THE FUTURE OF WORK WITH AI

You can start before you know how

I wanted to see how far I could get making something for 15 Days of AI using basically zero motion-design knowledge. So I gave Claude Code our brand, some illustrations, examples of motion I liked, the actual product context and a fairly loose brief. This is where I ended up.

I still would not describe myself as someone who “knows motion design” lol but I know considerably more about it than I did the day before. And I didn’t learn any of it before making something. I learned it because I was making something.

We used to learn before we could make. Now we can make in order to learn.

Before generative AI, I built a whole community around a very basic loop:

show up → make something → share what worked → come back tomorrow

Back then, the annoying bit was often that you had to learn enough of the tool before you could get anywhere near the thing you wanted to make.

Want to animate something? learn After Effects.
Want to build an app? learn to code.
Want to do proper analysis? time to develop a concerning relationship with Excel.

There was usually this big pile of learning the tool sitting between curiosity and doing anything useful with it. AI is starting to mess with that order.

You can start slightly above your current ability.

Now you can decide “I want to animate this” and start. And the project itself starts exposing what you need to learn next.

• You don't know why the animation feels wrong until you see it moving.

• You don't care what easing is until the transition annoys you.

• You don't know what information Claude is missing until it makes the wrong thing.

The questions arrive after you begin. Not because chapter four of a course says it’s time. Because the thing sitting in front of you sucks and you want to fix it.

That is a VERY different learning loop.

And motion design is just a particularly obvious example. YouTube announced conversational editing this week too. Describe the edit you want, see what comes back, then jump into the timeline and keep working from there.

You don’t always have to learn your way toward the project anymore. Sometimes the project teaches you what you need to learn.

You do not need to learn all of AI. You need to get comfortable starting things before you feel ready.

There’s a weird pressure around AI right now to know everything. It makes learning feel like a giant backlog you’re permanently failing to clear.

Every week there’s another model, tool, workflow, agent, benchmark and person claiming they’ve already replaced half their company with it.

You will lose that game. I will lose that game. Everyone loses that game because somebody ships something new while you're asleep.

But curiosity is much easier to maintain than completeness.

When something crosses your feed and makes you think wait…could I use that for my work?

1. Spend 15 mins on it.

2. Get annoyed.

3. Ask AI why it looks wrong.

4. Share it with someone.

5. Try again tomorrow.

That’s learning too.

And the best learning cultures I've seen aren't full of people who know everything. They're full of people constantly bringing little discoveries back.

That loop is also a big reason we built 15 Days of AI around short daily practice and people sharing what they're making with each other.

Not because everybody needs more AI content. God knows there’s enough of that. But because the habit of trying things compounds much better than the habit of keeping up.

How to AI

This week: Opus 5.5 edition 🎬

• How to create studio-quality motion from one prompt

• How to prompt Opus 5.5 (Anthropic’s Guide)

• How NOT to prompt Opus 5.5

+ a few things people are already making 👇

Twitter tweet

BEFORE YOU GO

See you next Sunday! ✌️

THE FUTURE OF WORK WITH AI

I think this is the bit that gets lost when every week brings another model, agent, or terrifying headline. The future of work doesn’t arrive all at once. Mostly, it looks like normal people noticing one annoying bit of their job, trying something different and getting a little better at it. Here are three examples I liked:

1. Use the 3× rule to find the right task

Alice Browne from Brainlabs is a pretty good example of what an AI-active People leader looks like right now. She’s been publicly sharing how she uses AI in her own work while Brainlabs rolls Claude across its ~1,000-person agency.

And the company has a brilliantly simple rule for figuring out what’s worth using AI for:

In four weeks, Brainlabs employees created around 400 reusable AI skills using that mindset.

You don’t need 400 skills but you can pick one thing you’ve already done three times, then paste:

2. Show AI what “good” looks like for better output

Douglas Brundage at Every has a simple fix for one of the most annoying AI problems: you know what you want. AI just keeps missing it.

When he uses AI to build slide decks, he doesn’t start with a giant prompt. He makes a few slides himself first, gives those to AI as the reference, then lets it build the rest. He edits what it gets wrong and asks AI to turn those corrections into rules for next time.

So instead of describing your taste over and over, you’re giving AI evidence of it. You can steal that for basically any kind of work:

That’s a lot more useful than typing “less corporate” for the seventh time.

3. Run the Vacation Test for reusable workflows

Emily Kramer’s three-person team at MKT1 had built a bunch of useful AI workflows. The problem: too much of it lived inside one person’s laptop, account or chat history. So they started using a very simple Vacation Test:

If the answer was no, the workflow wasn’t really reusable yet. That’s a useful test even if you’re working solo. Next time you finally get AI doing something well, paste:

Then save that somewhere you’ll actually find it again. A good chat is nice. A workflow you can reuse is much better.

The models are going to keep getting stupidly good, stupidly fast whether we keep up with every release or not. I’m much more interested in a future where non-technical professionals get better at experimenting with their own work without constantly feeling like they’re behind.

That’s the future of work with AI I want to be part of.

BEFORE YOU GO

Steal this if you feel personally attacked 👇

I’ll share where everyone lands next time.

See you next Sunday! ✌️

So today:

🔥 why software is starting to adapt to us, not the other way around
👀 what non-technical people are making at work
🧪 3 new ways to test AI on your work this week

THE FUTURE OF WORK WITH AI

This might be the AI story from this week that matters more to your actual job.

A developer spent a week vibe-coding an app then OpenAI shipped an update that basically made it obsolete.

At first this sounds like a pretty depressing advert for building anything with AI lol. But there’s another way to read it.

Software just became cheap enough to build in a week and disposable enough to replace the next one.

And I think that changes something much bigger than vibe coding. For most of our working lives, software has dictated how the work happens.

AI is starting to reverse that relationship. We’re seeing small versions of this inside 15 Days of AI all the time.

One person working in legal was frustrated by the world clocks available to her, so she made an editable one around the time zones she actually needed. Someone working in Workplace built a specialised Health & Safety assistant around information they were already responsible for. Others have been experimenting with contract review, spreadsheets, regulatory documents and stakeholder comms not because they woke up wanting to “build software”, but because they know where their own work is annoying better than anyone else does.

That’s why I think the whole everyone is a developer now framing is slightly wrong. You don’t need to become a software person. But knowing your own work deeply enough to say “this is stupid, it should work like this instead” is becoming a much more valuable skill.

HOW TO AI

Every week, this section is your shortcut. Here are a couple of ways you could use AI at work this week that are worth your time

📊 Ask your company data a normal human question

If your company uses ChatGPT Work, the new Data agent can connect to sources like Snowflake, BigQuery, Drive and SharePoint, investigate a question and build an interactive dashboard without you writing queries or learning another analytics tool.

✏️ Draw what you mean instead of writing long image prompts

ChatGPT Images 2.5 launched this week with Sketch, which lets you draw the composition you want and use that as the reference for the final image. To try it yourself⁠, just type “@Sketch” in ChatGPT

🫠 Stop jumping between five Google apps to finish one task

Google shipped new Gemini capabilities across Gmail, Drive, Docs, Slides and Chat this week. You can now ask from the app you're already in and have Gemini work across the others in the background including creating formatted docs, sheets and slides from the context it finds.

BEFORE YOU GO

Apparently this is how we’re coping with the AI discourse now:

See you next Sunday! ✌️

WINDOW INTO THE FUTURE OF WORK

Everyone is using AI. Sort of.

Everyone got the memo. Use AI more. Experiment. Find use cases. Become “AI native.” Whatever you do, don’t be the person who gets left behind. And apparently people are feeling that pressure.

Microsoft asked 20,000 people who already use AI at work about this earlier this year. 65% said they’re worried about falling behind if they don’t adapt quickly. Except 45% also said it feels safer to keep hitting their existing goals than to redesign how they work with AI. And only 13% said they’re rewarded for trying to reinvent the work when the experiment doesn’t immediately pay off.

This is a pretty cursed setup lol. You’re being told the way you work needs to change, while the safest way to succeed is to keep delivering exactly what you already deliver.

So of course people start optimising for the easiest thing they can prove. Prompts sent. Tokens burned. Copilot opened. Training completed. Agent deployed.

All very easy to count. Whether the work actually got better is where it gets annoying. So what does progress look like? 👇

When AI actually changes the work

Buffer basically stopped normal work for four days and told the whole company to build something useful with AI. By the end, 71 people had shipped 56 new internal tools.

And Hailley Griffis, their Head of Communications & Content, went into the week describing herself as non-technical and came out having shipped six internal tools and making small website fixes herself instead of immediately pinging an engineer.

This is obviously still a very AI-forward company. But Buffer isn’t the only team landing here. At LangChain, an AI drafts sales outreach, but it lands in the rep’s Slack for them to send, edit or cancel. Every edit teaches the system what to change next time. Mintlify literally has someone responsible for keeping the company knowledge their AI works from useful.

That is much closer to the AI future of work I’m seeing emerge. Specific bits of work now work differently and somebody is responsible for making sure it keeps working.

WHAT WE’RE TRYING TO MEASURE INSTEAD

I said on LinkedIn this week that I’m going to start sharing more of what we’re seeing inside companies trying to change how people work with AI. So here’s one that feels very relevant today.

At 100 School, AIQ is one of the ways we look at whether people are getting better at using AI in their work. It boils down to three pretty basic things:

The full thing gets more specific depending on the team. But those three questions are a lot closer to what we care about than how many times somebody opened ChatGPT this week.

HOW TO AI

The model war is back again.

Claude Fable 5.1 launched Tuesday. GPT-6 Astra followed Thursday. Try them but you probably don’t need to relearn how to use AI this week.

These are 4 guides that still matter whichever model you open:

• How to fix a bad AI answer.

• How to find a task worth using AI for.

• How to turn one good AI result into something repeatable.

• How to make AI check its own work against your quality bar.

BEFORE YOU GO

See you next Sunday! ✌️

WINDOW INTO THE FUTURE OF WORK

1. Use AI to practice the awkward stuff

47% of Gen Z sellers say they don’t get enough chances to roleplay before customer calls. 46% rarely get feedback on the conversations they do have.

That’s a problem AI is almost perfect for!

Harriet Hodgson-Grove, a partner at a UK accounting firm, uses AI voice mode to rehearse difficult conversations before having them for real. She’ll ask it to interrupt, challenge her or stay silent so she can test how the conversation might actually unfold.

And this goes way beyond sales. It works with a difficult feedback chat, defending an idea to your boss, a negotiation, or presenting to a sceptical client.

TRY THIS

Before your next presentation, pitch, interview or difficult conversation:

2. Don’t give AI work you can’t judge yet

From September, Deloitte’s new graduate auditors will learn some “old-school auditing” alongside AI-assisted work. One exercise has them do a cash reconciliation themselves (basically checking that the records and cash balances match) then review one produced by AI.

As AI takes on more of the manual work junior auditors used to learn on, Deloitte says the goal is to get new hires further up the learning curve sooner, doing more work that requires judgement, curiosity and scepticism.

There’s a useful rule hiding in that for the rest of us. You don’t need to manually do every spreadsheet, research pass or first draft forever. But if AI gave you a very convincing wrong answer tomorrow, would you catch it?

TRY THIS

For one task you use AI for regularly, build the checking habit before the shortcut:

3. Diagnose before you re-prompt

A bad AI answer turns a lot of us into slot-machine operators:

One thing we’ve learned from running 15 Days of AI is that structure beats endlessly prompting harder. In BCG’s People Team pilot, 122 participants produced 1,423 outputs using AI on actual work. Things like candidate-screening assistants, onboarding scripts, compensation analysis, HR dashboards and email-triage systems.

They didn’t just use AI more. They were given repeatable ways to think through the task first.

TRY THIS

When the next answer sucks, paste:

Then answer those questions before generating version two. A bad output is useful if it tells you what was missing. Otherwise you’re just pulling the lever again.

WANT TO START BUILDING WITH AI?

A lot of non-technical professionals can already see what in their work should be automated. Until recently, they just couldn’t build it themselves.

Harold Dijkstra and Kieran Ball are running a 3-week AI Builder Bootcamp starting Sept 7 to help you turn those ideas into working apps, automations and AI workflows.

ONE MORE THING WORTH READING 📬

If you’re thinking about your next career move, The Dose has a free 5-step planner for figuring out where you are, where you want to go, and what to do next.

You get it free when you subscribe to their weekly newsletter on careers, work and earning more.

_{not a paid ad btw}

BEFORE YOU GO

Anyway. Enjoy your Sunday. I’m off to let AI do the work while I provide the judgment. And judging by Thibault from OpenAI, apparently the humans are still doing quite a lot 😭

See you next Sunday! ✌️

WINDOW INTO THE FUTURE OF WORK

AI is extremely good at finishing the output, not the job.

Earlier this month, Epoch AI and Ipsos asked 1,106 professionals about how AI is showing up in 10 normal work tasks like reading documents, analysing data, maintaining records, designing systems, etc.

The good news is when AI does most or all of the task, people are much more likely to report saving time. 53% said those tasks now take less time, compared with 37% when AI only helps with part of the work.

However, roughly 1 in 6 AI-assisted tasks now takes longer than it did before.

Obviously this doesn’t mean AI is wasting everyone’s time. It means the extra work is going somewhere. So I went looking for it.

The clearest answer I found was annoyingly boring: bad project briefs 👇

A £6,000/month boring problem

The most useful example I found all week was not another autonomous agent demo. It was a marketing intake form.

E.ON Next’s go-to-market team had a very normal problem: people across the company would submit project requests without enough information, then producers would spend ages chasing context before anyone could start the work.

According to their case study, that manual back-and-forth was costing roughly £6,000 a month in producer time. So they built AI into the bit everyone usually skips over in the demo.

Someone submits a request. The AI checks it against an example of what a good request should contain. Missing context? It goes back and asks. Still missing something? It asks again. Only then does it write the brief.

I like this because there’s no “AI employee” doing everyone's jobs while the marketing team goes to the beach. They just found the part of the job that kept creating stupid back-and-forth and designed around it.

E.ON says the workflow now saves an estimated £5.5k/month. Across the wider setup, early-stage back-and-forth fell 25%. See exactly how E.ON built the workflow.

The AI output isn’t really the innovation here. The team first had to work out what information needs to exist before this job can move forward. That’s a very different question from what can we automate? And the actual process is stealable.

Steal this bit

AI SPEED VS. COMPANY SPEED

Last week I wrote about getting one good AI result to repeat. The annoying sequel is that you can get much faster while the system around you stays exactly the same. And this is probably why this post from Zara Zhang has been bouncing around X today:

I wouldn’t take the 10x vs 20% numbers literally, but the frustration feels very real.

You can 10x the person without 10xing the workflow.

Which is why I like the E.ON example. They didn’t just give the producer a faster way to write briefs. They went after the crappy handoff creating the work in the first place.

This gets even messier when the “workflow” involves 50 or 500 people rather than one very motivated ChatGPT user. That’s a lot of what we work on with teams at 100 School: moving beyond everyone individually getting faster and figuring out what changes in the work around them.

QUICK UPDATE 🎉

We shipped the first beta of our latest product this week!

It’s a simple experiment to help people spot the parts of their work that are taking too much time, effort or energy then figure out where AI might help.

If you want to try it, just hit reply.

BEFORE YOU GO ✌️

Anyway, one final diagram of the AI productivity revolution.

See you next Sunday!

WINDOW INTO THE FUTURE OF WORK

I heard Grace Clarke describe this really well on How I AI this week.

Grace teaches non-technical people how to use AI. She used to give students carefully engineered prompts so they could get a great result straight away. They told her it was actually making things harder.

What helped more was getting used to bringing AI into whatever they were already working on. Sometimes she’d literally send a Slack reminder saying: screenshot what you’re working on and drop it into Claude

Eventually they stopped needing the reminder. Grace calls it muscle memory.

You don’t need to become amazing at prompting before AI becomes useful. You need enough reps with the same kinds of work to start noticing what AI needs from you, what it keeps getting wrong and what’s worth keeping for next time.

And this is where I see people waste a surprising amount of time. So I’d do this instead.

How to make one good AI result reusable

1. Pick a task that gives you another rep

Don’t start with the task you hate most. Start with something that comes back and where you can tell fairly quickly if AI did a decent job. A good test is to ask whether you’d give this to a new teammate in their first week.

✅ A weekly project update that pulls from the same places and follows roughly the same format.
❌ A board report that changes every month and needs loads of judgement.

a16z found a similar pattern looking at AI workflows already running at scale: repeatable work with a clear path and an answer you can verify tends to work better. Good reps look more like:

• Teams interview → structured candidate notes you review

• 5 customer calls → repeated objections for the next campaign brief

• weekly project notes → drafts for social posts

2. Brief AI once instead of drip-feeding the job over six messages

I saved this prompt from Alex Prompter this week because it fixes a very normal problem: AI confidently starting before it has enough context.

His version makes AI ask questions one at a time instead of filling in the gaps itself. We use a similar exercise with non-technical professionals. For a task you want to repeat, try this:

💡 Voice works really well for this. You’ll usually mention the weird little details you’d never bother typing, and those are often the things AI needed in the first place. Once the summary looks right, keep it. That becomes the starting brief for the next rep.

3. Save the bits that changed the result

Don’t save every prompt. Save the things you had to teach AI before it became useful. For example:

Those are the things that make attempt #2 less random. A prompt can change completely while those four things stay useful.

4. Run the second rep

When the work comes back, don’t judge the setup by whether ChatGPT remembers your favorite adjective. Ask “What do you already know about how I work on this task, and what do you still need from me?” then run the real task.

DID I…

□ explain less?
□ repeat fewer old corrections?
□ get to something useful with less back-and-forth?

If you’re still explaining the same things, update the setup. If something new helped, keep it for rep #3.

WANT TO BUILD MORE OF THESE REPS?

This kind of repetition is a big part of our free 15 Days of AI challenge.

15 minutes each weekday practicing with AI on your real work. Over the 15 days, those little reps start turning into habits and reusable ways of working.

TRYING TO BUILD THE SAME HABITS ACROSS YOUR TEAM?

It gets harder when you’re trying to move hundreds of people beyond the occasional good ChatGPT conversation.

That’s what we work on with 15 Days of AI for Teams: repeated practice on real work until useful AI behavior and habits start sticking.

BEFORE YOU GO ✌️

See you next Sunday!

WINDOW INTO THE FUTURE OF WORK

Last week was about finding the right task for AI. This week is about turning that idea into something that works. We’ve started calling that the execution gap.

And in real work, that shows up as a bunch of very normal frustrations. So I pulled out five that felt especially familiar, plus the workflows and fixes people are using.

1. I keep having useful conversations that go nowhere.

Ryan Carr shared a tiny workflow this week that I loved because the frustration behind it is something we all experience: useful discussions happen in Slack but we keep losing them. So they created a system where whenever someone reacts with this ✏️ emoji in Slack, it triggers a Zapier workflow where Slack grabs the entire thread → Claude turns it into a ~300-word draft → the draft gets sent back to Slack, Google Docs or Gmail

Ryan says the draft usually gets them about 70% there, with roughly 15 mins of editing left. That’s the bit I like. Nobody on the team has to remember a new workflow. They keep talking in Slack like they already do and AI comes to the work instead of making everyone move their work to the AI.

They already knew the problem: we have useful thinking but we keep losing it. AI just gets one very specific job in the middle.

2. I have a lot of information. Turning it into something useful takes forever.

Nick Baumann from OpenAI shared a workflow recently that made me jealous. He records 50–60 messy video takes and dumps the whole batch in ChatGPT Work. Inside it, he adds his own “UGC Video” plugin he’s built from the instructions and corrections he’d accumulated while making previous videos.

From there, ChatGPT Work transcribes every clip, looks through the footage, uses cues like “bad take”/ “good take,” finds the strongest takes and pieces the story together. In the example below, 27 takes became a 43-second finished vertical video. It even checked for things like sensitive information and blurred them where needed.

I think this is where AI feels magical. Not when it invents more stuff but when it takes the messy material you already have and turns it into something you can actually use.

3. I get plenty of good answers from AI. I still have to decide what’s good.

I loved this interview with Hayashi Aki this week. She’s an ex-journalist who now works with around 15 companies. Her workflow is basically interview → transcript → give AI direction → decent draft → she edits the life back into it.

She says that AI workflow has taken her from drafting roughly one article a night to as many as ten. But the more interesting bit is what happened to her job after letting AI create her first drafts. Clients started asking for more strategy. She spends more time asking uncomfortable questions, drawing the actual story out of CEOs, spotting what they didn’t say, editing, managing risk, making the final call.

Her idea is AI can give you infinite plausible options but a human still has to choose. This reminded me of something I changed my mind about earlier this year. I used to like the rule: AI does the boring work. Humans do the thinking. Now I think it’s better to ask: which thinking should I keep?

I’m increasingly excited about that question and also slightly terrified by it.

4. I made the process faster but it’s still a terrible process

I watched Tomás Dostal Freire describe what happens when companies integrate AI into a broken workflow. He said it’s like putting a Formula 1 engine into a family car.

He calls it individual productivity vs company velocity: making one person faster vs actually removing steps, handoffs and bottlenecks from the work. We saw the opposite in one enterprise team we spoke to. A recruiting workflow with roughly 20 steps across two systems was redesigned down to about 8 using AI, without replacing the underlying software.

That’s what I meant when I wrote about “swapping the motor without redesigning the factory” back in March. I think I understand that line better now. Saving 20 minutes on a task is useful. Removing 12 unnecessary steps is something else.

5. I have the tools and training but my work still hasn’t really changed.

This one makes me a bit mad. One team recently told us they were struggling to get traction with their AI Academy despite a massive comms effort.

They’re not the only ones. Gallup’s workplace data shows how much the boring stuff around the AI matters. When people strongly agree their manager actively supports AI use, 79% are frequent users vs 46% when they don’t. And when AI actually integrates with the systems people work in: 86% vs 52%.

Which makes me increasingly suspicious of “we need more AI training” being the go-to solution for a lot of teams. Sometimes, yes. Sometimes people already know enough. They just don’t have the time, access, permission or manager support to actually change the work.

That’s not a skills problem. That’s the environment around the skill.

So what does good execution actually look like?

Nobody starts with “I would like an agent please.” They start with the thing that sucks, then get specific about what needs to change. That’s probably the execution gap: getting from “this sucks” to “this now works better.”

Still figuring it out, but I’m pretty convinced that’s where the interesting work is moving.

ONE MORE THING WORTH READING 📬

I read a lot of newsletters. Occasionally, one is too good to keep to myself.

TheFutureParty is a daily newsletter that makes it easier to keep up with business and tech news without making you feel like you missed a meeting.

They cover the latest stories and insights to help you understand the future of tech, business, and culture in one quick witty package. If you like the slightly weirder bits of this newsletter, you’ll probably like them too.

🤝 _{this is NOT a paid ad}

BEFORE YOU GO ✌️

Which AI frustration are you dealing with right now? Hit reply. I’m curious about the ones I missed.

See you next Sunday!

WINDOW INTO THE FUTURE OF WORK

Weirdly, OpenAI seems to have run into the same problem.

They spent years making the blank box smarter but still have a blank-box problem. Akshay Nathan, who leads core product engineering at OpenAI, was recently asked which automation he wishes worked better.

His answer was basically: bring me new ideas.

The same is true of finding a useful AI task. Your calendar, emails, Slack messages, half-finished documents and that spreadsheet called FINAL_ACTUAL_FINAL_v7 all know what happened last week.

ChatGPT doesn’t. At least not until you show it. So here’s a 10-minute exercise that turns one normal day into three clearly defined tasks and shows you what’s still missing before you test anything 👇

1. Reconstruct one normal workday

Pick the most recent day that felt fairly normal, included at least three different tasks and left enough receipts to reconstruct.

2. Pull out 3 pieces of work

• Open your: Calendar, sent emails, recent files…

• Write down one thing you finished (eg: the weekly report), one thing you chased (eg: chased interview feedback from hiring managers) and one thing you left unfinished (eg: left the campaign research half-done because the information was scattered).

• Don’t judge whether they are good AI tasks yet. Write them down roughly.

💡 Keep the source material open: the spreadsheet, report, PDF or email thread will tell you more than memory will. OpenAI’s own examples start with the actual file, then ask AI to extract changes, risks, dates or owners.

And if the evidence is scattered across files and apps, that’s the kind of longer, multi-step work ChatGPT Work is designed for. Eventually but first, name the task.

3. Turn broad areas into real tasks

This is the difference between naming an area of your job and naming an actual task. It is also one of the first habits we teach inside our free 15 Days of AI challenge for professionals (the next cohort starts August 10th).

💡 And a useful clue: if the same input → work → output repeats every week, it may eventually deserve a reusable Claude Skill but only once the process is clear enough to write down.

4. Find what’s still missing

Now you can open ChatGPT or Claude and paste this:

You should now have three tasks with the work, source material and gaps visible side by side. And that’s not just prompt theatre. MIT Sloan recommends breaking jobs and workflows into individual tasks before deciding which generative AI use cases are worth testing.

And that’s where this little rewind stops. Because a clear task is not automatically a good AI task. It may require judgment you should keep. It may be easy to automate and make almost no meaningful difference. Or it may be a great fit, but only after the missing information is fixed.

That’s also what the Institute for the Future of Work found in recent case studies: outcomes depend on which tasks change, which skills develop and how work is designed, not simply whether AI is introduced. Their research is worth reading.

Still not sure which task to choose?

That exercise makes the work visible but we are also testing a more detailed version that looks across your actual role, compares the opportunities and explains why something may be a strong fit, a partial fit or not an AI fix at all.

We’re still reviewing a mix of roles, so there’s time to put your name down before testing begins during the week of August 10.

Before you go ✌️

Window into the Future

The uncomfortable truth about AI at work right now is it doesn't always lift an entire team or business as much as it creates outliers. Boris Cherny, who built Claude Code, shared this week that he keeps hearing the same line from teams everywhere: one person 10x's their output with AI while the rest of the org never catches up.

Boris is talking about engineers. Swap engineer for a marketer, a finance analyst, the person who owns hiring, and the shape doesn't change. And notice that clearing one blocker just reveals the next, and it’s almost never the model or the tool but whether people can manage the thing in front of them.

Everyone's fighting over who owns the frontier, open weights, closed weights, Kimi, Fable, etc... but you can download the best open-weights model on earth tomorrow and your company still won't have changed how it works.

Say you run people at a company of a few thousand. Five of your staff are basically AI-native now who changed their work around AI without being asked to. Everyone else is still waiting to be told what to do with it. We watch this split all the time at 100 School. Giving everyone a licence or more AI training on its own doesn't fix that. The issue is not access or enthusiasm but that no one owns the next step:

• who redesigns the workflow?

• who checks the output?

• what does the manager reinforce?

• how does anyone know it worked? 

That last one is where a lot of the market cheats. Someone pulls up a usage chart: logins up, tokens up, and calls it ROI. I wrote a while back that usage tells you the tool got opened, not that anyone got better at their job. Boris, who has every reason to love a usage number also says usage "measures activity, not return”. Would you have paid a person to do this work? how many hours? that's the only number worth putting in front of your boss. Everything else is just people opening a tool.

The companies that get through this do it slowly, and someone owns it. We've run AI training inside HiBob for over 18 months: 9 departments, 1 function at a time, because what works for sales doesn't work for product. Nowhere near as fun to post about as a 2.8-trillion-parameter model. Also the only version I've seen that sticks.

Two weeks ago, I asked how you were feeling about and one of you said: “there's a new tool every day, and I'll always be behind”. You will. We all will. That race has no finish line so stop using “caught up” as the goal. Pick one workflow, decide who owns improving it, give them the time to try, check whether the result is actually better. The next model will arrive before you’ve finished. That’s fine. The advantage isn’t knowing every tool. It’s getting good at turning useful experiments into normal work.

How to AI 🤖

Every week, this section is your shortcut. Here are a couple of ways you could use AI at work this week that are worth your time:

• How to prepare your team for Agentic AI

• How to prepare HR for workplace AI rules

• How to redesign work before adding more AI

• How to start managing AI instead of just chatting with it

Before you go ✌️

I'll break down the results next Sunday. See you next week!

Window into the Future

ChatGPT Work is where this gets much bigger than the new voice model.

You can hand it a project, give it access to the files and apps you choose, and leave it working across the messy middle: research, analysis, documents, spreadsheets, presentations, websites. OpenAI says it can stay with complicated jobs for hours, splitting them into smaller steps and completing them without waiting for you after every click.

It’s basically the useful parts of Codex brought into the ChatGPT product almost everyone already understands well enough to open.

For non-technical people, that matters more than a benchmark score. You dont need to know how the agent is assembled. You can hand it the folder.

There are loads of examples in OpenAI’s launch post: complete slide decks, financial models, documents that follow an existing company format, interfaces it can inspect and refine instead of generating once and abandoning.

That part looks excellent.

But the most useful criticism I saw came from Ethan Mollick a few weeks ago, then again after testing ChatGPT Work this week.

Twitter tweet

His point is that these products still think like software tools.

In software, the codebase can serve as a source of truth. You can test whether something works. Changes are tracked. Failed branches, old versions and decisions are usually recoverable if someone cares enough to look.

Knowledge work is less cooperative. A polished report might contain 20 sources. It probably won’t tell you which two sources caused the author to change their mind, or that one important number came from a spreadsheet somebody described as “rough but basically right” in Slack.

A lot of the work is hiding around the artifact. Which means an agent can produce something that looks finished while missing the part that made the work make sense.

I keep thinking about my own newsletter archive.

There are dozens of published issues in there, plus screenshots, links, messy drafts, subject lines I dropped because they were repetitive or just a bit boring. If I ask ChatGPT Work to analyse it, I dont need a summary telling me we write about AI and the future of work.

I’d want it to tell me where an argument first appeared, how it changed across several issues, which sources kept coming back, whether I started repeating a sentence shape too often, and where it cannot know the reason because that reason only exists in somebody’s head or in an old Slack thread.

That’s closer to actual knowledge work.

The models can now work for hours. They can inspect more files than I could read in a week, keep several routes open at once, build the deck, fix the deck, then build a tiny website for the deck because apparently we’re doing that too.

I still have the same brain I had last Sunday.

Twitter tweet

The launch videos tend to end when the work is done but normal work doesn’t. And that’s a very different problem from “can the model do the task?” It can.

Can the rest of us keep up with what it did, understand enough of the route, and spend our attention in the places where a quick approval could turn into a very expensive afternoon? That’s the part I think teams will run into next.

There will simply be far more finished-looking work arriving than anyone is used to reviewing. We already have this with AI-written documents. Agents can multiply that asymmetry.

There’s a version of this that works brilliantly, obviously. The agent does the slow collection and construction, then returns the small number of decisions that actually need you. That’s what the good examples are moving toward.

But I want the receipt attached. So the first thing I’m trying with Work is a very unglamorous instruction at the end of the task:

Alongside the final output, give me a one-page work log showing:

- the sources you relied on most
- anything you ignored or could not access
- the assumptions you made
- decisions you made without asking me
- anything that still needs a human to check

I just want to know where to look before I approve something built from 70 files and three years of context.

This is also a lot of what we’re working through with teams at 100 School. Once the tools can do more, somebody still has to agree what needs checking, what needs recording and what “finished” means. Because once the job gets more complicated than counting the e’s in seventeen, “looks finished” isn’t enough.

How to AI: GPT-5.6 + Work Edition 🤖

Every week, this section is your shortcut. Here are a couple of ways you could use GPT-5.6 at work this week that are worth your time:

• How to choose between Chat, Work and Codex.

• How to start your first ChatGPT Work project.

• How to turn a repeated part of your job into a Work loop.

• How to choose between Sol, Terra and Luna.

• OpenAI Build Week starts tomorrow.

Before you go ✌️

See you next week!

Window into the Future

First, the boring-but-important bit: Fable is included on most paid Claude plans through Tuesday, July 7. If you want the full story of what changed, The Neuron wrote a good plain-english version, but the short version is all you need: you can use the best model there is, on your normal plan, until Tuesday.

And the smart way to spend the next two days is setting up a few things that keep paying you back long after Tuesday. Five worth stealing:

1. Stop over-explaining it. 

Your instinct is to write a giant prompt with fifteen rules. Anthropic's own guide says that makes Fable worse. The elaborate prompt setups you built for older AI are the thing dragging it down. Tell it what you’re actually doing and who its for, hand over the mess, and get out of the way. Their exact template, which you can paste right now:

im working on [the bigger thing] for [who its for]. they need [what the output lets them do]. with that in mind: [your actual ask].

so instead of a numbered checklist, its: "im writing the Q3 board update for our exec team, they've got five minutes and care about pipeline and churn, here’s everything. Write it."

2. Give it the whole pile at once. 

The one thing Fable does that a normal model can’t is hold a huge amount at once without losing the thread. So stop feeding it one doc at a time. Drop in the whole folder, the call notes, the horrible spreadsheet, and ask for the finished piece, not a summary. This is the sit-down job you’ve been putting off for a month.

3. Give it a memory. 

Keep one plain running doc where it notes how you like things, what it got wrong last time, the calls youve already made, and have it read that every session. Anthropic recommends this outright, and you can set it up with no code in about ten minutes. It means you stop re-explaining yourself in every new chat, which, if you’re anything like me, is half the reason AI starts to feel like a chore.

4. Brainstorm with the cheap model, then have Fable build it.

This is the one people are calling the smartest move of the week. You’ve only got Fable at half your usual limit this week, so don’t burn it on the thinking. Open a chat with Opus 4.8 and have it interview you about your day (where you lose time, what you do over and over, what drives you mad) until you land on one small tool that'd actually help. Then, while Fable's still free, have Fable build the thing. Cheap model to figure out what you want, Fable for the part that lasts.

In Claude Opus 4.8: 

"Interview me about my work, one question at a time, about where I lose time, what I repeat every week, and what annoys me most, until we land on one small tool or system that would genuinely help. Then write a clear, detailed brief for building it." 

Then paste that brief into Fable 5: "Build this for me, end to end. Here's the brief: [paste]."

5. Build yourself a "does this pass?" check. 

My favorite, and the one I’m stealing. A creator fed Fable all their past posts, had it work out what made the best ones land, and turned that into a scorecard. You can copy his prompt to build you own:

Here are my three best [posts]: [PASTE THEM]

1. Study them and write a scoring rubric out of 100. Base the categories on what makes these good: [hook, structure, voice, clarity]
2. Draft a new one on [TOPIC]
3. Score the draft against the rubric as a strict reviewer
4. Under 95: rewrite and score again. Repeat until it passes
5. Show me only the final version and its score

swap "posts" for emails, designs, proposals, reports. I immediately wanted one for this newsletter.

Notice what all five have in common? none of them are really about Fable. They’re about building yourself a little system, a way of describing things, a memory, a reusable playbook, a quality bar, that keeps working after the model changes. And it will change. Fable might be behind a paywall by Wednesday. Thats the whole thing I keep banging on about with AI-proofing your work: the tool is temporary, the way you work is what compounds. (It’s most of why 100 School exists.)

(and yeah, there’s a loud fight this week about whether Fable came back weaker, its real, but its a coding argument, and for everything mentioned above, that shouldn’t be a concern)

So don’t rush, don’t feed it junk. Set up one of these five properly and let it earn its keep long after Tuesday. And build yourself something fun like the koi pond too, for morale.

How to AI: Fable 5 Edition 🤖

Every week, this section is your shortcut. Here are a couple of tips/ ways you could try Fable 5 this week that are worth your time:

• How to use Claude Fable 5 without going bankrupt

• How to prompt Claude Fable 5 (Anthropic’s full guide)

• How to use Claude Fable 5 (unlearn your ChatGPT habits)

• How to create cinematic websites with Claude Fable 5

Before you go ✌️

Whatever gets launched or pulled back next week, the habit matters more than the tool, so I’m curious where you actually are right now?

none of those? hit reply and tell me where you really stand. I read every one.

See you next week!

Window into the Future

On Tuesday, Anthropic launched Claude Tag: a version of Claude that lives inside a Slack channel as a shared teammate. You tag it the way you'd tag a colleague, hand it a task, and it works away while you get on with something else. There's even an "ambient" mode where it doesn't wait to be tagged. It watches the channel and jumps in when it reckons it should.

For over a year the question was "are you using AI?" but now it’s "how does your team work when there's an agent in the room?"

The people who build this stuff are excited, and their reason is more interesting than "cool robot." The argument, made by Andrej Karpathy and the Latent Space crowd, is that an agent was never really held back by how clever the model is. It was held back by where it lived.

A chatbot in a separate tab only helps if you remember to open it, paste in the context, and copy the answer back out. An agent that already sits where the work happens, with the memory and the access, skips all of that. Karpathy called it the third big redesign of how we work with AI. He also said that the payoff only shows up once the dull, hard plumbing is done aka the security, the memory, the permissions.

Not everyone's sold though. The worry, from people like Ashwin Gopinath, is about what your Slack becomes over time, the place that holds every decision, every "we tried that in Q2 and it died," every half-finished thread. Let one vendor's agent be the thing that remembers all of it and you stop renting a clever model and start renting your own company's memory back from them. The open-source camp, Santiago and others, say own the layer yourself and never get married to one provider. Reasonable. Though it's worth noticing that most of the loudest critics happen to sell the alternative.

Ethan Mollick made a good point. While leadership debates whether to build an in-house AI stack, the employees have already decided. They want the same tool they use at home, and they'll keep asking IT until they get it. Nobody's waiting for permission. People are already using AI at work on their own. Claude Tag just makes it official.

The change that matters more and that’s easy to miss in my opinion is the second you tag an agent and hand it a job, you've stopped using a tool and started managing one. You start feeding it context, setting limits, checking its work, figuring out when to trust it and when it's wrong. That's management. And most people on your team have never managed anyone. Claude Tag doesn't just drop a new worker into the channel. It makes everyone in there a first-time manager, overnight, with nobody to show them how.

You can't buy a team that manages this well, and you can't get there overnight. You build it on purpose. So the ones who come out fine will be the ones who already manage well, or who decide to get good at it. Anthropic's own blog basically says this: what makes a human-and-agent team work are clear roles, writing things down, building trust over time, keeping a person on the calls that matter which is just good management. None of it is new. The agent only makes skipping it more expensive. (Which is most of why 100 School exists. The behaviour change is the part that sticks when the tool gets swapped out from under you.)

How to AI 🤖 

Every week, this section is your shortcut. Here are a couple of ways you could try AI this week that are worth your time:

• How to build a human-and-agent team that actually holds together.

• How to set up Claude Tag in your team's Slack.

• How Claude Tag actually works from the people who built it.

• How to use ambient mode in Claude Tag.

Before you go ✌️

If an AI showed up in your team's channel tomorrow, would your people know how to work with it? Hit reply and tell me, I read every one.

See you next week!

Window into the Future

Everyone keeps asking what AI's going to do to their job. I get the worry but a more useful question is who on your team is going to be the person that gets everyone else to use this stuff?

because that person is going to matter way more than people realise right now and at the moment, almost no one is doing the job.

Think about where last week's money actually went. $60 billion on Cursor, 300,000 consultants lined up to go and set the software up. And getting normal people to change how they work with this day to day gets almost nothing. I’m a bit annoyed at myself for how long it took me to properly see that.

BCG actually put a number on this that stuck with me. They reckon about 70% of the value companies get from AI comes from people changing how they work, not from the tech itself and yet the spending is almost all going to the tech.

but how do you become that person? A few things, and none of them are technical, I promise.

1. First, stop talking about AI and just show it. Pick the one task your team openly hates and do it with AI, right in front of them. People copy people. A colleague sharing in Slack how they use AI to get something done does more than any all-hands slide on AI strategy ever will.
   
   …but research out of Atlassian found that if you admit you used AI, your colleagues rate you as 10 times lazier and are far less likely to put you forward for the good, visible work. However, in teams where people use it out in the open and the wins get talked about, that flips completely, and the AI users end up looking like the sharp ones. So dont hide it. the more openly you use it, the easier you make it for everyone around you to do the same.

   
   

   

   
   

2. Second, sit with the AI sceptic. the person you want is the one working around the tool the old way, nodding along in the meeting and ignoring the whole thing the moment it ends. Winning that one person over is most of the job. Think back to Morgan Stanley: the four months wasnt a tech problem, it was a trust problem.
   

3. Third, and this is the one that wears me out: please dont run a training. Im so tired of the one-hour AI workshop everyone forgets. Just sit next to one person and do one real task on their actual work, together. What you're after is a habit, and habits dont come from watching a video.
   

4. The fourth one is the one everybody skips. Stick around long enough that the new way just becomes the normal way. One of our clients, HiBob, has been doing this with us for 18+ months. 15 mins of AI a day is the habit they kept feeding, which is exactly why it worked.

   
   And the role itself is starting to exist now: Stripe is hiring someone to embed with a team of marketers and coach them, one workflow at a time, until working with AI is simply how that team operates. This is a job that didnt have a name a year ago.
   
   

   

   
   

5. and the last one might be my favorite: when one person finally cracks it, don’t keep it to yourself. Show the next person, or better, get the one who cracked it to show them. We ran our 15 day of AI challenge with one content team at Headway, and two other teams signed themselves up based on the content team’s results. No one told them to, they just watched it working for their colleagues and wanted in.

which brings me back to those Morgan Stanley numbers. The money this year is still pouring almost entirely into the tech. Soon enough everyone's working off the same models and the same apps, and having the best one wont count for much. What sets one team apart from the next will be something far less shiny. Whether there’s someone there who can get a room full of people to work differently. You dont need a budget, or a title, or anyone's permission to be that person. You can start this Monday.

Before you go ✌️

When's the last time you sat down with a colleague and went over one task with them? what happened?

Our team at 100 School meets once a month for what we call “AI Show & Tell” where each one of us shares how they’ve been using AI in their work. Highly recommend!

See you next week!

Window into the Future

The best thing I read all weekend wasn't really about the ban. It was Daria Cupareanu on what she built in the 72 hours before it.

She's not a developer. For months she'd been trying to get her own numbers out of Substack, which doesn't make that easy, and had given up. This time she just described what she wanted, walked away, and Fable built the whole dashboard for her. It's live on her site now, refreshing every morning.

She's in Romania. So she's also one of the people who lost it on Saturday.

According to Greg Isenberg, the smart move from here is to get good at local models. The ones you download and run on your own machine so nothing can ever switch them off. He's right that it's a real option now. You can run capable models at home, and his thread walks you through how to start. Alex Finn went even further encouraging you to own your own compute, before that gets taken away too.

Bans aren't the only risk. Mike Taylor flagged that the cheap era is ending. Right now AI companies subsidise your subscription but that won't last. Microsoft just moved Copilot to pay-per-use, and some bills jumped from $39 to over $3,000 a month. Funny timing, then, that the UK government landed on the answer the same week. What Works for AI Upskilling in the UK studied ten organizations: Airbus, Roche, the NHS and found the same thing every time: the tools were never the problem. Everyone had tools. What mattered was building skill that sticks with you, not one product that's gone in six months. We were one of the ten (still can't believe it) and their numbers, not ours: one team went from 49% to 86% confident using AI.

Easy to say, though. So here are three ways to actually build that confidence with AI regardless of whichever model is up, down, or suddenly charging you:

1. Get good at the skill, not the tool. That 49-to-86 jump was people getting better at directing AI (like what to ask it and how to tell when it's wrong). That works in Claude, ChatGPT, anything. Last week I called it keeping the first and last mile. Try this: run the same task through two models this week, so switching feels normal.

2. Save your prompts, not just the answers. Most people lose their best prompts. You write something that works, get a great result, then never find it again. So when a prompt nails something, paste it into one running note. Five seconds now, and it saves you redoing the thinking later. The wording might need a tweak when you move to a new model, but you're starting from something that already works instead of a blank box.

3. Keep your own data. You know the sinking feeling when an app you relied on shuts down, changes its plan, or locks your stuff behind a paywall? Same risk here. Keep your files, your numbers, and your important docs somewhere you control. That's Isenberg's real point: own enough of your setup that losing one tool doesn't wipe you out.

That's it. None of this is about trusting AI less. It's what lets you lean on it harder, because losing any one tool stops being a big deal.

Before you go ✌️

Did you get to try Fable before it disappeared? And if you did, what did you build, or what were you about to? Hit reply, I read every one.

See you next week!

Window into the Future

There's an answer everyone's landed on. Dylan Field, who runs Figma, put it this week: when execution's cheap, design and creativity are the edge. Swap in "taste," swap in "judgment". That's most of my feed right now. The doing is the easy part, your value is knowing what's worth doing.

Twitter tweet

I buy a version of it. I've built part of a company on it. But it caught on so fast that "taste is the edge" has become its own kind of slop because nobody says what “taste” actually looks like.

Since agentic AI turned up, app releases have shot up but the apps people use stayed flat, and reviews even dropped. There's even a study on it, with a title that gives it away: Writing Code vs. Shipping Code.

The researchers have a name for it: the weak link. AI crushed the easy part like building, so all the slow human parts that come after like getting anyone to use these apps became the bottleneck. Cheap execution didn't hand everyone an edge. It just shows how important distribution is.

The thing that turns AI from a fast typist into better work is knowing what to ask for, catching it when it's wrong, and reworking it until it's yours. You build that by doing the work, not by watching AI do it. a16z had a piece this week about handing a dentist's insurance paperwork to an agent. Fair enough, nobody trains for years to chase reimbursements. But the junior stuck with that job was also learning how the business actually works. Hand it all over on day one and the judgment everyone's calling the edge never gets the chance to form.

Which is how you get what people are calling AI dumping: someone forwards you a doc Claude spun up and you can't tell if a brain was ever in the room. I do it too, when I'm tired. The fix isn't clever. Treat anything AI hands you like a first draft from a sharp intern, and never ship it unread. Argue with it. Cut the lines that sound smart and say nothing. Your name's on it, so the thinking has to live somewhere.

It's also why your feed reads like it was written by one tired robot. LinkedIn said this week it'll start burying posts that sound machine-made including the "it's not X, it's Y" move, now officially a tell (a sentence shape I'll be avoiding for the rest of my life). Ethan Mollick caught the everyday version: it's bleak to open an app and hit a menu full of Claudisms. When everyone runs their thinking through the same model, polished writing stops being proof that anyone thought. And proof is all it ever was.

Twitter tweet

And it's about to get blurrier. OpenAI gave ChatGPT a memory upgrade this week (they’re calling it dreaming) so the more you use it, the more it answers like it knows you. Want to see that line? Drop the same prompt into a logged-out incognito window and compare. The difference is how much was just it having learned you.

So you have to decide what stays yours on purpose.

The rule I've landed on, stolen from Alex Lieberman, who broke down his content process this week: keep the first mile and the final mile, give AI the middle. First mile is deciding what's worth doing and handing over the context. Final mile is the final call. The boring 70% in between, let the machine have it.

A visual inspired by Alex Lieberman’s Content Machine.

So this week: pick one thing you've been handing over whole. Write the messy first draft yourself, let AI clean it up, then do the last read by hand. Keep the two ends. That's the part you'll want to still be good at next year.

For your coffee break 🎧

Last week, John Emmerson had me on his leadership series and I had a genuinely good time. We got into the stuff like why teams buy the right AI tools and still don't change how they work and why you can't give people chaos, you need to give them pathways. Worth a listen if you've got a spare half hour.

How to AI 🤖 

Every week, this section is your shortcut. Here are a few ways you could try AI this week that are worth your time:

• How to connect Claude to your apps

• How to stress test business ideas with Perplexity

• How to get started with Claude Code

• How to build a content machine with AI

Before you go ✌️

What's a bit of your work you still do by hand on purpose, even though AI could take it now? Hit reply, I read every single one.

See you next Sunday!

Window into the Future

First, the money, because it's hard to even picture. Anthropic is now worth around $965 billion, more than OpenAI, and bringing in roughly $47 billion a year. A year ago a number like that would've sounded made up. But the giant number isn't really the part that changes your week.

The same day, they put out a new version of Claude (Opus 4.8, if you're keeping count). The scores nudged up a bit, nothing dramatic. What people care about is the stuff around it.

There's a little dial now that lets you tell it how hard to think, from "quick and cheap" to "really chew on this." There's a faster, cheaper setting. There's a mode where it splits a big job across dozens of copies of itself, all working at once. And there's another change I think is the most interesting one. It's now a lot better at telling you when it's not sure, instead of confidently handing you something that's wrong.

Reactions were all over the map.

Dan Shipper, who runs Every and tests this stuff obsessively, loved it and said it's good enough they may as well have skipped ahead and called it Opus 5. Anthropic's own CTO said the "tells you when it's unsure" part matters to him more than any test score. Then you scroll over to where the heavy users hang out, and they're fuming. They loved the old version, they reckon the new one burns through their usage, and they're roasting it for sounding fake-humble (it keeps opening replies with "let me be honest…"). The best example going around is a guy who asked it for a quick review, and Claude spun up 45 copies of itself and torched his entire usage allowance in 43 minutes.

One reply called it "the new YOLO mode for when you've got usage left before the weekly reset."

Meanwhile, the most-shared setup guide for Opus 4.8 going round this week barely talks about the flashy stuff. It spends most of its time on that little dial, and the big tip is that most people will just leave it cranked all the way up and never touch it. The actual trick is turning it down for the easy stuff and only leaning on it for the hard stuff. Which, from the other direction, is exactly what nuked that guy's afternoon. The AI can't sort that out for you. Same with the "tells you when it's unsure" bit. It only helps if you're the kind of person who reads the part where it says so.

Funny thing is, Anthropic's own CEO more or less called this almost a year ago, before any of it existed. On a podcast last summer he said AI would mess up way less often than people, but in weirder, sneakier ways, because it sounds just as sure of itself when it's wrong as when it's right. The fix, he said, was teaching it to double-check its own work. That's basically what this update is. The hard thing he described is now a setting you can switch on.

And the people who actually do this work for a living keep landing in the same spot. Duolingo's CEO spent this spring walking back his big "AI-first" announcement, now openly admitting the stuff still spits out a lot of junk. A founder who's about as pro-AI as they come told his team a line I liked: AI should get you to great work faster, not drop you straight into mediocre. And Klaudia, a team manager at Ocado who just finished our 15 Days of AI for Teams, said it best. The win for her wasn't AI doing her job. It was setting it up to handle the boring lifting so she could put her energy into the part that's actually hers.

So that's the week. A near-trillion-dollar price tag, a new Claude half the internet loves and the other half wants to undo, a house that'll take your startup shares, and a pile of smart people agreeing the AI was never the hard part. There'll be a slightly better one next month, there always is. The thing I'm watching is whether "when do I lean on this, when do I check it, and what's even worth making" starts to feel like the real skill. For most teams, not yet.

How to AI 🤖 

Every week, this section is your shortcut. Here are a few ways you could try AI this week that are worth your time:

• How to buid a Single Company Brain using AI

• How to use Claude Code as a video editor

• How to build an AI-native paid Ads system

• How to get maximum quality for minimum cost from Claude Opus 4.8

Before you go ✌️

Quick gut check. When AI hands you something, do you actually read it before you pass it on? When's the last time you nearly sent something and caught a problem just in time? Hit reply, I read every one.

See you next Sunday!

Window into the Future

The thing I really can't stop thinking about isn't the cuts though. It's that the two stories about them landed in completely opposite directions, and both of them got passed around like crazy.

Intuit's CEO Sasan Goodarzi went on CNBC and said that "None of it had to do with AI. Everything was about how we become more effective." He framed the 3,000 cuts as cutting management layers, closing some offices, simplifying the org. Now, whether you fully buy that is another thing. Intuit also announced multi-year partnerships with OpenAI and Anthropic the same day. But the choice to not reach for AI as the public reason, when reaching for it is basically free this year, is one of the more interesting things any CEO has done in 2026.

ClickUp's Zeb Evans went the other way. Like, the complete other way. He laid off 22% of his company by way of a 2,000-word manifesto on X about something he calls the 100x org, with million-dollar salary bands for people who "create outsized impact using AI", plus a line about how "most companies are delusional" if they think they can keep iterating on what they already have. Cuts and manifesto in the same tweet, on the same morning Meta was sending out its 8,000 notices. By the next day, it was already a meme.

It's kinda funny that ClickUp posted this 8 days before laying off 22% of its staff expecting a 100x return.

Original video by greg_tube, hailey_janemarie and brittanyroo_

LinkedIn

And honestly, if you kept scrolling, the same split was showing up everywhere. Elena Verna, a growth advisor working with a bunch of scale-ups, wrote a piece this week saying most people in tech won't have the same job by the end of 2027. Dan Shipper, who runs Every, wrote pretty much the opposite from inside his own company. They've automated everything they possibly can and somehow grown from 4 humans to 30.

Which brings me to the practical bit. There's been a lot of advice flying around your feed this week about what to actually do with all this. I want to walk you through the three takes I think are worth your time.

1. Claire Vo, three-time CPTO, currently building ChatPRD. Dropped a tactical playbook right after Coinbase that picked up a lot of reactions. Her advice if you find yourself laid off: download Codex and Claude Code, build a dozen skills, ship them to GitHub, throw together a portfolio site with an agent per skill, post the link, then start messaging the founders of every "AI for [your job]" startup you can find.

2. Lenny Rachitsky, who many of you probably already read, came at it from a totally different angle in his most recent State of the Product Job Market piece. His take basically: ignore the headlines for a second and look at the actual hiring data, because the data tells you a different story. PM openings are at their highest level in over three years. The companies adopting AI properly aren't shrinking. They’re going out and hiring more people, especially ones with what he calls AI fluency. So basically the panic is louder than the data. Go figure out where the hiring actually is before you spiral.

3. Holly Lee, who spent years as a global recruiting leader at Amazon, Google, Meta and Microsoft and now coaches careers, made a point earlier this year that I think holds up best this week. "Layoffs shock people's self-esteem. But you've got to be innovative and open-minded. Because when the market picks up, and someone wants to hire you, they'll ask: 'What did you do during that time?' You can't say, 'I've been applying for thousands of jobs.'" So what Holly is suggesting is to make whatever gap you've got visibly count for something.

Someone trying to hire you in six months wants to see what you did with this moment. And honestly, the same thing applies if you haven't been cut, which is most of us reading this. The people I see getting ahead are the ones who've had AI on their laptop every single day for a while, doing real work alongside them, well before any of this week's news hit their inbox.

There's actual data on what that looks like, by the way. Asana's CMO Shannon Duffy shared this research her team did with Anthropic. The jump from weekly to daily is bigger than the one from never to weekly. Most of what makes someone genuinely AI-native isn't which model they're using or which course they took. It's whether the tool is in their hands every single day, on the work they were already going to do anyway.

Join a 5-day Build Sprint 🛠️

If you've been meaning to actually build something, this is your sign. Harold and Ash are running a five-day vibe coding sprint (May 27–31) made for non-technical people, the whole point being that you can ship a working app without writing code.

There are live workshops along the way too, then a demo day to show what you made. Sign up here to join The Vibe Coding Games.

Alongside the sprint, the team will also host live workshops full of practical tips, AI workflows, and live demos.

🗓️ 27 May | Kickoff: Build Sprint Starts
🗓️ 29 May | Workshops
🗓️ 1 June | Build Sprint Deadline
🗓️ 9 June | Final Demo Day

How to AI 🤖 

Every week, this section is your shortcut. Here are a couple of ways you could try AI this week that are worth your time:

• How to design using Nano Banana Pro

• How to run a workday in 47 mins with Claude Cowork

• How to make Claude brutally honest

• How to audit Claude’s context for you and your work in 15 mins

Before you go ✌️

Which of the three takes lands for you this week, Claire's, Lenny's, or Holly's? Hit reply, I read every one.

See you next Sunday!

Window into the Future

So if the tools were settled in five minutes, what filled the rest of the room? The culture. And under all of it, one question nobody could really answer: what does good AI use actually look like here? Some treat heavy AI use as a kind of cheating. Others treat it as survival. Often it's the same company and sometimes even in the same team.

Zapier did something about that. They made AI fluency a requirement for every new hire, then published an update to the rubric they score people against and the telling part is what they added. The original test had three parts: mindset, strategy, building. The new fourth part is accountability.

Their line for it: with AI you can hand off the work, but not the responsibility for it. "Capable" (you use AI with purpose and can show real impact) used to be a decent score but now it's the minimum. And the tier below it, the one they call unacceptable, sounds almost exactly like Romero on slop: work that reads like obvious AI, and looks the same before AI as after.

So who owns this? I argued recently that the most underrated AI job in most companies sits in HR. Not because HR is technical, but because it already touches every employee and every review cycle. Zapier seems to agree. Their Chief People Officer, Brandon Sammut, is now Chief People and AI Transformation Officer, and that rubric runs through every hire.

McKinsey's new Rewired makes the same case in consultant language. Every digital transformation is really a talent transformation and adds the uncomfortable part: as AI takes on more, teams get smaller, and someone has to plan for that honestly. You have to decide what good looks like then build it into how you hire and promote.

The good news: a rising bar is progress. For years "I used ChatGPT" counted as an achievement. The bad news is a bar also sorts people.

Deedy Das wrote a much-shared post on the mood in San Francisco where a small group got rich fast and everyone else is wondering if the ladder they're climbing is even the right one. Jasmine Sun was circling the same thing on the fear of a "permanent underclass". So if "good at AI" becomes a hiring filter, it matters who got a fair shot at getting good and who didn't.

And maybe not everyone needs to be good at AI at everything. Eleanor Warnock makes the point that enjoyment counts for something. The baker doesn't want to vibe-code an inventory tool, they want to make a great croissant. Claude for Small Business will take that kind of work off their plate, and that's genuinely useful. But Warnock's instinct is right: being able to do something and wanting to do it are not the same thing.

(Every team I talk to now have the tools and the budget but are missing a shared answer to "what does good look like here". Ianos, in delivery management, finished our 15 Days of AI this week and posted that the surprise was how much of it showed up in his actual day. So yes buying the tools is easy but the capability and the measuring are what we keep building 15 Days of AI around.)

Which brings me back to that fake Monet. The crowd were handed a label and confidently picked apart flaws that weren't there. Companies are about to do a version of that. Deciding what counts as good AI use before anyone's agreed what they're even looking at.

Join a 5-day Build Sprint 🛠️

If you've been meaning to actually build something, this is your sign. Harold and Ash are running a five-day vibe coding sprint (May 27–31) made for non-technical people, the whole point being that you can ship a working app without writing code.

There are live workshops along the way too, then a demo day to show what you made. Sign up here to join The Vibe Coding Games.

Alongside the sprint, the team will also host live workshops full of practical tips, AI workflows, and live demos.

🗓️ 27 May | Kickoff: Build Sprint Starts
🗓️ 29 May | Workshops
🗓️ 1 June | Build Sprint Deadline
🗓️ 9 June | Final Demo Day

How to AI 🤖 

Every week, this section is your shortcut. Here are a couple of ways you could try AI this week that are worth your time:

• How to Build a Personal AI System With Claude

• How to automate marketing assets with Chatgpt Image 2.0 

• How to do Prospect Research using AI

• How to master Claude Code in 30 mins (workshop from Anthropic)

Before you go ✌️

Has anyone where you work actually said out loud what good AI use looks like? Or is everyone just guessing? Hit reply, I read every one.

See you next Sunday!

Window into the Future

While all that was unfolding, Claude landed inside Microsoft 365 (Excel, PowerPoint, Word, Outlook) with one conversation carrying context across all four. ChatGPT landed inside Sheets and Excel. And GPT-Realtime-2, a voice model with GPT-5-level reasoning baked in, is openly aimed at the call centre.

Nothing to install. It's just there.

AI used to be a separate tab. You'd open Claude, paste your context, ask the question, copy the answer back. There was always a moment of switching that filtered out the casual stuff. If it wasn't worth opening another tab, you didn't bother. That filter is now gone.

So if you're wondering how to read all that, here are three habits I keep coming back to that are not new but very useful:

1. Think before you prompt.

Old advice but there's a recent study Think First, ChatGPT Later that puts numbers on it. The think-first group retained more, made better decisions on follow-up tasks. The AI-first group was faster on the original task and worse on everything after. We've covered the difference between using AI and getting better at it before. The research is finally catching up. (Vlad at Lucanet put it well after he finished our 15 Days of AI for teams this week and posted that what stuck for him were the frameworks, not the tools, which tracks.

2. Make it show its working. 

Whatever AI gives you, ask it to walk you through how. What it assumed, what it ruled out, what trade-offs it made. Not for the model. For you. Microsoft's Work Trend Index this year calls this human agency, which is on the nose but accurate. (We did a whole issue on the same idea applied to Skills.)

Painfully accurate. Which brings me to the third one.

3. Decide what shouldn't exist. 

Philippa Hardman wrote this week that when anyone can build a course in an afternoon, the real job becomes deciding which ones shouldn't. Same logic for almost anything you ship now. AI will produce a hundred decent versions of anything in the time it takes to drink coffee so the bottleneck moves to the front of the process. What to make at all, for whom, why this and not the other ninety-nine. An earlier issue made the same point: the teams getting traction aren't the ones using AI most but the ones still asking what's worth doing.

(Speaking of which, we just kicked off a partnership with Ocado Retail to roll 15 Days of AI into their UK team. It’s our flagship enterprise product. If you want it for your team, reply and let's talk.)

That's the texture of the week. Signals in every direction. Pick whichever habit looks doable and try it.

How to AI 🤖 

Every week, this section is your shortcut. Here are a couple of ways you could try AI this week that are worth your time:

• How to research 1, 10, or 100 companies using Claude Code.

• How to vibe code and build your own ideas.

• How to build real products by yourself (no coding skills needed).

• How to use Claude Design for real product design workflows.

• How to design your brand in one weekend.

Before you go ✌️

Now that AI's inside Excel, Word, Sheets, and your inbox, what's the part of your job you'd rather keep doing yourself? Hit reply, I read every one.

See you next Sunday!