idea-os Skill
当用户分享一个原始产品想法或问题陈述,希望获得一条结构化流水线——从澄清问题到深度研究、PRD 和分阶段执行计划——并以可继续推进的文件形式输出时使用。这是一个从想法到可构建状态的端到端工作流,不是快速反馈技能(那请用 `idea-refine`),也不是 PRD 编辑器(那请用 `product-management`)。触发语包括 "I have an idea for…"、"help me build X"、"validate and plan this concept"、"what should I build?"——适用于用户明确想要完整多阶段产出物而非一次性回答的场景。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
idea-os
An operating system for turning a raw idea into a build-ready plan. Takes a rough problem statement and produces four files: clarifying questions, deep research, a PRD, and a phased execution plan with platform/stack picks, a user-journey diagram, and kill criteria.
What it produces
Four files in the current working directory (use pwd). Filenames are canonical.
questions.md— sharp clarifying questions (skip if the brief is already dense or run is autonomous)research.md— market, competitors, SWOT, JTBD, distribution, risks, insightsPRD.md— problem, users, scope, non-goals, metrics, solution shapeplan.md— user journey + mermaid, platform + stack, phased build, kill criteria, next steps
Workflow — 5 phases, in order
Do not skip or reorder. Weak inputs produce weak outputs; each phase feeds the next.
Phase 1 — Triage
Classify the idea on two axes. Depth of research, PRD, plan, and question-count scale with tier; vocabulary scales with sophistication.
Idea tier (T1/T2/T3) — complexity:
- T1 (Simple) — single-user utility, weekend hobby, CLI, small app with no network effects
- T2 (Moderate) — multi-user app, SaaS MVP, AI wrapper with distinct users and workflows
- T3 (Complex) — marketplace, two-sided network, B2B SaaS with procurement, regulated or deep tech
Sophistication (S1/S2/S3) — builder vocabulary:
- S1 — first-time builder, non-technical. Explain every term, no frameworks by name
- S2 — hobbyist or junior PM. Introduce frameworks with one-line definitions
- S3 — founder, senior PM, shipped before. Use the full vocabulary
State your classification in one line (e.g. "T2 · S2 — moderate SaaS, builder has shipped before") before proceeding. If ambiguous, ask at most one question.
See references/clarifying-questions.md for the question bank tiered to these levels.
Phase 2 — Clarify
Write questions.md using assets/questions-template.md. Pick 4–18 questions (count scales with complexity) matched to the idea type — SaaS, marketplace, consumer, AI wrapper, dev tool, hardware, content. Group into three buckets: Who and Pain · Scope and Wedge · Constraints and Goals.
Every question must be actionable — the answer must change what you build. Generic questions ("who is your user?") are rejected; force specifics ("describe the last time your target user hit this pain — what did they do instead?").
After writing, stop and wait for answers. Do not research yet.
Skip questions.md entirely if the brief is already dense; state assumptions and proceed.
Phase 3 — Research
Write research.md using assets/research-template.md. Use WebSearch + WebFetch for real signal. Never hallucinate numbers; flag [assumption] or cut.
Pre-flight checklist (a floor):
- ≥5 WebSearches + 2 WebFetches on named competitors (pricing, positioning, traction)
- ≥1 search on market size / category growth with a sourced number
- ≥1 search on "why now" — tech, behavioral, regulatory, or cost unlock
- 1 community scan (Reddit, HN, Discord, FB group) where the target user congregates
- Note the date of every source
Sections required (depth scales with complexity):
- Problem validation →
references/research-frameworks.md - Jobs-to-be-done →
references/research-frameworks.md#jtbd - Market (TAM/SAM/SOM, segmentation, WTP) →
references/market-analysis.md - Competitors (direct, indirect, substitutes; positioning map; feature matrix) →
references/competitor-analysis.md - SWOT →
references/research-frameworks.md#swot - Distribution (first 100 users, channel CAC, founder actions) →
references/distribution.md - Risks
- Insights — 3–7 non-obvious findings that change the build. Most valuable section; no filler.
Phase 4 — PRD
Write PRD.md using assets/PRD-template.md. Anatomy + tier-scaled sections + anti-patterns in references/prd-structure.md. Non-goals section is mandatory — it's where bad PRDs die.
Phase 5 — Plan
Write plan.md using assets/plan-template.md. Must include:
- User journey (text + mermaid) →
references/user-journey.md - Platform recommendation with reasoning tied to research →
references/platform-recommendation.md - Stack recommendation (conservative / modern / cutting-edge matrix) →
references/stack-recommendation.md - Phase breakdown (MVP → v1 → target) with scope, success metrics, kill criteria, distribution per phase →
references/mvp-slicing.md - Metrics per phase (North Star + inputs + counter) →
references/metrics-framework.md - Immediate next steps (3–5 actions for the next 1–2 weeks)
Autonomous mode
When there's no live user (batch run, test, dry-run, or the user said "just proceed"): still write questions.md to make decisions legible, then simulate plausible answers consistent with a stated persona at the chosen tier. At the top of research.md, add an Assumptions in lieu of answers block listing each assumed answer and flagging the load-bearing ones. The reader should see exactly what you decided for them.
Output conventions
- Write to
pwdunless user specifies - Use the target-market currency consistently (USD, INR, EUR, etc.)
- Relative links between files (
PRD) - Header block at top of each file: idea title, date (ISO), tier (T1/T2/T3) + sophistication (S1/S2/S3)
- Render on GitHub — mermaid, tables, clean headings
Top anti-patterns
- Generic SWOT ("good team" — cut)
- Unsourced TAM numbers — source it or flag
[assumption] - PRD with no non-goals — reject
- Plan with no kill criteria — not a plan
- MVP as a feature checklist — MVP is a thin vertical slice, one user one job end-to-end
- Platform recommendation before understanding user context — research first
More in references/prd-structure.md (PRD anti-patterns) and references/mvp-slicing.md (MVP traps).
Example
examples/example-notion-for-dog-trainers.md is a reference-only walkthrough showing the quality bar. Do not load it unless you need calibration — it's 500+ lines.