---

name: skills-discovery description: Find and evaluate community skills, plugins, and marketplaces. Use when searching for existing skills, evaluating safety, or when "find skill", "discover plugin", "community skills", or "marketplace" are mentioned. metadata: version: "1.0.0" related-skills: - skills-dev - claude-plugins allowed-tools: Read WebFetch WebSearch Bash(gh:*)

Skills Discovery

Find community skills and plugins, evaluate quality and safety before use.

<when_to_use>

• Searching for existing skills before building from scratch
• Evaluating community plugins for safety and quality
• Finding inspiration for skill design patterns
• Auditing plugins before installation

NOT for: creating new skills (use skills-dev), validating your own skills (use skills-dev)

</when_to_use>

Discovery Workflow

1. Search — Find candidates via GitHub topics or code search
2. Filter — Apply quality heuristics to shortlist
3. Audit — Security review before installation
4. Adapt — Customize or extract patterns for your use

GitHub Discovery

Topic Pages

High-signal discovery starting points:

Topic	Content	URL
claude-code-plugin	Plugins	https://github.com/topics/claude-code-plugin
claude-code-plugin-marketplace	Marketplaces	https://github.com/topics/claude-code-plugin-marketplace
claude-code-skills	Skill packs	https://github.com/topics/claude-code-skills
claude-code-skill	Individual skills	https://github.com/topics/claude-code-skill

Code Search Patterns

Precise searches for specific artifacts:

# Find SKILL.md files in .claude/skills paths
filename:SKILL.md path:.claude/skills

# Find marketplace configurations
".claude-plugin/marketplace.json"

# Find plugin manifests
".claude-plugin/plugin.json"

# Find hook configurations
"PreToolUse" AND hooks

# Find skills with specific features
filename:SKILL.md "context: fork"
filename:SKILL.md "allowed-tools"
filename:SKILL.md "disable-model-invocation"

Recency Filters

Focus on actively maintained projects (adjust dates as needed):

# Updated in last 90 days (calculate: date -v-90d +%Y-%m-%d)
pushed:>YYYY-MM-DD

# Updated since plugins era (Oct 2025+)
pushed:>2025-10-01

Official Sources

Source	Trust Level	Notes
anthropics/claude-plugins-official	High	Curated, reviewed
agentskills/agentskills	High	Spec + reference skills
platform.claude.com docs	High	Official patterns
Community topics	Medium	Popularity ≠ quality
"Awesome" lists	Low-Medium	Curated but not audited

Quality Heuristics

Real Usage Signals

Signal	Good	Suspicious
Updates	Recent commits, active issues	Stale for 6+ months
Stars	Steady growth	Sudden spike (star farming)
Issues/PRs	Open and being addressed	Many open, no responses
Install docs	Uses official commands	"curl | bash" installs
Dependencies	Minimal, explained	Many unexplained deps

Content Quality

Check	Good	Bad
Description	Clear WHAT + WHEN + TRIGGERS	Vague "helps with files"
allowed-tools	Minimal, justified	Full tool access
disable-model-invocation	Used for side effects	Missing for deploy/commit
Scripts	Documented, minimal	Obfuscated, complex
Hooks	Obvious purpose	Hidden network calls

Marketplaces

Good Sign	Red Flag
Version pinning	Floating branches
Listed sources visible	Opaque references
Clear update policy	Silent auto-updates
Curated with criteria	"Everything goes"

Security Audit

Threat Model

Installing skills/plugins = running code. Treat with same care as npm packages.

Surface	Risk	Mitigation
Skills with Bash	Command execution	Review allowed-tools
Hooks	Lifecycle interception	Review hook scripts
MCP servers	External connections	Review endpoints
Preprocessing !	Shell before thinking	Review commands

Audit Checklist

Before installing, review:

For Skills:

• Read SKILL.md frontmatter (allowed-tools, disable-model-invocation)
• Check for scripts/ directory — review any scripts
• Search for ! ` preprocessing commands
• Verify no secrets/credentials in files

For Plugins:

• Read .claude-plugin/plugin.json
• Check for hooks/ — review hook scripts
• Check for .mcp.json — review MCP endpoints
• Review all referenced skill SKILL.md files

For Hooks:

• Understand exit code semantics (0=allow, 2=block)
• Check for network calls in hook scripts
• Verify no data exfiltration patterns

Sandboxing

When running untrusted skills:

1. Restrict tools — Start with minimal allowed-tools, expand as needed
2. Isolate context — Use context: fork to limit blast radius
3. Block side effects — Add disable-model-invocation: true initially
4. Monitor first run — Watch tool calls on first execution

Safe First Run

# Test skill in restricted mode:
---
name: untrusted-skill-test
allowed-tools: Read, Grep, Glob  # read-only first
context: fork                     # isolated
disable-model-invocation: true   # explicit only
---

Expand permissions only after reviewing behavior.

Use Case Catalog

Common skill categories with examples (for inspiration, not endorsement):

Workflow Automation

Pattern	What It Does	Key Features
PR workflows	Summarize, review, update PRs	Preprocessing with gh
Issue pipelines	Triage → implement → ship	Artifact-based state
Release automation	Preflight → deploy → verify	Side-effect gates

Code Quality

Pattern	What It Does	Key Features
Spec gates	Verify scope before coding	Fork for clean analysis
Adversarial review	Security-focused code review	Threat model in artifacts
Refactor loops	Safe read-only explore first	Tool restrictions

Domain Skills

Pattern	What It Does	Key Features
Framework-specific	Rails, React, etc conventions	Nested skill discovery
DB-aware	Schema injection for queries	Preprocessing with psql
Platform integrations	Jira, Linear, GitHub	MCP or API wrappers

Safety & Guardrails

Pattern	What It Does	Key Features
Safety nets	Block irreversible operations	PreToolUse hooks
Hardstops	Require human acknowledgment	Exit code blocking
Test gates	Enforce tests before commit	Hook enforcement

Context Management

Pattern	What It Does	Key Features
Memory plugins	Persist across sessions	MCP-backed storage
Context ledgers	Rolling state in files	Hook-driven updates
Constraint files	Minimal "always load" context	Shared conventions

Extraction Patterns

When you find a useful skill, extract patterns rather than copying wholesale:

1. Identify the pattern — What makes it work?
2. Adapt to your context — Match your conventions
3. Minimize scope — Take only what you need
4. Document provenance — Note where the pattern came from

<rules>

ALWAYS:

• Verify recency (prefer active projects)
• Review security surfaces before install
• Start with restricted permissions
• Document what you installed and why

NEVER:

• Blindly install from unknown sources
• Trust stars as quality signal
• Run obfuscated scripts
• Skip hook script review

</rules> <references>

• discovery-patterns.md — Detailed GitHub search patterns
• security-checklist.md — Full audit checklist
• use-cases.md — Extended use case catalog

</references>