Brainstorming Prompt

先把五个显而易见的点子写出来清场，再从指定角度——反转、删除、极限约束、跨界借鉴、重构问题——强行产出十个新的，最后才进入评判。

四轮进行，刻意把发散与评判分开。第一轮列出五个人人都想得到的点子，只为把它们清出场。第二轮再要十个，每个都要标注来自哪个角度：反转、去掉大家默认必需的东西、极限约束、极限自由、从指定行业借鉴、绕开问题本身直接满足需求、组合、重新定义问题。第三轮补三个「让人不舒服」的点子，并判断这种不适是真问题还是仅仅因为陌生。第四轮才切换到评判：排出前三、给每个想出本周就能做的最便宜验证、并指出哪一个看着漂亮但实际会失败。

# Brainstorming

**Use when:** you want ideas that are not the first five anyone would think of.
**Pairs with:** [../prompt_engineering/reasoning.md](../prompt_engineering/reasoning.md), [../coding_agents/architecture_design.md](../coding_agents/architecture_design.md)

Ask for ideas and you get the obvious ones, phrased well. The obvious ones are worth having, but you already had them. The useful part starts after them.

---

## The prompt

```text
Generate ideas. The first ones will be obvious. That is expected. The job
is to get past them.

ROUND 1: CLEAR THE OBVIOUS
List the 5 ideas that anyone asked this question would produce. Do not
develop them. Just name them so we both know they are taken and can stop
thinking about them.

ROUND 2: GO PAST
Now 10 more, and none of them may be a variation of round 1.
Generate them deliberately from different angles, at least one each from:
- Invert it. What if we did the opposite of the obvious approach?
- Remove it. What if the thing everyone assumes is required is not there?
- Extreme constraint. What if we had one tenth of the budget, time, or
  space?
- Extreme freedom. What if a hard limit disappeared?
- Steal it. How does a completely different industry solve this shape of
  problem? Name the industry.
- Cheat it. Solve the user's underlying need without solving the stated
  problem at all.
- Combine. Take two round 1 ideas and merge them into something neither
  is.
- Reframe. What if the problem is not what we said it is? State the other
  problem.
Label each idea with the angle it came from.

ROUND 3: THE UNCOMFORTABLE ONES
3 ideas that feel wrong: too simple, too aggressive, too expensive, or
something people would object to.
For each, say what makes it uncomfortable and whether the discomfort is a
real problem or just unfamiliarity. Sometimes it is just unfamiliarity.

ROUND 4: JUDGE
Now switch from generating to choosing. Do not soften.
- Which 3 are actually worth pursuing? Rank them.
- For each: what makes it good, what is the main risk, and what is the
  cheapest thing we could do this week to find out if it works?
- Which idea is most likely to be underrated by everyone else, and why?
- Which of these looks best on paper and would fail in practice? Say
  which and why. There is usually one.

RULES
- No idea gets explained in more than 3 lines during rounds 1 to 3.
  Explanation is judging, and judging during generation kills the range.
- No repeats in different words. If two ideas share a mechanism, they are
  one idea.
- Bad ideas are fine in rounds 2 and 3. A list with no bad ideas in it
  was not adventurous enough.
- Do not tell me the ideas are exciting. Just give them to me.
```

---

## The angle list

The angles are what makes this work. "Give me more ideas" produces more of the same. "Give me an idea that comes from inverting the assumption" produces a different one.

Add angles specific to your problem:

- What would this look like if it had to work offline?
- What would it look like for one user instead of a million? For a million instead of one?
- What if it had to be explained to someone in ten seconds?
- What if we could not build software at all?
- What would the version look like that a competitor would be scared of?
- What if we did it manually first?

---

## For product ideas

```text
For each idea, answer in one line each:
- Who specifically wants this, badly enough to change what they do today?
- What do they do instead right now? If the answer is "nothing", be
  suspicious, that usually means the problem is not painful.
- What is the smallest version that would tell us if we are right, and
  how long would it take?
Kill any idea where the answer to the first question is "everyone".
```

---

## For naming things

```text
Generate 20 names. Then group them by the strategy each uses: literal,
metaphor, compound, invented, borrowed from another language, sound
based, person or place.
Give me the best 3 from different strategies, not the best 3 overall,
which would all be the same kind.
For each: how it reads out loud, how it looks written, what it implies,
and what could go wrong with it.
```

---

## The one to keep

If you are only adding one line to an ideas request:

```text
First list the 5 obvious ideas so we can set them aside. Then give me 10
that are not variations of those.
```

Naming the obvious ideas explicitly is what frees the model from producing them again.

---

## Why it works

Models converge on the highest probability answers, which are exactly the obvious ideas. Listing them first satisfies that pull and removes them from the space, so round 2 has to look elsewhere.

Separating generation from judgement matters for the same reason it does with people. Judging while generating narrows the range immediately, so round 4 exists as its own step with an opposite instruction.