| name | skill-curator |
| description | Maintain skill ecosystem quality, discoverability, and security. Auto-runs after skill creation/modification to validate conventions, check security, detect similarity overlaps, generate searchable catalog, suggest MCP tool integrations, track usage analytics, and optimize skills for token efficiency. Use when creating/reviewing skills, generating skill catalog, checking for duplicates, maintaining ecosystem health, or reducing token footprint of a skill.
|
| trigger | - "review this skill"
- "is there a skill that does X?"
- "generate skill catalog"
- "check skill ecosystem health"
- "find duplicate skills"
- "suggest MCP tools for this skill"
- "optimize skill tokens"
- "compress this skill"
- "reduce skill token count"
- "how many tokens does this skill use?"
- Auto-trigger after creating/modifying any skill
|
| examples | [{"input":"Review the newly created flake-hunter skill","output":"Runs security review (via skill-security-review), validates frontmatter (name, description, trigger), checks conventions compliance, suggests MCP tools (agent-builder-skill-dev for trace analysis), outputs quality report with remediation steps"},{"input":"Is there a skill that does test migration?","output":"Searches skill catalog for 'test migration', finds cypress-to-scout-migrator (95% match), shows description, usage examples, and trigger phrases"},{"input":"Generate skill catalog","output":"Scans 34 skills in ~/.agents/skills/, categorizes by domain (Kibana-specific: 8, Development workflow: 12, Quality: 6, Documentation: 4, Security: 4), generates ~/.agents/SKILL_CATALOG.md with searchable index"},{"input":"Optimize skill tokens for the flake-hunter skill","output":"Counts ~340 tokens (260 words * 1.3), identifies 4 opportunities (redundant headings, verbose examples, duplicate triggers, filler prose), shows before/after diff with ~38% savings, applies changes after approval"}] |
@skill-curator
Purpose: Comprehensive skill ecosystem quality assurance, discoverability, and security. Maintains skill quality gates, generates searchable catalogs, detects overlapping skills, suggests MCP tool integrations, and tracks usage analytics.
Context: You have 34+ skills in ~/.agents/skills/ and the ecosystem is growing. Need automated curation to prevent: (1) security vulnerabilities, (2) duplicate/overlapping skills, (3) convention violations, (4) missed MCP tool integration opportunities, (5) unused/stale skills.
Integration with MCP: Uses agent-builder-skill-dev MCP server for similarity detection and ecosystem audits.
When to Use
Automatic activation:
- After creating a new skill (post-create hook from
create-skill)
- After modifying a skill's SKILL.md
- When user asks "is there a skill for X?"
- When user mentions skill quality, ecosystem health, or catalog
Manual invocation:
/skill-curator [mode]
Modes:
review <skill-path> — Full quality review of a single skill
catalog — Generate searchable skill catalog
similarity <skill-path> — Check for overlapping skills
ecosystem-audit — Full ecosystem health check
suggest-tools <skill-path> — Suggest relevant MCP tools
usage-analytics — Analyze skill usage from conversation history
auto-update <api-name> — Find skills using deprecated API
token-optimize <skill-path> — Analyze and reduce token footprint of a skill
Core Workflow
See references/mode-workflows.md for detailed step-by-step workflows.
Mode 1: Post-Creation Review
Goal: Validate a newly created/modified skill meets all quality gates.
Steps:
- Run security review (via
skill-security-review)
- Check convention compliance (frontmatter, structure, examples)
- Suggest MCP tool integrations (based on domain keywords)
- Detect similarity with existing skills (via
agent-builder-skill-dev)
- Generate quality report with health score
Output: ~/.agents/skills/<skill-name>/QUALITY_REPORT.md
Quality gates:
- Security: 0 CRITICAL, ≤2 HIGH
- Conventions: ≤2 violations
- Similarity: <85% overlap (70-85% = WARN)
- SKILL.md: ≤500 lines
Mode 2: Catalog Generation
Goal: Generate searchable index of all skills by category.
Steps:
- Scan all skills in
~/.agents/skills/
- Parse frontmatter, extract domain keywords
- Categorize by domain (Kibana, Dev workflow, Quality, Docs, Security, etc.)
- Generate
~/.agents/SKILL_CATALOG.md with descriptions, triggers, MCP tools, usage stats
Categories: Kibana-specific, Development workflow, Quality assurance, Documentation, Refactoring, Observability, Security
See references/catalog-template.md for full template.
Mode 3: Similarity Detection
Goal: Find overlapping skills using semantic similarity.
Steps:
- Use
agent-builder-skill-dev.analyze_skill_similarity
- Get similarity scores with existing skills
- Classify risk: >85% CRITICAL, 70-85% HIGH, 50-70% MEDIUM, <50% LOW
- Suggest: merge, clarify distinction, or keep as-is
Output: Similarity report with recommendations
Mode 4: Ecosystem Audit
Goal: Full pairwise similarity audit to find all overlapping skills.
Steps:
- Use
agent-builder-skill-dev.audit_skill_ecosystem
- Identify highest-risk overlap pairs
- Check security issues across all skills
- Analyze usage (most used, unused)
- Calculate ecosystem health score
Output: ~/.agents/ECOSYSTEM_HEALTH_REPORT.md
Mode 5: MCP Tool Suggestions
Goal: Suggest relevant MCP tools based on skill domain.
Steps:
- Analyze skill description for domain keywords
- Map to MCP tools via heuristics (see
references/mcp-tool-mappings.md)
- Generate integration snippets for SKILL.md
Example: "Agent Builder + eval" → suggest agent-builder-skill-dev, langsmith
Mode 6: Usage Analytics
Goal: Analyze skill usage from conversation history, identify unused skills.
Steps:
- Use
cursor-chat-browser to search for skill invocations
- Count usage in last 30/60/90 days
- Identify deprecation candidates (0 usage in 90d or 180d)
Output: ~/.agents/SKILL_USAGE_REPORT.md
Mode 7: Auto-Update
Goal: Find skills using deprecated API, suggest updates.
Steps:
- Grep all SKILL.md files for deprecated API
- Generate update plan with suggested replacements
- Show affected skills with line numbers
Example: yarn test:jest --config → yarn test:jest <path>
Mode 8: Token Optimization
Goal: Reduce the token footprint of a SKILL.md by 30-70% without losing meaning or trigger coverage.
Trigger phrases: "optimize skill tokens" | "compress this skill" | "reduce skill token count" | "how many tokens does this skill use?"
Steps:
- Count tokens — estimate:
word_count * 1.3. Report as "~N tokens (M words * 1.3)".
- Audit for opportunities — scan the skill body for each technique below and list findings.
- Show before/after diff — present a side-by-side summary of proposed changes with estimated per-change savings.
- Get approval — ask "Apply these optimizations? (yes / yes, all / skip N)" before writing.
- Apply — rewrite the file with approved changes.
- Verify triggers — confirm every trigger phrase still appears (or is covered by a condensed variant) in the updated file.
- Report savings — "Reduced from ~N to ~M tokens (~X% savings)."
Optimization techniques:
| Technique | When to apply | Typical savings |
|---|
| Remove redundant headings | Section title restates what the content already makes obvious | 5-15 tokens/heading |
| Compress examples | Multi-line code block can be expressed inline | 10-30 tokens/example |
| Deduplicate trigger phrases | Two triggers convey the same intent ("fix tests" + "fix the tests") | 5-10 tokens/duplicate |
| Replace prose with tables | 3+ parallel items described in sentences | 20-40% of that block |
| Remove filler phrases | "This skill will help you to", "In order to", "Please note that" | 3-8 tokens/phrase |
| Consolidate step lists | Steps 2-4 are a single logical action split arbitrarily | 10-20 tokens/merge |
| Move data to frontmatter | Trigger examples duplicated in both frontmatter and body | Dedupe body entries |
| Prune rare examples | Keep the 2-3 highest-frequency triggers; cut niche edge-case examples | 20-50 tokens/example removed |
Constraints (never violate):
- Every unique intent covered by the original trigger list must remain triggerable after optimization.
- Do not remove required frontmatter fields (
name, description, trigger, examples).
- Do not merge steps that have distinct preconditions or failure modes.
- If a quality gate (
≤500 lines) is already met, still report token count — caller may want the reduction for other reasons.
Output format:
Token Optimization Report — <skill-name>
=========================================
Current: ~680 tokens (523 words * 1.3)
Target: ~400 tokens (estimated)
Opportunities found:
[1] Remove filler preamble in Purpose section → -18 tokens
[2] Deduplicate 3 trigger phrases (review/check) → -12 tokens
[3] Replace 4-paragraph prose with table (Mode 3) → -55 tokens
[4] Compress verbose example #2 to inline → -28 tokens
[5] Prune 2 rarely-triggered edge-case examples → -42 tokens
Projected savings: ~155 tokens (~23%)
Apply? (yes / yes,all / skip 3,5 / no)
Quality Gates (Auto-Enforced)
After every skill review, enforce these gates:
| Gate | Threshold | Action if Failed |
|---|
| Security (CRITICAL) | 0 | ❌ BLOCK: Fix before proceeding |
| Security (HIGH) | ≤2 | ⚠️ WARN: Fix recommended |
| Convention violations | ≤2 | ⚠️ WARN: Fix recommended |
| Similarity (HIGH overlap >85%) | 0 | ❌ BLOCK: Merge or clarify |
| Similarity (MEDIUM overlap 70-85%) | Allowed | ⚠️ WARN: Clarify distinction |
| SKILL.md length | ≤500 lines | ⚠️ WARN: Use progressive disclosure |
| Token footprint | ≤1500 tokens (~1150 words) | ℹ️ INFO: Run token-optimize mode |
| MCP tool suggestions | ≥1 (if applicable) | ℹ️ INFO: Review suggestions |
Integration with Other Skills
After create-skill:
- Auto-run
/skill-curator review <new-skill-path>
- Block commit if CRITICAL security issues found
After modifying SKILL.md:
- Auto-run
/skill-curator review <skill-path>
- Warn if similarity increased (new overlap detected)
After major API changes:
- Run
/skill-curator auto-update <api-name>
- Generate PR with suggested changes
Output Artifacts
- Quality Report:
~/.agents/skills/<skill-name>/QUALITY_REPORT.md
- Skill Catalog:
~/.agents/SKILL_CATALOG.md
- Ecosystem Health:
~/.agents/ECOSYSTEM_HEALTH_REPORT.md
- Usage Analytics:
~/.agents/SKILL_USAGE_REPORT.md
Success Metrics
- Security: 0 CRITICAL issues across all skills
- Conventions: 95%+ compliance
- Discoverability: Catalog updated weekly, <5 min to find relevant skill
- Overlap: <10% of skills have HIGH overlap risk
- Usage: >70% of skills used in last 90 days
- Health Score: >80/100 ecosystem health
Notes
- MCP availability: Uses
agent-builder-skill-dev MCP server for similarity/audit. If not available, falls back to manual keyword-based similarity.
- Conversation history access: Uses
cursor-chat-browser MCP tool for usage analytics. If not available, skip analytics mode.
- Security reviews: Always delegates to
skill-security-review skill for deep security analysis.
- Progressive disclosure: For large catalogs (>50 skills), split by category into separate files.