Brainstorm (platform skill — single source)

Who runs: the coding agent, not the human. Never tell the user to copy-paste run.sh as the primary path. Never replace this with a solo idea dump and call it a brainstorm — the point is multiple independent personas + cross-model, then a judge.

This directory is the only skill + runner unit. Other paths are symlinks (installed by scripts/dev/link-platform-skills.sh). Do not maintain a second SKILL.md body.

Who runs what

When to load / run

• User wants ideas, options, approaches, or angles on an open-ended question
• User explicitly says 头脑风暴 / brainstorm / 发散 / 想点子

Skip / redirect:

• Concrete diff / existing code → adversarial-review (this skill is generative, not a defect finder)
• Someone's design doc / RFC / ADR / 方案 when the intent is only to evaluate that proposal → default 设计评审 (main-agent structured challenge; host validation.md → Design proposal review). Do not default to a full diverge pass that ignores the proposal
• Same design doc when the user still wants alternatives → this skill is appropriate: phrase problem + extract hard constraints from the doc; prefer challenge-heavy use; full diverge only if the problem space is still open

Three stages (divergent → convergent)

Built-in personas (lenses): Moonshot Innovator · Pragmatic Builder · End-User Advocate · Contrarian/First-Principles. Configurable in lib/personas.conf — add a lens = one line, no code edits.

Reasoning effort per stage (passed to the provider CLI): diverge=low (favour breadth/speed over deep reasoning), challenge=medium, converge=high (scoring + synthesis is where depth pays off).

Why stateless (no session resume)

Each stage is a fresh provider call; the prior stage's JSON is passed as INPUT, not via claude --resume / codex exec resume. Deliberate:

• Diverge personas must not see each other — independence is what defeats groupthink / anchoring; resume would share one context.
• Resume binds a single provider; brainstorm wants cross-model diversity (personas + challenger + judge on different backends when available).
• Stateless calls stay reproducible and each stage's JSON is schema-checked.

(Same rationale as adversarial-review — see memory adversarial-review-no-resume.)

Agent procedure

1. Phrase the problem in one line; gather any hard constraints.
2. Run (defaults: 4 personas rotated over all available providers, challenger = 2nd available, judge = 3rd available):

   "$TOOL_HOME/run.sh" "<the problem>" --constraints "<optional>"

   Pin backends only if the user named them:

   "$TOOL_HOME/run.sh" "<problem>" \
     --diverge claude,codex --challenger codex --judge grok

   长跑解耦(host 支持后台时)：多角色 × 慢后端可能数分钟。若 host 支持后台执行 （Claude Code 的 Bash run_in_background），后台跑 run.sh 且 --json > <file>， 完成后读该文件再汇报，别同步阻塞会话；host 不支持（如 Codex-TUI）则同步跑并 先给「耗时约 N 分钟」预告。
3. Relay the top ideas + synthesis + key tradeoffs, with the disclosure block below. Honest note if a stage was skipped (single-model / unavailable).

$TOOL_HOME = this directory (where run.sh lives); resolve via $ADV_REVIEW_HOME-style discovery or $HOME/github/wezterm-config/skills/brainstorm.

Options (see run.sh --help)

--problem-file F · --constraints TEXT / --constraints-file F · --personas N (1–4) · --diverge CSV · --challenger P · --judge P · --top N · --json · --dry-run

Mandatory disclosure (paste into 结果)

## 头脑风暴披露
- diverge personas: <N> over <providers>
- challenger: <provider>
- judge: <provider>
- consensus: 单盲评委推荐，非多模型共识（cross-model UNREVIEWED）
- notes / skipped: … | 无   (single-model? stage skipped?)
- top 想法(带 score/verdict) + synthesis + key tradeoffs

Helper commands (agent-only)

"$TOOL_HOME/run.sh" selfcheck claude codex grok
"$TOOL_HOME/run.sh" "<problem>" --dry-run
"$TOOL_HOME/test.sh"           # offline smoke — PROVIDER_MOCK=1, no LLM calls
"$TOOL_HOME/test.sh" --live    # same checks, but against real providers

Install / refresh user-level discovery (idempotent):

./scripts/dev/link-platform-skills.sh

Don't

• Don't ask the human to run run.sh as the main path
• Don't collapse to a solo idea list and call it a multi-persona brainstorm
• Don't use it to review existing code — that's adversarial-review
• Don't default to full multi-persona diverge just to critique one closed proposal (that is 设计评审, not brainstorm) unless the user wants alternatives
• Don't claim cross-model diversity when only one provider was available (the runner marks single-model in notes)

Related

• Runner (this dir): run.sh, lib/ideas-schema.json, prompts/
• Sibling skill (provider layer source): ../adversarial-review/
• Design-doc routing (no dedicated skill): host validation.md → Design proposal review
• Link installer: scripts/dev/link-platform-skills.sh
• Memory: adversarial-review-no-resume (why stateless, not resume)