brainstorm

What it does

Enforces the rule that idea generation and judgement must be separate phases (Guilford), and that LLM ideation must be diversity-forced or it homogenizes (Anderson et al. 2024). It orchestrates the toolkit's ideation/critique agents so you get spread AND a decision, not just a list.

Pipeline

1. Reframe. Turn the request into 1-3 sharp "How might we ___?" questions. Confirm the framing with the user if the problem is ambiguous (a wrong frame wastes the whole pipeline).
2. Diverge — run the ideator agent. Over-generate diverse options using ordinary-persona sampling and axis variation. Do not judge yet.
3. Expand (optional). For an existing artifact, also run scamper-style transformation; for a stuck space, force lateral/analogy ideas. Skip if the diverge step already spans the space.
4. Stress-test. Run premortem (how does the leading direction fail?) and assumption-surfacer (what must be true for it to work?) on the strongest 2-3 candidates. Kill or revise options whose load-bearing assumptions fail.
5. Converge — run the synthesizer agent. Cluster, score against explicit criteria, recommend a shortlist with the "fails if…" condition for the winner.

Orchestration notes

• This skill runs in the MAIN thread, so it can spawn the agent subtasks; the agents themselves cannot. Dispatch ideator/premortem/assumption-surfacer/synthesizer as subagents and synthesize their returns.
• Keep generation and judgement strictly ordered — never let the ideator see the scoring criteria, never let the synthesizer add new ideas.
• Scale to the task: a quick "name this" needs steps 1-2 + a light converge; a "which architecture" needs the full pipeline.

Output

FRAMED: How might we <…>?
DIVERGED: <N ideas across K axes>  (from ideator)
STRESS-TESTED: <which survived; what assumption killed the rest>
SHORTLIST: <2-3 ranked options with criteria>
RECOMMENDATION: <winner> — fails if <assumption>.

Anti-patterns

• Don't collapse to convergence too early — let the diverge step over-generate first.
• Don't skip the stress-test on high-stakes decisions; the cheapest time to kill a bad idea is before building it.
• For a single contested yes/no, this is overkill — use debate.