Agent Skills
› mohitagw15856/pm-claude-skills
› which-skill
which-skill
GitHub智能路由技能，将模糊需求匹配至最合适的具体工具。提供首选建议、备选方案及多技能工作流推荐，通过识别交付物和搜索目录实现精准分发，避免用户手动浏览海量选项。
触发场景
用户不确定应使用哪个技能
描述任务但未指定具体技能
请求可能对应多个相似技能
安装
npx skills add mohitagw15856/pm-claude-skills --skill which-skill -g -y
SKILL.md
Frontmatter
{
"name": "which-skill",
"description": "Route a fuzzy request to the right skill in this library. Use when the user is unsure which skill fits, asks 'which skill should I use for X', describes a task without naming a skill, or when a request could plausibly match several skills. Produces a best-fit recommendation with the inputs to gather, a runner-up with the tie-breaker, and a workflow recipe when the job spans multiple skills."
}
Which Skill Router
Given a fuzzy professional ask ("my boss wants an update on the Q3 launch"), pick the single best skill in this library to run — and say why — instead of making the user browse 400+ options.
What This Skill Produces
- The best-fit skill for the request, with a one-line justification
- The inputs to gather before running it (from that skill's Required Inputs)
- A runner-up skill and the tie-breaker that separates them
- A workflow recipe recommendation instead, when the job genuinely spans 3+ skills
Required Inputs
Ask for (if not already provided):
- The task in the user's own words (even one sentence is enough)
- Who the output is for (audience changes the pick: a board deck is not a team update)
- One-off or recurring? (a monitor/briefing skill differs from a one-time analysis)
Routing Method
- Name the artifact. What lands on someone's desk when this is done — a PRD, a ranked list, a briefing, a plan? Route on the deliverable, not on topic keywords.
- Search the catalog — never route from memory. Read
SKILLS.md(the auto-generated listing grouped by domain), or search withnpx pm-claude-skills list/ the MCPsearch_skillstool. Match the user's phrasing against skilldescriptiontrigger phrases. - Prefer the specific skill over the general one. A skill built for the exact artifact (e.g.
ab-test-readoutfor analysing a finished test) beats a broader neighbour (experiment-designer). - Check the disambiguation table below for the known look-alike clusters before answering.
- Escalate to a workflow recipe (see
WORKFLOWS.md, e.g./ship-a-feature,/launch-a-product) when the ask needs 3+ chained skills — don't recommend the skills one by one. - Recommend, don't interrogate. Ask at most one clarifying question, and only when the answer would change the pick.
Disambiguation Table — look-alike clusters
| You want… | Use | Not |
|---|---|---|
| A one-off deep teardown of a rival (SWOT, positioning map) | competitor-teardown |
competitive-analysis |
| A full landscape doc: feature matrix, win/loss, battlecard inputs | competitive-analysis |
competitor-teardown |
| A recurring "what changed in the market this week/month" briefing | competitive-intelligence-monitor |
competitor-signal-tracker |
| A read on one specific competitor announcement | competitor-signal-tracker |
competitive-intelligence-monitor |
| Release notes straight from a raw git log / commit list | changelog-generator |
changelog-writer |
| A Keep-a-Changelog entry from an already-curated change list | changelog-writer |
changelog-generator |
| Positioning, messaging pillars, use cases — the GTM content | go-to-market |
go-to-market-planner |
| A tiered launch plan with cross-functional coordination — the GTM operation | go-to-market-planner |
go-to-market |
| Themes from interview transcripts specifically | user-interview-synthesis |
user-research-synthesis |
| Synthesis across mixed sources (surveys, feedback, transcripts) | user-research-synthesis |
user-interview-synthesis |
| Pure RICE scoring of a backlog | rice-prioritisation |
feature-prioritisation |
| Choosing/applying a framework (RICE, MoSCoW, Kano, ICE) | feature-prioritisation |
rice-prioritisation |
| RICE blended with strategic-fit weighting | rice-impact-matrix |
rice-prioritisation |
| A summary of an existing document for executives | executive-summary |
executive-update |
| A standalone product briefing written for the C-suite | executive-update |
executive-summary |
| A BLUF-style project status update for stakeholders | stakeholder-update |
executive-update |
| Designing an experiment before it runs (sample size, guardrails) | ab-test-planner |
ab-test-readout |
| Analysing a finished test and making the ship/no-ship call | ab-test-readout |
ab-test-planner |
Output Format
Skill Recommendation
Best fit: skill-name — [one line: why this artifact matches the ask]
Before you run it, have ready:
- [input 1 from that skill's Required Inputs]
- [input 2]
Runner-up: other-skill — pick this instead if [the tie-breaker condition].
Run it: /skill-name in Claude Code, or open it in the Playground.
(If a workflow fits better) This is a multi-skill job — run /recipe-name (chains a → b → c), because [why the chain beats a single skill].
Quality Checks
- The pick was verified against the live catalog (SKILLS.md / search), not recalled from memory
- Every look-alike cluster the ask touches was checked against the disambiguation table
- The recommendation names the concrete artifact the user will get, not a topic
- The runner-up includes a real tie-breaker condition, not "also good"
- Multi-skill jobs point to one workflow recipe, not a list of 4 skills to run manually
Anti-Patterns
- Do not recommend more than two skills — a router that returns a list has not routed
- Do not route on topic keywords ("competitor" ≠ always
competitive-analysis); route on the deliverable - Do not ask a chain of clarifying questions — one at most, and only if it changes the pick
- Do not invent skill names — if nothing in the catalog fits, say so and suggest
SKILL_REQUEST.md - Do not recommend a general skill when a specific one exists for the exact artifact
版本历史
- a38bc30 当前 2026-07-05 11:46
