find-skills

Find Skills

You've been asked to find existing skills relevant to a description. Search the sources below in order, rank by relevance and quality, produce a discovery document, then hand off to skill-creator on confirmation.

Sources

Work through these in order — earlier sources carry richer quality signals:

#	Source	How
1	skills.sh (SkillsMP)	Fetch https://skills.sh/ and search the leaderboard; this is the primary marketplace with install counts
2	Skills CLI	Run bunx skills find {keywords} (fall back to npx skills find {keywords} if Bun isn't available)
3	GitHub code search	WebSearch for site:github.com filename:SKILL.md {keywords} — finds skills not yet in the marketplace
4	GitHub collections	WebSearch known skill collections: ComposioHQ/awesome-claude-skills, vercel-labs/agent-skills, xixu-me/skills, openai/skills
5	Local	Scan installed skills in ~/.claude/skills/, .claude/skills/, and the project's skills/ directory

From each skill found, extract:

• name — skill identifier from frontmatter
• author — publisher, GitHub org, or username
• url — skills.sh page or direct link to SKILL.md
• installs — from skills.sh if available, otherwise N/A
• description — one sentence from frontmatter description: field

Quality signals

Use these to rank, not to hard-exclude:

• Install count — prefer 1K+; flag anything under 100 as low-confidence
• Publisher reputation — known publishers (vercel-labs, anthropics, openai, microsoft, xixu-me) rank higher
• Repository quality — check GitHub stars; treat skills from repos with <100 stars with skepticism

A niche but highly relevant skill from a small repo beats a popular but off-topic one.

Ranking

Score each candidate against the user's description. Weight:

1. Relevance to the task domain (highest weight)
2. Quality signals above
3. Breadth — broader coverage is more reusable

Pick the top 10. If fewer than 10 exist, include all.

Output

1. Write the discovery document

Write to ./docs/skill-discovery.md (create ./docs/ if needed):

# Skill Discovery: {user's description}

_Sources: skills.sh · Skills CLI · GitHub code search · GitHub collections · Local_  
_Add comments anywhere in this file, then reply "build it" to proceed._

---

## Top 10 Relevant Skills

| # | Skill | Author | Source | Installs | Description |
|---|-------|--------|--------|----------|-------------|
| 1 | [skill-name](url) | Author | skills.sh / GitHub / Local | 12K / N/A | One sentence |
...

---

## Synthesis

[3–5 thematic paragraphs in the style of a literature review. Group skills by what they
emphasise rather than listing them one by one. Note variation in approach, scope, or
workflow philosophy across the set. Write in flowing prose — no sub-headers, no bullets,
no gaps or recommendation section. The goal is a readable overview of the landscape.]

---

## Your notes

<!-- Add comments here, or inline above any row in the table -->
_Nothing yet._

2. Tell the user

• Confirm the file was saved
• Name the top 3 picks with one sentence each in the conversation
• Say: "Full table and synthesis in ./docs/skill-discovery.md — add notes directly there, then reply build it to hand off to skill-creator."

On confirmation

When the user replies "build it" (or equivalent):

1. Re-read ./docs/skill-discovery.md to pick up any notes they added
2. Build a skill-creator prompt from:
   • The original description
   • Patterns and approaches from the synthesis worth borrowing
   • Specific techniques to adopt, with source URLs
   • Any user notes from the document

Handoff prompt structure:

Build a skill for: {description}

Patterns worth borrowing from existing skills:
- {pattern} — from [{skill-name}]({url})

User notes: {anything from the discovery document}

3. Invoke /skill-creator with that prompt.

What not to do

• Don't skip skills.sh — install counts are the strongest signal of real-world adoption.
• Don't run bunx skills add — this skill only discovers; the user decides what to install.
• Don't fabricate URLs — if a URL can't be confirmed, mark it (unverified).
• Don't make the synthesis a disguised bullet list or a rehash of the table — it should surface patterns that only emerge when you read the skills as a set.