ideate
Generates, stress-tests, and evolves ideas for a stated problem with research-grounded divergent/convergent loops — persona generators, independent judges, recombination — at quick or deep depth. Triggers on "brainstorm", "give me ideas", "ideate on", "what could we build", "/ideate".
mthines · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add mthines/agent-skills --skill ideate技能说明
name: ideate description: > Generates, stress-tests, and evolves ideas for a stated problem with research-grounded divergent/convergent loops — persona generators, independent judges, recombination — at quick or deep depth. Triggers on "brainstorm", "give me ideas", "ideate on", "what could we build", "/ideate". disable-model-invocation: false argument-hint: '[quick|deep] [--n ] [--no-framing] []' license: MIT metadata: author: mthines version: '1.1.0' workflow_type: orchestrator tags: - ideation - brainstorming - divergent-thinking - idea-evaluation - creativity - agent-loops - confidence-gate - innovation
Ideate
Turns a problem statement into a small set of validated, evolved ideas by simulating a nominal group of independent generators, scoring with independent judges, and breeding the winners — instead of asking one context for a list and polishing it.
This
SKILL.mdis a thin index. Phase procedures live inrules/*.mdand load on demand. The evidence behind every numerical default lives inreferences/ideation-research.md— section references below (§) point there, and its defaults tables flag which numbers are study values vs the skill's own operationalizations.
Contents
- Mode Detection
- Workflow
- Quick vs deep
- Composition
- Core Principles
- Self-Improvement
- Anti-patterns
- Definition of Done
Mode Detection
Parse $ARGUMENTS:
| Mode | Default | Trigger |
|---|---|---|
quick | quick token, or auto-triage says small. | |
deep | deep token, or auto-triage says open-ended. | |
| auto | yes | No mode token — run the depth triage table below. |
| Flag | Default | Meaning |
|---|---|---|
--n <count> | unset | Optional cap on the Lead-finalists list. Unset by default — the finalist count is emergent from the quality bar (see idea-scoring.md § Selection rules). --n only narrows; it never manufactures finalists the bar did not admit. |
--no-framing | off | Skip Phase 1 (problem is already well-framed). |
Everything else in $ARGUMENTS is the problem statement.
If it is missing, ask for it — never ideate on a guessed problem.
Depth triage (auto mode)
First match wins:
| # | Signal | Depth |
|---|---|---|
| 1 | User asks for thorough / extensive / "really new" ideas, or the problem combines ≥ 2 domains ("X meets Y"). | deep |
| 2 | Outcome is a product, business, strategy, or roadmap decision. | deep |
| 3 | Problem is a local decision: naming, small feature shape, workaround, single component. | quick |
| 4 | User is mid-conversation and wants options fast ("any ideas?", "what do you think?"). | quick |
| 5 | Unsure. | quick, and name the escalation path: "run /ideate deep <problem> for the full pipeline". |
Cost expectation: a deep run with a typical 2–4-idea finalist band dispatches ~17 subagents (2 bursts × 5 generators, 1 pool judge, 1 breeder, 1 variant judge, 3 panel judges, 1 pre-mortem) — prefer quick for casual or budget-sensitive asks. A richer pool that admits more finalists adds only panel-judge cost, not generation cost.
Workflow
| Phase | Name | Rule file | Gate |
|---|---|---|---|
| 0 | Intake & triage | (inline below) | Problem restated in one sentence; success criterion named; depth chosen; lessons read. |
| 1 | Frame | (inline below) | 3–5 "How Might We" framings at different widths; one selected and logged. |
| 2 | Diverge | rules/divergence.md | Pool ≥ 12 (quick) / ≥ 25 (deep) unique ideas; zero evaluation happened during generation. |
| 3 | Score | rules/idea-scoring.md | Every pooled idea scored on 4 independent axes by a non-generator judge. |
| 4 | Evolve | rules/evolution-loop.md | ≤ 3 rounds; stopped on flat external scores, never on self-assessed improvement. |
| 5 | Validate | rules/idea-scoring.md § Finalist validation | Every finalist has an executability probe; confidence(analysis) ≥ 70 on the recommendation. |
| 6 | Report | templates/ideation-report.md | Report emitted; finalist count reflects the quality bar; high-novelty wildcard included; "Ideas worth revisiting" filled; verdict question asked. |
| 7 | Learn | rules/self-improvement-loop.md |