How IdeaStew works, and why it works that way

A long read, on purpose. The short version is on the front page.

IdeaStew is for the stage before you can explain what you want. You bring half an idea. It throws words at you. You react — keep, spark, drop — and by the end you have something you could hand to anyone: a finished brief, or a shortlist you'd actually use.

The problem this is built for

Everyone has been told how to get more out of an AI assistant: be specific, give context, state your constraints, say what you don't want. It's good advice. It's also slightly circular, because specificity is not a writing skill — it's the result of having already thought something through.

So the useful question isn't "how do I write a better prompt." It's the one underneath: how do you think something through when you don't yet know what you think? You sit down to describe the business, the book, the name, the plan — and what comes out is the vague version, because the sharp version doesn't exist yet. Ask an AI at that stage and you get a generic answer, which is the correct response to a generic question.

IdeaStew attacks that from a different angle, using one fact about how minds work:

That's the whole design. Everything below is machinery in service of it.

A session, step by step

1. Start with almost nothing

Type an idea, or just a few keywords. The bar is deliberately on the floor — "a name for my new bakery", "my business needs another income stream", "a gift for someone who has everything", "plot a historical romance" all work. Vague is expected. Vague is the input this is designed for.

2. Words arrive as cards. React to each one.

Four responses, and each means something different:

Go on instinct. A second a card is plenty, and the speed is a feature rather than impatience — deliberating too long invites you to talk yourself out of things, which is exactly what a first pass shouldn't do. On a phone you can swipe instead of tapping: left keeps, up sparks, right drops.

Spark is the one people underuse and it's the most powerful control in the app. Keep says yes, that's what I meant. Spark says I don't know what that is yet, but go there. The second one is where sessions turn interesting.

3. Set how far the words travel

A slider above the cards runs from close to home to makes you reach, and it changes what gets generated on the very next batch:

• 1 — words you'd expect. Near-synonyms, the direct parts of your topic.
• 2 — one step out. Related practices, tools, neighbouring ideas.
• 3 — a beat of thought away. Oblique connections and metaphors from next-door fields. The default, and usually the best place to sit.
• 4 — surprising, but it connects. Distant fields, structural likenesses.
• 5 — a long reach, but it's there. Provocations from unrelated worlds.

Nothing on this dial is random. Even at 5 every word is one a person could build a link to — what changes is how far it travelled and therefore how much of the connecting you do. Move it whenever the words feel too obvious or too silly.

4. Dig into a word that landed

Click any word you kept or sparked and you get six new words radiating from that one instead of from your original topic. It's how you follow a hunch without abandoning the session.

5. Change the question — that's normal

Click your idea at the top of the screen and you can reword it mid-session. Sessions drift, and the drift is usually the good part: you begin asking one thing and by card thirty you're really asking another. Nothing on the board is lost when you reword — what you've kept so far is precisely why you're rewording — and the next words follow the new wording.

6. Taste the stew

Part way through, a spoon appears under your question. Press it and you get two boxes: your question as you typed it, and underneath, the same question rewritten using what you've kept and sparked so far — with your own words highlighted inside it, so you can see exactly which of your choices did the work. Take the sharper version, edit your own, or ignore it and carry on.

It exists because the finish used to be the only moment the app proved it was working. This is a look at the middle: evidence that the thing in your head is getting sharper, before you've finished.

7. Ask for Ideas

Once you've flicked enough, the Ideas ✦ button lights up. It reads everything you kept and sparked and writes back real, sentence-length ideas built out of your choices. These arrive as cards too, so you triage them the same way — and keeping an idea is a much stronger signal than keeping a word. Ask for ideas as many times as you like; each round is built from everything that's happened since the last one.

8. Turn an idea over

Any idea you keep gets a ⇄ button. It finds one analogy — not a resemblance, but a shared structure borrowed from a completely different field, ending in a line about what that would mean you'd do differently. Press it again for another, from somewhere else again.

9. Finish

Finish reads what you actually asked for and gives you one of two things.

If you commissioned something specific — a name, a title, a gift, a pick — you get a shortlist of eight to twelve candidates, best first, built from the words and ideas you endorsed, each with a line on why it works, ending with the single strongest pick. You asked for a thing; you get the thing.

If you were exploring — a business, a problem, a design space — you get a developed prompt: one self-contained brief you can paste into any AI assistant to take the idea further. One click switches between the two if it guessed wrong.

Why the prompt is the point

This is the part most people skip past, so it's worth being precise about.

The limit on what you get from an AI assistant is almost never the model. It's the brief. A capable model given a thin question produces a thin answer, confidently. And the brief is the hard part, because writing a good one requires exactly the clarity you didn't have when you sat down.

Your finished prompt is better than the one you would have typed at the start, and the reasons are concrete rather than mystical:

It carries decisions you made but never said out loud

You sparked borrowed and dropped storage. You kept commons and threw away platform. Nobody asked you to justify any of that and you probably couldn't have — but it is real information about what you're after, and it ends up in the brief. Preference demonstrated is more reliable than preference declared.

It fixes the vocabulary

The words you kept set the register the assistant works in. Keep commons and drop platform, and what comes back won't read like a pitch deck. That's a level of control most people never get, because it depends on knowing which words you dislike — which you only discover by seeing them.

It keeps your open questions open

The things you sparked go in as live tensions rather than settled points, so the assistant explores them instead of confirming them back to you. A prompt written from scratch tends to hide its uncertainty; this one states it.

It's written for a machine, not for a conversation

People instinctively write to a chatbot the way they'd write to a person — hedged, polite, a bit apologetic, holding back detail to seem reasonable. All of that wastes the turn. A composed prompt is dense and unambiguous, and spends every line on the job.

One more practical thing: the prompt is plain text, and it's yours. Paste it into Claude, ChatGPT, Gemini, whatever you use — or into a document, or an email to a collaborator. It isn't locked to this site, and it doesn't stop working when you close the tab.

Where your work goes

Nothing is lost

Every session saves itself as you go, including the cards you hadn't decided on yet. Leave mid-flick, come back later, pick it up. Every finished prompt and shortlist is kept under Past prompts.

The junk drawer

Star any word, idea or analogy and it goes in a drawer for later. Drawer items age visibly — a fresh one is a spark, an older untouched one an ember, one that's been sitting a long while a museum piece. Not a to-do list; a compost heap you can dig through, and clicking one starts a new session from it.

Projects

Promote anything worth working on into a project. A project holds an editable next-steps checklist, notes, a plan it can draft for you, turn it over for analogies, and put it in a nutshell — a plain-English version written for a human rather than an AI, for when you need to explain it to someone.

Brainstorming a single step

Every open line on a project checklist carries a ✦: brainstorm that step. The session is seeded with the step, and carries the project as background so it knows what the step is a step of. Every word that comes back is a candidate answer to the step rather than something merely on-topic for the project. Finish, and the result attaches to the project under its thinking so far — so a project accumulates a history of being thought about.

Making it know you

Words get better when the app knows something about you. About you has five areas you can fill in — by typing, or by flicking cards about yourself, which is faster. Name your skills and it stops explaining them to you. Name your audience and the vocabulary shifts to match. Name what you're working toward and every round leans that way. It also remembers the words you've already been shown across sessions, so you don't see the same ones twice.

What gets recorded, and why

The app keeps what you type and what you tap — which words you kept, sparked, dropped or didn't recognise. That record is what makes the next round better, what lets a session resume, and what makes the finish possible at all. It's described in full, in plain English, on the what we keep page.

There's a longer-term idea behind it, and it's fair to say it's an ambition rather than a feature today: the record of how you thought is worth something on its own. Which words you reach for, which ones you consistently refuse, how your questions change mid-session — none of that survives anywhere else. Whether that becomes genuinely useful to look at depends on having enough history to be worth looking at, and that takes time.

What it costs, and what it isn't

It's free, and there's no account — no email, no password. Give a first name if you want your work filed under it; skip it and everything still works.

Because every session costs real money to run, there's a daily ceiling. If a day gets busy you may see a note saying the kitchen is full; the next day it's open again.

It isn't a note-taking app, a project manager, or a place to keep anything important. It's a tool for the messy front end of an idea, and it's honest about being early.