skill-discovery

Skill Discovery

Search across all Claude Code skill sources, compare to what's installed, and install the best match — or identify gaps where a custom skill is needed.

Quick Start

# Unified search across all sources (skills.sh + GitHub + SkillsMP)
uv run mde-py research skill-discover <query>

# JSON output for programmatic use
uv run mde-py research skill-discover <query> --json

# Via mise task
mise run mde:research:skill-discover -- <query>

The Discovery Pipeline

1. INVENTORY — Check what's already installed locally
2. SEARCH — Query working sources (skills.sh, skillkit, GitHub)
3. COMPARE — Rank results by adoption and relevance
4. DECIDE — Install best match, or identify gap for custom skill
5. INSTALL — skills add + symlink + agent wiring
6. VERIFY — Confirm skill file exists and agent references it

Step 1: Check Local Inventory First

Before searching externally, check what's already installed. This avoids duplicates and may reveal the skill already exists.

# Search installed project + user skills by keyword
find .claude/skills .agents/skills -name "SKILL.md" 2>/dev/null \
  | xargs grep -li "<query>" 2>/dev/null

# Count total installed skills
skillfish list 2>&1 | tail -3

# Search installed plugins
claude plugins list 2>&1 | grep -i "<query>"

Step 2: Search External Sources

Run all agent-accessible tools in parallel, then guide the user for interactive/web tools.

Agent-Accessible Search Tools (run these via Bash)

# 1. skills.sh — BEST source. 22K+ skills with install-count rankings.
#    Fastest, most reliable, has adoption metrics as quality signal.
skills find "<query>" 2>&1 | head -30

# 2. skillkit — 15K+ skills, cross-platform (13+ AI tools).
#    Slower (downloads 19MB on first run), but finds results skills.sh misses.
skillkit search "<query>" 2>&1 | head -30

# 3. GitHub code search — deepest, finds niche/personal skills.
#    Searches ALL public SKILL.md files on GitHub.
gh search code "<query>" --filename "SKILL.md" --limit 15 2>&1

User-Interactive Search Tools (user runs these in their terminal)

These tools have interactive TUIs that require a real terminal. Ask the user to run them via the ! prefix (which runs in their shell session):

# 4. skills-installer — 203+ results for popular queries, interactive TUI selector.
#    Searches all public GitHub skills via agentskills.io index.
! skills-installer search "<query>"

# 5. skill.fish browser — 31-33K skills, visual search.
#    User opens https://skill.fish in browser and searches there.

Tell the user: "I can search skills.sh, skillkit, and GitHub automatically, but skills-installer and skill.fish need your terminal. Run ! skills-installer search '<query>' to search the agentskills.io index (203+ results for terraform, 33 for chezmoi)."

API-Key-Required Sources (setup needed first)

# 6. SkillsMP API — 546K+ skills, largest catalog.
#    Requires API key: sign up at https://skillsmp.com/docs/api
#    Set SKILLSMP_API_KEY env var, then use skillsmp-mcp-lite MCP server.
curl -s -H "Authorization: Bearer $SKILLSMP_API_KEY" \
  "https://skillsmp.com/api/v1/skills/search?q=<query>" 2>&1 | head -50

Skillfish (installer only, search broken upstream)

skillfish search returns empty results — the backend API at mcpmarket.com returns 403. But skillfish add, skillfish list, skillfish install (from skillfish.json manifest), and skillfish bundle all work. Use it for installing/managing, not searching.

# List all installed skills across all agents
skillfish list 2>&1 | tail -5

# Install a specific skill from a known repo
skillfish add <owner/repo/skill-name> 2>&1

GitHub Collection Search (curated quality)

For high-quality skills from known collections, search their repos directly:

for repo in wshobson/agents VoltAgent/awesome-claude-code-subagents \
  affaan-m/everything-claude-code ComposioHQ/awesome-claude-skills \
  anthropics/skills BehiSecc/awesome-claude-skills \
  travisvn/awesome-claude-skills; do
  echo "=== $repo ==="
  gh api "repos/$repo/git/trees/main" --jq '.tree[].path' 2>/dev/null \
    | grep -i "<query>" | head -5
done

Step 3: Compare and Rank

Build a comparison table from all search results:

Skill	Source	Installs/Stars	Already Installed?	Author
name	skills.sh / GitHub	count	YES/NO	ecosystem fit?

Ranking criteria (in priority order):

1. Relevance — Does it actually solve the use case? Read the SKILL.md content.
2. Adoption — Install count from skills.sh is the best quality signal
3. Ecosystem fit — Same author as other installed skills? (terrylica has both chezmoi + mise)
4. Recency — When was it last updated?
5. Not a duplicate — Check it isn't already installed under a different name

Step 4: Decide

A. Strong match found → Install (Step 5)

B. Partial match → Install best available + note gaps for a custom companion skill

C. No match → Report: "No existing skill covers this. Closest are [X, Y]. Recommend creating a custom skill covering [gaps]."

Step 5: Install

# Install from skills.sh
skills add <owner/repo> --skill <name> --yes 2>&1

# Verify it landed in .agents/skills/
ls .agents/skills/<name>/SKILL.md 2>/dev/null

# Symlink to .claude/skills/ if not auto-linked
if [ ! -e ".claude/skills/<name>" ]; then
  ln -s "../../.agents/skills/<name>" ".claude/skills/<name>"
fi
echo "Symlink: $(ls -la .claude/skills/<name>)"

Wire to Agent (if target agent specified)

# Read current skills field
grep "^skills:" .claude/agents/<agent-name>.md

# Add the new skill to the YAML list (edit the frontmatter)
# Example: skills: [existing-skill] → skills: [existing-skill, new-skill]

Step 6: Verify

# 1. Skill file exists
head -5 .claude/skills/<name>/SKILL.md

# 2. Agent references it (if wired)
grep "^skills:" .claude/agents/<agent-name>.md

# 3. Skill resolves through symlink
test -f ".claude/skills/<name>/SKILL.md" && echo "RESOLVES" || echo "BROKEN SYMLINK"

High-Adoption Skill Authors

Track these authors — they produce consistently popular, quality skills:

Author	Ecosystem	Top Skills (installs)
terrylica/cc-skills	DevOps + dotfiles (20 plugins)	mise-tasks (114), chezmoi-workflows (85)
samhvw8/dotfiles	mise specialist	mise-expert (98)
bobmatnyc/claude-mpm-skills	Package management	homebrew-formula-maintenance (84)
hashicorp/agent-skills	Terraform (official)	terraform-style-guide (2K), terraform-test (1.2K)
github/awesome-copilot	Azure/cloud	terraform-azurerm (7.3K)

Web Marketplaces (manual search only)

These have large catalogs but no working CLI. Use a browser or MCP when available:

Marketplace	URL	Catalog Size
SkillsMP	https://skillsmp.com	546K+ (API key required)
LobeHub	https://lobehub.com/skills	229K+
Skill.Fish	https://skill.fish	31-33K
awesomeclaude.ai	https://awesomeclaude.ai/awesome-claude-skills	Curated directory
ClawHub	https://clawhub.ai	NEW, vector search

For the full source registry with URLs, sizes, and search commands, see references/sources.md.

Project-Specific Policies

• Library-first: Always search before building custom skills
• Mise-first: If a skill needs a CLI tool, install via mise not npm -g
• Chezmoi pipeline: Tool installs flow through chezmoi templates
• Source catalog: Log every URL discovered to docs/research/source-catalog.md
• Provenance: Write findings to docs/research/trail/findings/finding-<slug>.yaml