back1ply/skill-curator

Skill Curator

A comprehensive skill auditing system for AI coding assistants. Automatically discover, analyze, and optimize skills across any platform.

Supported Platforms

| Platform | Skill Locations | Config Format | |----------|-----------------|---------------| | Gemini CLI | .gemini/skills/, ~/.gemini/skills/ | SKILL.md + JSON | | Claude Code | .claude/, plugins, CLAUDE.md | Markdown + JSON | | Cursor | .cursor/rules/, .cursorrules | Markdown | | GitHub Copilot | .github/copilot-instructions.md | Markdown | | Windsurf | .windsurfrules, .codeium/ | Markdown | | Custom | User-specified paths | Various |

---

Marketplace Preference Hierarchy (Claude Code)

When analyzing Claude Code plugins, always prefer official sources over third-party alternatives.

| Priority | Type | Source | Trust Level | |----------|------|--------|-------------| | 1. Official | Made by Anthropic | @claude-plugins-official | Highest | | 2. Endorsed | Third-party in official marketplace | Listed in marketplace.json | High | | 3. External | Third-party marketplaces | Other @marketplace sources | Medium | | 4. Custom | User's own plugins | Local/personal repos | User-managed |

Preference Rules:

1. If official alternative exists, recommend replacing external with official
2. If endorsed alternative exists, recommend replacing external with endorsed
3. If duplicate across marketplaces, keep the higher-priority source
4. If no official/endorsed equivalent, keep external (note it in report)

Fetch the authoritative list from: https://raw.githubusercontent.com/anthropics/claude-plugins-official/main/.claude-plugin/marketplace.json

For full marketplace listings, identification methods, and overlap resolution rules, see references/marketplace-reference.md.

---

Phase 0: Quick Start Mode

Offer analysis depth at the start:

Quick or Full Analysis?

[Q] Quick: Auto-detect everything, use smart defaults (2 min)
[F] Full: Complete profile questionnaire + deep analysis (5 min)

Quick Mode: Auto-Profile from Workspace

Detect automatically instead of asking questions:

| Detection | Files to Check | Inference | |-----------|----------------|-----------| | Tech Stack | package.json, requirements.txt, *.csproj, go.mod, Cargo.toml | Primary language/framework | | Workflow | .github/, .gitlab-ci.yml, CODEOWNERS | Solo vs team indicators | | Priority | Existing skill categories | Security skills = security priority | | Platform | .agent/, .claude/, .cursor/ | AI assistant in use |

Quick Mode Weights: Relevance 35%, Uniqueness 25%, Quality 20%, Efficiency 15%, Usage 5%

---

Phase 1: User Profile Discovery

Skip in Quick mode.

Gather these four data points to personalize recommendations:

1. Tech Stack — Primary languages/frameworks (Python/FastAPI, TypeScript/React, Go, Rust, etc.)
2. Workflow Type — Solo developer, small team (2-5), or large team/enterprise
3. AI Usage Priorities — Rank: Speed, Quality, Security
4. Primary Use Cases — Writing code, code review, debugging, testing, documentation, planning

---

Phase 2: Automated Discovery

Step 1: Detect Platform

Check for platform indicators in this order:

1. .gemini/skills/    -> Gemini CLI
2. .claude/           -> Claude Code
3. .cursor/           -> Cursor
4. .github/copilot-*  -> GitHub Copilot
5. .windsurfrules     -> Windsurf
6. Ask user           -> Custom/Unknown

Step 2: Scan Skill Directories

Execute discovery based on detected platform:

Gemini CLI: Scan .gemini/skills/**/SKILL.md and ~/.gemini/skills/**/SKILL.md, check MCP server configurations, parse each SKILL.md frontmatter.

Claude Code: Scan .claude/settings.json for plugins, check CLAUDE.md files, parse plugin manifests, fetch official marketplace.json for source classification, classify each plugin as Official/Endorsed/External/Custom.

Cursor: Scan .cursor/rules/*.md and .cursorrules in project root.

Copilot/Windsurf: Scan instruction files in standard locations, parse markdown content for capability definitions.

Step 3: Build Skill Inventory

| # | Skill Name | Source | Description | Est. Tokens |
|---|------------|--------|-------------|-------------|
| 1 | skill-name | path   | description | ~500        |

---

Analysis Phases

After discovery, run these analysis phases in order:

1. Deep Skill Analysis — Read full skill content, extract trigger keywords, anti-patterns, dependencies, token estimates, and complexity scores
2. Semantic Similarity — Extract capability signatures ([Actions] x [Domains]), detect overlaps (exact duplicate, superset, partial, complementary), group by semantic category
3. Conflict Detection — Find contradictions between skills (opposing rules, style clashes, trigger conflicts, version conflicts)
4. Weighted Scoring — Calculate composite score per skill: Relevance (30%), Uniqueness (25%), Quality (20%), Efficiency (15%), Usage (10%)
5. Dependency Mapping — Map requires/enhances/conflicts/MCP relationships, identify orphans, critical nodes, broken dependencies

For detailed methodology, scoring tables, and detection methods, see references/analysis-methodology.md.

For the standard report output template, see references/report-template.md.

---

Operational Rules

• Offer Quick vs Full mode at the start of every audit
• Use automated discovery — never ask users to paste configuration
• Calculate semantic overlap, not just name matching
• Check for conflicts between skill rules
• Provide weighted scores with justification
• Give exact commands, not vague instructions
• Show dependency relationships for critical skills
• Never recommend removing a skill without understanding the user's workflow
• Never assume platform — detect or ask
• Ask about tech stack before stack-specific recommendations
• Prioritize removing high-token duplicates over low-token unique skills
• Consider that some overlap is intentional (tiered tools for different contexts)
• Prefer official/endorsed plugins over external alternatives (Claude Code)
• Fetch the official marketplace.json to identify endorsed plugins
• Recommend replacing external plugins when official/endorsed equivalent exists

---

Example Workflow

Scenario: Gemini user, Python/FastAPI stack, solo developer, quality-focused

Discovery finds: 3 code review skills (similar), 2 Python skills (both relevant), 1 JavaScript testing skill (wrong stack), 1 planning skill (unique)

Result: Remove 3 skills (2 duplicate code reviewers + wrong-stack JS skill), save ~800 tokens, 43% reduction. Keep best code reviewer, both Python skills (complementary), and unique planning skill.

---

Reference Files

For detailed methodology, templates, and extended references:

• references/analysis-methodology.md — Deep analysis, semantic similarity, conflict detection, weighted scoring, and dependency mapping methodology
• references/report-template.md — Standard audit report template with backup commands, consolidation suggestions, and verification steps
• references/marketplace-reference.md — Complete official/endorsed plugin listings, identification methods, replacement lookup, and overlap resolution rules
• references/portfolio-templates.md — Recommended skill sets by developer role (Python Data Scientist, TypeScript Full-Stack, Power BI/DAX, DevOps, Solo Generalist)