
Hi AInauts,
Welcome to the latest issue of your favorite newsletter.
Anthropic goes public today. OpenAI tried to steal a little of the spotlight by releasing GPT-6.2. The early reactions are lukewarm; Claude Fable 5.2 is still ahead. Anthropic could hardly have asked for a nicer IPO gift.
Claude itself has been busy too: Meetings, Whiteboards, Sheets, and Data. We showed you what they're good for on Monday.
So today, no feature fireworks. Instead, three topics that go against the grain.
All three tell the same story: what you give the AI model matters more than the model itself. For ideas. For design. And for the question of what an AI employee or AI agent actually is.
Here's what's inside:
Let's go!
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
🎨 The 6-3-5 method: Can we make AI creative?
If you've been reading us for a while, you know we love using AI to challenge ideas and develop new ones.
The models have become incredibly powerful and can take over plenty of tasks. One thing they still often do a really poor job of: CREATIVITY.
We're always trying to make the models more creative, testing techniques that have worked for people.
A few weeks ago, we shared the Six Thinking Hats as a prompt. Today, another technique from the toolbox. Plus a test that turned out very differently from what we expected.
What is the 6-3-5 method?
The aim is to generate as many creative ideas as possible about a topic.
Management consultant Bernd Rohrbach invented the method in 1968. Six people each write three ideas on a sheet of paper. Time allowed: five minutes.
Visual note: the German diagram shows six people writing three ideas in five minutes, passing the sheets through six rounds to produce up to 108 ideas without talking.

Then every sheet moves one seat along. The next person reads what's there and builds on it. Develop it, turn it around, combine it.
After six rounds, you have up to 108 ideas on the table.
The trick: nobody talks. Nobody goes first. And nobody pins everyone else to their idea.
Sounds perfect for AI. Six participants you don't have to get into a room first. We tried a few different approaches.
Our test
We wanted to test it on a task of our own that Fable had failed at last week: cover ideas for a guidebook, including a title and subtitle.
The topic: a career restart after 45 in the age of AI.
We tried the method with all kinds of technical setups. From simple to complex. From "Give me 108 ideas" to six models from six providers passing the sheets around through OpenRouter.
Something interesting happened in round one.
Visual note: this German test sheet contrasts repeated titles from six separate Fable chats with more varied ideas from six different models. The translated examples are in the text below.

We gave Fable the same task six times in six chats. And the first idea on all six sheets was: "Halftime. Switch sides."
Lesson: use the same model every time and you get very similar answers.
What happens if we do the opposite and use six different AI models?
Muse writes: "Not too late. Just new."
Grok turns that into: "Plug in port 45 yourself."
Mistral: "Resolder the cables."
GPT: "Rewire your experience."
And Gemini lands on: "Kintsugi for your career."
Plenty of nonsense, plenty we couldn't use. But we're definitely getting more creative.
We then ramped up the test scenarios, hoping for something both creative and good:
1 prompt
6-3-5 in a prompt (one model)
6-3-5 round by round (one chat, one model)
6 × Fable (6 chats)
6 × Fable (6 chats, 6 roles)
6 models
6 models with roles
And who won?
At the end, all six models and we ourselves voted blind on all 756 ideas: which would you print? Which was the most original?
Visual note: the German chart shows 240 AI-jury votes across 756 ideas. In the order of the seven setups listed above, the vote counts are 3, 20, 34, 50, 64, 30, and 39. Fable with roles wins.

Here's the most interesting part:
The method beats the model. "Give me 108 ideas" gets 3 out of 240 votes. The same Fable with 6-3-5 gets up to 64. Fable wasn't the problem.
A single model needs roles. Six separate calls without roles often write the same thing. There are still good ideas, but roles or different models produce far fewer copies.
Six models aren't automatically better. They build the wildest chains. But Fable with roles was even better.
One important thing we saw: the AI jury isn't creative either.
Its favorite title for "would sell": "Starting Over at 50." One of the most boring titles imaginable. But Fable with roles was our clear favorite too.
What that means for you: AI gives you the volume. You still have to choose.
And you don't need an elaborate technical setup. Even one prompt can genuinely help.
The prompt to copy
In a normal chat, one round per message worked best. The model can see everything already on the page and only has to write 18 ideas per round instead of 108. Our result: not a single copy.
We're doing a 6-3-5 brainwriting session based on Bernd Rohrbach's method. You play all six participants.
The task:
[WHAT IT'S ABOUT, WHO IT'S FOR, WHAT MAKES A GOOD IDEA]
How it works:
1. Cast the group: six participants with as little in common as possible (occupation, age, country, way of thinking). One sentence each. At least one comes from a completely unrelated industry.
2. Each has a separate sheet. In round 1, each writes three ideas on their sheet.
3. Then each sheet moves one seat along. Everyone reads the sheet they now have and writes three new ideas below. Build on the ideas above: develop, reverse, combine, exaggerate.
4. Six rounds, until every sheet has six rows.
Important: do exactly one round per response, meaning 18 ideas. Then stop and wait. I write "continue," and you do the next round. Take time with each idea; good is better than fast.
Rules for every round:
- No criticism, no evaluation.
- Wild ideas are welcome. Better too bold than too safe.
- No idea repeats another, even in different words. That includes ideas on other sheets.
- Each idea is 30 words at most, in the participant's way of thinking, not yours.
After round 6: show me the ten strongest ideas, with one sentence explaining why for each. At least three should feel a little too bold to you.After each round, you just write "continue." And under "The task," more context is better. Who it's for, the constraints, what makes a good idea.
Our take: The method beats the model zoo
Our original hypothesis was that six models together would be more creative than one. We got a different result.
A better one, really. We don't have to wire up every model through OpenRouter. We can use one really good model, Fable in our case.
We'll keep exploring the model mix anyway. More creativity techniques are coming soon, too.
Attio is the #1 fastest-growing CRM on Brex and Ramp.
Built for the most ambitious companies in GTM, Attio is an end to end revenue platform that allows teams to run multiple plays across different revenue motions and lifecycles.
Teams at the frontier are already building on it: Granola, turbopuffer, Modal, and many many more.
🖌️ Why all AI designs look the same
Ask Claude to build a landing page or presentation. Looks really good.
But somehow just as "good" as everything else floating around the web right now.
Nate Parrott is the Anthropic designer behind Claude Design and Artifacts. In his interview with Peter Yang, he explains why.
Every request starts in a fresh Claude chat.
And every time, it thinks: I'm going to invent the most brilliant style ever. Serif type, cream background, orange, colored stripes along the edges of cards.
Because, as far as Claude knows, it's only done that once!
Sound familiar? Same here. In the 6-3-5 test above, six separate Fable calls wrote "Halftime. Switch sides." six times.
Nate's five tips
Here are Nate's tips for better AI designs:
Provide context: screenshots, sites you like, your own work.
Use familiar patterns: have Claude look at how established products solve the problem.
Ban Claude's quirks: the model doesn't know its own preferences. Write them down in a design file. Our design tool can help with that (the interface is in German).
Layout first, type and color second: work out the structure before the look. Otherwise, you fall in love with colors on a page that doesn't work.
Five variations, then mix: take the best of all five, then make another five from that.
We tested it
The same task for Fable 5.2: a landing page for a fictional trades-business app.
First, a normal request in three fresh chats. Then Nate's method, all in one chat.
Visual note: these German landing-page mockups compare three similar yellow designs with five distinct alternatives: a DIN-style letter, typography, a Dieter Rams-style device, an Otl Aicher-inspired grid, and a price poster.

The basic request: all three pages had the same layout, yellow everywhere. And the sample customer was called "the Brandt family" all three times. In three chats that knew nothing about each other.
For Nate's method, we gave Claude three things:
Context: German trades businesses, formal German address, German design rather than US-startup styling.
Real content: our sample business, our voice message, our price. Nothing left to invent.
The quirks from round one: the Brandt family, yellow, the same layout. All banned.
Then: design five variations that differ substantially in layout and style. First as short descriptions, then build them one at a time.
We got a DIN-style business letter, a type-only page, a device in the style of Dieter Rams, an Otl Aicher-inspired grid, and a price poster. Five distinct pages.
The trick is the same as in 6-3-5: in one chat, Claude sees what it's already built. In separate chats, it starts from zero every time.
By the way, just copying Nate's ban list into the prompt isn't enough. We tried that too: it changes the paint, not the house.
How to spot your model's quirks
Make the same request in three fresh chats.
Put the results side by side. Whatever appears in all of them is your model's default mush.
New chat: add context and your real content, ban the quirks. Then ask for five fundamentally different variations.
Our take: Your taste belongs in the folder
AI has become really good at design. But it doesn't know it's repeating itself.
Only you notice that. That's why your taste belongs in your folder. More on that in a moment.
Save it once. Provide it every time.
📁 The folder is the AI employee
Recently, we sat down with a team at a midsize company. Not techies. Ordinary people from sales and administration.
Management had given them an assignment: "Think about which AI agents would be useful in your daily work."
We hear this kind of thing all the time. Everyone's making AI agents, so we need some too. Let's think about which ones...
Nobody on the team really knew what that meant. The assumption in the room: surely you'd need the IT department. Coding. A project, integrations, a budget...
Nope. You don't.
We hear this in the AInauten community too, so let's finish with our simplest AI-agent setup. And how you can "build" one yourself right now.
What an agent actually is
Countless tools call themselves AI agents. And in some way, they're all right.
But in our simple, non-techie world, we like this definition: an agent is a model with enough context that you don't have to explain everything from scratch each time.
In other words: an AI employee is a folder with three things in it.
A job description. A text file: what you're here for, what a good result looks like, what must never happen. In Claude Code, it's called CLAUDE.md; in the ChatGPT app, AGENTS.md.
Reference material. Good old examples, price lists, templates, knowledge.
Repeatable processes. Anything that should run the same way every time goes into a Skill.
The inbox operating manual from Monday is exactly this kind of "job description."
Visual note: this folder screenshot includes German folder names. The highlighted instruction files are CLAUDE.md and AGENTS.md; workflows are stored alongside them.

Your AI tool just needs access to the necessary tools and the folder, and you have your own AI agent.
Magic.
Same model, different folder
Let's make that last point clear.
You can use the same model every time. Claude, ChatGPT, Grok, whatever.
In one folder, it wears the developer hat. In another, it handles ongoing operations. Simply because the folder contains something different.
At that midsize company, you'd have a "Quotes" folder with a price list and the best past quotes. Beside it, a "Complaints" folder with goodwill policies and reply templates.
Same model. Two employees. No IT department.
What matters: build it first, use it yourself, then learn to trust it. Only automate after that.
One step at a time. Eventually, you have a whole team of AI employees.
How to start
Create a folder for a task that comes up every week.
Put three to five really good examples in it.
Write the job description.
Open the folder in Claude or the ChatGPT app and give it the task. Write your corrections back into the job description.
It takes half an hour. And if you already have a folder like this, have it checked:
Before changing anything: explain how this folder is organized. What lives where? Which file is the authoritative guide, which is only a copy or draft, and where are there duplicates?
Then tell me honestly: what would you need to know to work here like an experienced employee, and what's missing? Show me the three most important gaps and the smallest step to close each one.
Don't move, rename, or delete anything until I've approved it.Our take: An AI agent isn't an IT project
We've been preaching files over tools since February. Honestly, we thought everyone knew by now.
That meeting proved us wrong. To most people, "AI agent" still sounds like an IT project. And of course it can be.
But it doesn't have to be. It can be very simple.
That's it for today. And let us know what you think about Anthropic stock: leave a comment below or just reply to this email.
Thanks for reading!
Reto & Fabian from AInauten




