Why one-chat brainstorming disappoints
Ask a single chat for ideas and three things bias the output toward mush:- Typicality — alignment training makes the most familiar ideas the most probable (mode collapse).
- Anchoring — every idea is generated after, and conditioned on, the previous ones. The list converges as it grows.
- Self-censorship — one call playing both generator and critic pre-filters toward defensible, safe suggestions.
The manual workflow (works in any AI chat)
1
Write the problem once, cleanly
One paragraph: the goal, the constraints, what “good” looks like. Strip hints toward your current favorite solution — they anchor.
2
Open 4–6 fresh chats and force a vantage point in each
Same problem, different lens per chat. Steal these from ADHD’s frame library:
- “You think in latency, memory layout, physical constraints. Re-ask this as a hardware problem.”
- “You audit systems for failure modes. What must be provable or refusable here?”
- “You are a hostile competitor. How would you break the obvious solution? Invert into ideas.”
- “Ask the OPPOSITE question — how would we guarantee failure? Negate each answer back.”
- “$0 budget, one hour. What’s the crudest version that still works?”
3
Ban the obvious and forbid evaluation
Append to every chat: “The first three answers everyone would give are banned. Do not evaluate, rank, or hedge — just generate 6 short, distinct ideas.” Generation and judgment must not share a pass.
4
Converge in one final session
Paste everything into a fresh chat: “Score each idea 0–10 on novelty, viability, and fit. Flag any idea that looks attractive but is a trap, with a one-line reason. Cluster by underlying angle. Commit to a top 3.” Demand an opinion — “here are 20 ideas, you decide” is a cop-out.
The automated version for coding agents
If you work in Claude Code, Cursor, Codex, or another agent, the ADHD skill runs this exact loop as parallel isolated agent calls — 5 frames, ~30 ideas, scored, clustered, traps pruned, top 3 deepened into sketches with risks and first steps:When NOT to brainstorm with AI this way
Skip the ceremony for lookups, bugs with a known root cause, or anything with one canonical answer — a direct question wins there. The test: if a junior would Google it and find the answer, ask directly. If a senior would say “let me think about this differently for a minute,” diverge. More in When to use ADHD.Frequently asked questions
How many ideas should I generate before converging?
Around 20–30 for a real decision (4–6 lenses × 5–6 ideas). Stop early when new ideas start repeating the shape of existing ones — the space is mapped; padding to a number adds noise, not options.Should I show the AI my favorite idea upfront?
Not during divergence — it becomes an anchor and every “alternative” orbits it. Introduce your favorite in the convergence pass and ask the critic to score it against the field on the same axes.Why ban the first three obvious answers?
The obvious answers are the ones you already had — a senior engineer produces them in thirty seconds. Banning them forces the model past the peak of the distribution into the awkward middle, where the ideas you couldn’t have listed yourself live.What if the results are too weird?
Weird is a feature during divergence and a bug after convergence. That’s what the critic pass is for: absurd ideas earn their place by seeding viable ones, then get pruned with a reason. Never ship the wide set raw — always converge.Further reading: How the diverge/focus loop works · The 15 cognitive frames