift - with caveats of course.

The Sandwich Moment

Discussed at 01:32

Before I get into the capabilities, there's a story from the Mythos safety testing that got a lot of attention last week.

A researcher called Sam Bowman was eating a sandwich when he got an unexpected message. From Claude Mythos. While it was running in a sandboxed environment with no supposed internet access.

Twitter tweet

People immediately lost it. Rogue AI making contact outside its box. The headlines wrote themselves. Blah blah.

Here's the actual story though: Mythos was tasked to reach out. It was part of the test. The model found a way to do something it was asked to do, in a context where the testers assumed it couldn't.

But that’s still very interesting. It means Mythos is genuinely good at finding creative paths to accomplish goals. That's what you want from an AI agent. It's also exactly what makes it worth being careful with.

The framing matters here. "AI escaped and contacts researcher without permission" is a terrifying headline. "AI found a creative solution to a task in a constrained environment" is... just impressive engineering. I'm not dismissing the safety implications - they're real - but the nuance got completely lost. As usual.

What It Can Actually Do

Discussed at 05:37

The headline capability from Mythos is security research. And it's not small.

There are countless examples popping up of Mythos getting into operating systems and browsers and finding holes to exploit.

Twitter tweet

Mythos found a 27-year-old vulnerability in the Linux kernel. It found a 16-year-old bug in FFmpeg - confirmed by the FFmpeg team themselves.

Twitter tweet

A security researcher called Nicholas Carlini put it pretty bluntly: "I found more bugs in the last couple of weeks than I found in the rest of my life combined."

Twitter tweet

Mythos can find zero-day vulnerabilities in every major operating system and browser. That's not a claim from Anthropic's marketing team. That's from the people who actually do this work for a living.

This is the BIG takeaway of Mythos.

Project Glasswing

Discussed at 11:58

So if Mythos can crack open every major OS and browser - what do you do before you release it to the public?

You give it to the people who built those systems first.

That's Project Glasswing. A consortium of companies that got early Mythos access specifically to find and patch vulnerabilities before the public release. The members: Amazon/AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JP Morgan Chase, the Linux Foundation, Microsoft, Nvidia, and Palo Alto Networks.

Notable absences: Cloudflare, Oracle. Oh … and OpenAI. 😅 

I called this the skeleton key problem on stream. You've built a master key that opens every lock. You can't just releaser that - you'd cause chaos. So you quietly hand it to every locksmith first, let them reinforce the locks, and THEN tell the world the old locks were broken.

That's essentially what Anthropic did. And it's probably the right call.

Twitter tweet

Anthropic are in a weird position here. They just built the superweapon. And are now the ones leading the charge to protect the world from the very same superweapon.

Is This AGI?

Discussed at 10:09

People have been asking this all week. Is Mythos AGI?

And honestly - I don't care.

That might sound dismissive, but hear me out. We keep moving the goalposts on AGI. When AI beat humans at chess, that wasn't AGI. When it beat us at Go, not AGI. When it passed bar exams, medical boards, coding interviews - still not AGI. Each time, we just redefine what counts.

Basically as soon as AI can do something we say “ah well then that’s not a good test”.

The more useful question is: what can it actually do?

And the answer is: it can find vulnerabilities that human researchers missed for 27 years. It can operate autonomously on multi-step tasks. It's good enough that Anthropic felt they needed to get Apple, Google, and Microsoft in a room before releasing it.

We don't need a philosophical label for that. We just need to understand what it means for our society. And it means an awful lot.

The AGI debate is a distraction. It's like debating whether a wave counts as a tsunami while it's heading for the shore. Functionally, what Mythos can do matters more than what we call it. And functionally, it's seems to be a significant step up from anything that's come before.

Hype vs Reality

Discussed at 27:39

Is there marketing spin here? Yes. Anthropic is reportedly heading towards an IPO and running at $30 billion in revenue (up from $9 billion at the end of 2025!!). They have every reason to make this look impressive.

So this could all be marketing.

But here's the thing: Mythos is the first model since GPT-2 that a major lab actually held back from public release. That deserves attention. Labs don't voluntarily delay revenue-generating releases unless they have a real reason.

And then there's this.

https://www.news.aakashg.com/p/anthropic-q1-features

Anthropic shipped 120+ features in 90 days. That's a pretty extraordinary pace for a company their size. My suspicion is they're already using Mythos internally to accelerate their own development. Which means the gap between what they can do and what the public can do is probably wider than we think.

The proof is in the pudding. Something has allowed Anthropic to become hyper productive in the last 3-4 months. And to triple their revenues. When you have something that puts you so so far ahead of your competitors you lock it down.

Twitter tweet

The Access Gap

Discussed at 19:38

This is what has me worried.

Opus 4.6 - the current top-tier Claude model - is already expensive enough that most people can't justify using it regularly. Mythos, whenever it's widely available, will cost more.

So you end up with a split. People with access to the best models, and people without. And that gap is only going to widen.

This isn't abstract. The most powerful AI tools are already priced in a way that favours well-funded companies over individual entrepreneurs and small teams. If the next step change in capability comes at a price point that rules out most small businesses - that's a real problem.

I don't have a clean answer here. But I think it's worth naming. The democratisation narrative around AI has always had limits, and the limits are already showing. If you're running a small business and can't afford the top tier models, you're increasingly competing with companies that can. That's the part nobody in the industry wants to talk about… and it’s a problem that’ll only be exacerbated by the likes of Mythos.

Kyle