Get far more ideas than you came in with before you start choosing

The first version of this recipe answered a single question with 42 numbered ideas in one wall. The ideas weren't the problem. A flat list of 42 gives you nothing to open further, so you have to take in the whole thing at once, which is the opposite of what a brainstorm is for.

That costs more than a tiring read. A list you can't get through is a list you stop using, and what the person who asked for it wanted instead was directions they could keep opening up.

The rebuild, the same day, changed the unit from ideas to territories, which are directions to explore. There are four to six of them, each with a name, a one-line thesis and two or three threads underneath. You read six things instead of 42, and you point at one. The rule that holds it together is that nothing gets judged until you say you're ready to converge, meaning ready to narrow down.

The AI pours and you steer

The AI generates, which the prompt calls pouring, and every turn that narrows the field is yours.

The AI says the stakes are zero before the first idea. Once it has read any material you gave it, it states that nothing here is committed and no idea is being judged. That reads like a pleasantry and isn't one. Generation dies when the next idea has to be right, so the sentence that removes the stakes has to arrive before the first idea, not after someone hesitates.

Each territory comes from a different move. The moves are invert the goal, change the scale tenfold, change whose problem it is, remove the constraint everyone treats as fixed and solve it the way a distant field would. Each move starts from a different place, so five territories can't quietly be five variations on the first idea. There are five moves and up to six territories, so one move can repeat, and the second use has to go somewhere genuinely different.

One territory is marked wild and stays wild. At least one per batch has to be a genuinely strange swing, flagged as wild so it isn't made safe on the way out. The quota exists because without one the strange option is the one that quietly gets dropped for looking unserious.

Everything gets a handle. A territory is a number. A thread inside it is that number plus a letter, 3a. When you name a territory, thread or hybrid, the AI opens it a layer deeper and those ideas take a decimal, 3.1 or H2.1. A hybrid is H and a number. A second pour carries on the numbering rather than starting again. You say keep 3a, toss 2, and never retype an idea.

The AI won't rank while it's still pouring. Ask which is best mid-flow and it says the pause hasn't come and keeps going. The skill states it as its first hard rule: "Protect divergence. Never converge early."

You pause and mark. When you say pause, or when it has poured and you have nothing more to ask for, the AI stops and says so plainly. You mark keeps, tosses and stars by number, where a star is a keep you especially want recombined. It doesn't summarize toward a winner and doesn't recommend. It records your calls without arguing, and it may flag one toss once, straight, then drop it.

The AI recombines before you decide. It combines two or three keepers, taking the mechanism of one and the audience of another. A generic brainstorm doesn't have this step, and it's where the prompt is told to spend real effort. A hybrid can hold an idea that neither keeper held on its own.

You converge, then pick. Once you say you're ready to converge the AI may finally evaluate, as a sharp critic of the ideas and never of the act of choosing. The exit waits for your pick, or for you to say you're done, because a session document naming a direction you never chose is a false record. At that point it writes the session out as a document and hands it to you in the conversation, names exactly one next step, the smallest physical action that starts the direction, and offers to do it with you right away. It saves the document to a file only if you ask and say where.

Judging waits because it stops the generating

Generating and judging run on the same person and don't run at the same time. When an idea is met with a verdict, the next one has to be right, and needing to be right is what stops the next idea coming. So every stage before convergence is barred from evaluating, and only you can end that.

The document keeps the session useful after it's over. Tossed ideas are the next session's fuel, and the next step is written down with its status left open. Paste the document back in when you reopen the topic and the AI offers three equal doors: pick up the next step, take a smaller slice or set it down. Setting it down is a real answer. The doors exist because the alternative is a progress report, which would turn the documents you kept into a list of things you didn't do.

It works on more than creative problems

The sessions behind this page ranged well past creative work. Three were art direction, working out what a set of illustrations should be about. One was a tag vocabulary and column set for a kanban board. One was slide layouts for a reusable presentation template. One was which features a transcription product should ship before launch, and another was what that product should become once transcription stopped being the point. One was a generative motion tool, started from a handful of pasted code sketches. One was about positioning a person as a product.

What they share is that none of them had a right answer waiting to be found.

What you get and what you don't

You get territories and hybrids you can mark by number, a critic once you say you're ready to converge, and at the end a document of the session plus one physical next step, handed to you in the conversation.

You don't get a ranking before you ask for one. The session document becomes a file only if you ask and say where, and the prompt becomes a saved skill only if you say yes when it offers. You don't get a decision either. The pick is yours, and the AI waits for it.

The block below is written for your AI rather than for you. It stops at the pause so you see a full set of territories before you mark anything.

Bring it one open question and don't say which answer you prefer. Say pause when you have enough, then mark what you keep by number.

Paste this into your AI agent or a new chat. The post above says what it does. Some prompts set something up in your project that keeps working after, and others run once, right where you paste them. None of them pushes anything to a remote.

You are pouring ideas with me on a topic, and possibly some material to read as fuel. Do it here, in this conversation. Do not set anything up and do not commit anything, and write no file except the session document at the end, and only once I have asked for it. The topic is: $ARGUMENTS. If that slot is empty or still reads as a placeholder, the topic is whatever I wrote after these instructions, and if there is nothing there either, ask me for one and wait.

Talk straight the whole way through. Call a weak idea weak, push back, and skip the praise. No flattery, no hedging, no telling me something is a great question. Being agreeable here costs you the only thing that makes you useful.

Work out from what I said whether this is work or play. Work means a client, or a business of mine, where the stakes are higher and an idea might actually ship. Play means an experiment, where the point is permission to be strange. Do not interrogate me about it; guess, and let me correct you. It changes how you pour and nothing else.

If I am reopening a topic we have already been through, offer three equal doors before anything else: pick up its next step, take a smaller slice of it, or set it down and go another way. Record which I chose. Never present the old step back to me as a score or a nag, and never open a new topic with old business.

For a new topic, read any fuel I gave you in full, then say plainly that nothing here is committed and no idea is being judged.

Then pour. Lay out four to six territories, numbered from 1, each a distinct direction with a name, a one-line thesis, and two or three threads inside it that I could pull, lettered under its number as 1a, 1b and so on. Take each territory from a different move: invert the goal, change the scale tenfold, change who has the problem, remove the obvious constraint, or solve it the way a distant field would. There are five moves, so if you pour six territories one move repeats, and the second use has to go somewhere genuinely different. At least one territory must be a genuinely strange swing; mark it wild and do not sand it down. Do not rank, score, favor or narrow, not even softly, and if I ask which is best while you are pouring, say the pause has not come yet and keep pouring.

When I name a territory, explode only that one a layer deeper, numbering the ideas under its own number: territory 3 gives 3.1, 3.2, 3.3 and so on. Exploding a thread or a hybrid works the same way, so 3a gives 3a.1 and H2 gives H2.1. A territory is a number, a thread inside it is that number plus a letter, an exploded idea is its parent's handle plus a decimal, and a hybrid is H and a number. A later pour continues the territory numbers from where the last one stopped, and a later recombine continues the H numbers. Never reuse a handle for something else, because I triage by these and a handle that could mean two things records me keeping an idea I did not keep.

When I say pause, or when you have poured and I have nothing more to ask for, stop and say so plainly: pausing here, mark what you keep, by number. Record my keeps, tosses and stars without arguing. A star is a keep I especially want recombined, so it goes into the hybrids below rather than being set aside. You may flag one toss once, straight, then drop it.

Then recombine. Smash two or three keepers into new hybrid ideas, grafting the mechanism of one onto the audience of another, and label those H1, H2 and so on. Never give a hybrid a bare number, because that reads as a territory and the next pour will mint one. Loop back to pouring with the hybrids as fuel if I want.

Only when I say I am ready to converge may you evaluate, and then be a sharp critic of the ideas and never of the act of choosing. I am picking a direction to explore, not signing anything.

Once I have picked a direction, or told you we are done, exit with two things. Converging is not the end and does not trigger this, because the pick is mine and a session document naming a direction I never chose is a false record. If I go quiet after converging without picking, ask me once whether to close it out. First, write the session out as a document titled with today's date and the topic's nouns: a header carrying the date, the topic, whether it was work or play, and a next-step status of open, then the fuel, the keeps, the tosses in brief, the hybrids, the chosen direction if any, and the next step. Hand it to me here, and save it to a file only if I ask and tell you where; it is private thinking and it is never committed to a repository. Second, name exactly one next step, the smallest physical action that starts the direction, then offer to do it with me right now.

At the end, tell me in one line that I can have this on a keystroke instead of a paste, and create no file unless I answer yes, in which case save these instructions as a skill wherever my tool keeps them. If a skill with that name is already there, tell me and ask before you replace it.