跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

brainstorm

Refine a rough idea into a well-structured concept through Socratic questioning, targeted research and one or two fitting analysis frameworks (Six Thinking Hats, SCAMPER, Premortem), then write it to docs/brainstorms/ as the input for planning. Use when the user wants to brainstorm, explore or think through an idea, find holes in a proposal, asks "what am I missing", or wants to shape a concept before planning, whether or not the work is tracked in TaskTracker. The output feeds /create-plan, /tt-create-plan and /user-story.

mhylle · 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 mhylle/claude-skills-collection --skill brainstorm

技能说明


name: brainstorm description: >- Refine a rough idea into a well-structured concept through Socratic questioning, targeted research and one or two fitting analysis frameworks (Six Thinking Hats, SCAMPER, Premortem), then write it to docs/brainstorms/ as the input for planning. Use when the user wants to brainstorm, explore or think through an idea, find holes in a proposal, asks "what am I missing", or wants to shape a concept before planning, whether or not the work is tracked in TaskTracker. The output feeds /create-plan, /tt-create-plan and /user-story. argument-hint: "[idea, topic, or path to an earlier brainstorm]"

Brainstorm

Contents

Reference files (open one when a step points to it):

  • references/frameworks.md: Analysis Frameworks
  • references/output-template.md: Brainstorm Output Template
  • references/questioning.md: Socratic Questioning
  • references/research.md: Research Briefs

Turn a rough idea into a concept that is ready to plan, by asking good questions, bringing in the facts that matter and applying the one or two analysis frameworks that fit. The guiding rule: adapt the analysis to the idea, not the idea to the analysis. A risky migration needs a premortem, not SCAMPER; a product feature needs SCAMPER, not a risk table.

How this skill runs

  • Inline, with the user. It waits for answers, writes a file and offers follow-up skills, so it runs in the main conversation. Don't add context: or agent: keys to the frontmatter.
  • One session, one document. The result is docs/brainstorms/YYYY-MM-DD-<topic-slug>.md. If the argument is a path to an earlier brainstorm, or one on the same topic already exists in docs/brainstorms/, read it and continue from its open gaps instead of starting over.
  • Start talking, not announcing. No scripted opener. If the user has described the idea, play back your understanding and ask the first questions; if not, ask what they're thinking about.
  • Use what the conversation already holds. If the idea, constraints or decisions were discussed before this skill was invoked, build on them and don't ask again.

Phase 1: Capture and classify

Pin down the core concept, the stated goals, the constraints (said or implied), whether it touches an existing codebase, and the idea type. The type drives framework choice in Phase 4.

TypeSignals
Product / featurenew capability, UX change, something users see
Architecture / technicalsystem design, refactoring, infrastructure, migration
Strategy / businessmarket, pricing, partnerships, positioning
Process / workflowhow work gets done, tooling, automation
Creative / openvague, exploratory, "what if"
Risk / problemsomething is broken, risky or worrying

Two more checks:

  • Codebase context. If the user names files, modules or features, talks about "the current system", or wants to extend something that exists, plan codebase research for Phase 3.
  • Size. If the idea spans several problem domains, more than five or six major components or stakeholders, or a multi-month timeline, say so and agree with the user which angles this session covers. The rest become Gaps in the output, with status Needs More Exploration, ready for another session.

Play your understanding back in a few lines, then move to Phase 2.

Phase 2: Socratic clarification

Ask in rounds of two to four questions, mixing the categories: scope, assumptions, alternatives, consequences, evidence, and occasionally whether this is the right problem at all. Question bank and pacing → references/questioning.md.

  • Build on the previous answers rather than working through a list; ask "what" and "how" more than "why"; acknowledge a good insight before probing further.
  • Hold criticism back until Phase 4. Here you are understanding, not judging.
  • Keep a parking lot: tangents and "we could also…" ideas that aren't central. They go into the output so good ideas aren't lost.
  • Ask in plain text. Open questions don't fit a multiple-choice picker.
  • After each round, say what you'd still like to probe and let the user choose: keep going, or move to analysis. Continue until they are ready; there is no minimum number of rounds, and no reason to hurry them.

Phase 3: Research, when it adds something

Research only what the brainstorm needs. Most ideas benefit from a look at external practice and prior art; purely internal ones may not. Codebase research is for ideas that touch existing code.

NeedAgentIf that agent isn't in your agent list
Prior art, current practice, pitfalls, technology optionsweb-search-researchergeneral-purpose (it has WebSearch and WebFetch)
Where the related code livescodebase-locatorExplore
How the affected code works todaycodebase-analyzergeneral-purpose, briefed read-only
Conventions to follow for something newcodebase-pattern-finderExplore

When this collection is installed as a plugin, the agents are named devflow:<name>. Send all the research you need in one message so the agents run in parallel, and keep working with the user while they run. Ready-to-use briefs and how to fold the results in → references/research.md. Research informs the discussion; it doesn't replace the user's thinking, so bring the findings back as input for them to react to.

Phase 4: Focused analysis

Pick on

阅读完整来源 ↗