search-skill

name	search-skill
description	Discover and evaluate Claude Code skills from curated sources. Use when searching for existing skills, finding skill inspiration, or before creating a new skill. Triggers include "search skill", "find skill", "skill for X", "what skills exist for", "discover skills".
allowed-tools	["WebSearch","WebFetch","Bash","Read","Glob","Grep"]
model	sonnet

Search Skill

Discover existing skills from curated sources, evaluate quality with toothbrush filter, output comparison table. Philosophy: inspect, adapt, own — never import blindly 🪥

Workflow

1. Parse query from $ARGUMENTS (e.g., "pdf converter", "git workflow")
2. Search tiered sources (parallel where possible):

Tier	Source	Method
1	Anthropic official repos	WebSearch: "$QUERY claude code skill site:github.com/anthropics"
1	VoltAgent curated	WebSearch: "$QUERY site:github.com/VoltAgent/awesome-agent-skills"
2	GitHub community	WebSearch: "$QUERY claude code SKILL.md site:github.com"
2	Awesome lists	WebSearch: "$QUERY awesome-claude-skills OR awesome-claude-code"
3	SkillHub rated	WebFetch: skillhub.club search
3	Broad web	WebSearch: "$QUERY claude code skill 2026"

3. Fetch SKILL.md for top candidates via WebFetch on raw GitHub URLs
4. Apply toothbrush filter — see reference.md for rubric details:
   • Token efficiency: <500 ideal, hard fail >2000
   • Clear triggers: keywords + file types required
   • Single responsibility: one capability only
   • Anthropic patterns + progressive disclosure (soft)
5. Output comparison report:

🔍 Found N skills for "$QUERY":

| Skill | Source | ~Tokens | Quality | Verdict |
|-------|--------|---------|---------|---------|
| name  | source | ~N      | ⭐⭐⭐⭐  | 🟢/🟡/🔴 Study/Adapt/Ignore |

💡 Recommended approach:
- [What to study from best matches]
- [What patterns to adapt]
- [What to ignore and why]

📎 Key patterns found:
- [Reusable patterns across results]

Rules

• Output verdicts (study/adapt/ignore), NOT install commands
• Skip tiers 2-3 if tier 1 has strong matches
• Fetch actual SKILL.md source when possible — evaluate code not marketing
• See reference.md for full source registry and quality rubric weights