skillgen-ai 0.3.2

Analyze any codebase and auto-generate AI agent skill files for Claude Code, Cursor, and other AI tools

skillgen

Teach AI assistants your codebase conventions — automatically.

One command. Your conventions. Every AI tool on your team.

---

The Problem

Every time an AI assistant touches your code, it guesses at your conventions. Wrong quote style. Wrong test patterns. Wrong import order. You fix it, it forgets, you fix it again.

skillgen reads your actual codebase and generates convention files that Claude Code, Cursor, and other AI tools understand natively. No hand-writing rules. No guessing. Every line is backed by evidence from your code.

---

Install in Claude Code

/plugin marketplace add mmoselhy/skillgen
/plugin install skillgen@skillgen-marketplace

That's it. Now you have two slash commands in every project:

/skillgen:skillgen              Analyze codebase, generate .claude/skills/*.md
/skillgen:skillgen enrich       Find community skills for your stack

No Python or pip required. The plugin runs entirely inside Claude Code — Claude reads your code and generates skill files directly. For even better results, also install the CLI (pip install skillgen-ai) to enable hybrid mode (CLI stats + Claude semantics).

---

Install as CLI

pip install skillgen-ai
skillgen .

Generates .claude/skills/, .cursor/rules/, and AGENTS.md. Deterministic output, works offline, runs in CI.

Optional extras:

pip install skillgen-ai[tree-sitter]  # AST-based analysis (more accurate)
pip install skillgen-ai[llm]          # LLM-enhanced output (requires API key)

Requires Python 3.11+.

---

What It Produces

skillgen scans your code and generates 8 convention categories:

Category	What It Captures
Naming	snake_case vs camelCase, class suffixes, file naming patterns
Error Handling	Exception hierarchy, try/except patterns, error propagation style
Testing	Framework, fixtures, assertion style, mocking patterns
Imports	Absolute vs relative, grouping order, key dependencies
Documentation	Docstring format, coverage %, type annotation usage
Architecture	Directory layout, layering, where new code goes
Code Style	Line length, quotes, trailing commas, formatter config
Logging	Library, logger init pattern, structured vs unstructured

Every bullet includes prevalence stats ("82% of files use snake_case") and real examples from your code. No generic advice.

Demo

$ skillgen . --dry-run --format claude

 Scanning files and detecting languages...  0:00
 Analyzing patterns...                      0:00
 Synthesizing conventions...                0:00
 Generating skills...                       0:00

--- naming-conventions (dry run, not written) ---
# Naming Conventions

## Function Naming
- **82% Functions use snake_case** (14/17 files)
  - Examples: `analyze_project`, `detect_project`, `validate_input`

## Class Naming
- **41% Classes/types use PascalCase** (7/17 files)
  - Examples: `Language`, `PatternCategory`, `OutputFormat`

--- code-style (dry run, not written) ---
# Code Style

## Formatters & Linters
- **ruff** -- line-length: 100, select: E, F, W, I, N, UP, B, SIM, RUF
- **mypy** -- python_version: 3.11, strict: true

Done! 8 files would be generated.

CLI vs Plugin

	CLI	/skillgen Plugin
Install	pip install skillgen-ai	/plugin install skillgen@skillgen-marketplace
How it works	Regex + tree-sitter AST	Claude reads your code
Output style	Statistical ("82% snake_case")	Semantic ("verb_noun pattern: get_user")
Formats	.claude/ + .cursor/ + AGENTS.md	.claude/ only
Speed	< 1 second	5-15 seconds
Deterministic	Yes	No (richer, but varies)
Context cost	Zero	~1,700 lines (hybrid) / ~10K (standalone)
Best for	Teams, CI, multi-format	Individual devs, quick setup

Hybrid mode: When both are installed, /skillgen uses CLI stats as the backbone and Claude adds semantic enrichment — best of both worlds.

Output Formats

Claude Code .claude/skills/*.md — one file per category:

<!-- Generated by skillgen v0.3.0. Do not edit manually. -->
# Code Style
## Quote Style
- **88% Prefers double quotes** (15/17 files)
## Formatters & Linters
- **ruff** -- line-length: 100, select: E, F, W, I, N, UP, B, SIM, RUF
- **mypy** -- python_version: 3.11, strict: true

Cursor .cursor/rules/*.mdc — same conventions, Cursor-native frontmatter.

AGENTS.md — single Markdown file at repo root. Uses <!-- skillgen:start/end --> markers so your handwritten sections are preserved.

Community Skills

After generating local conventions, pull in community-curated skills for your stack:

$ skillgen . --enrich

Found 3 community skills for Python + FastAPI:

  #  Skill                   Description
  1  FastAPI Conventions      Router patterns, DI, HTTPException
  2  Pytest Best Practices    Fixtures, parametrize, conftest
  3  SQLAlchemy Patterns      Session management, eager loading

  Skipped (already covered locally): naming, code-style, imports

$ skillgen . --enrich --apply --pick 1,2

207 community skills across Python, TypeScript, JavaScript, Go, Rust, and Java — sourced from Anthropic, GitHub Copilot, and awesome-cursorrules with trust tiers (official, community, contributed).

Supported Languages

Language	Frameworks Auto-Detected
Python	Django, FastAPI, Flask
TypeScript	Next.js, React, Angular, Vue
JavaScript	Express, React, Vue
Go	Gin, Cobra
Rust	Actix, Tokio
Java	Spring
C++	—

skillgen reads your tool configs (ruff, prettier, eslint, mypy, golangci-lint) and embeds the actual settings in generated skills.

How It Works

path ──> DETECT ──> ANALYZE ──> SYNTHESIZE ──> GENERATE ──> WRITE
          |          |            |               |            |
       languages   patterns    conventions     skills       files
       frameworks  evidence    prevalence      confidence   .claude/
                   per-file    config values   meters       .cursor/

1. Detect — scan file tree, read manifests, identify languages + frameworks
2. Analyze — sample up to 50 files/language, extract patterns across 8 categories
3. Synthesize — deduplicate, compute prevalence, parse config files
4. Generate — render evidence-only skills with confidence scoring
5. Write — atomic writes, orphan cleanup, --dry-run support

CLI Reference

skillgen <path> [flags]

Output:
  --format, -f <claude|cursor|all>   Target format (default: all)
  --diff                             Show what AI learns vs blank slate
  --dry-run                          Preview without writing files
  --json                             Export analysis as JSON
  --verbose, -v                      Detailed analysis output
  --quiet, -q                        Errors only

Analysis:
  --no-tree-sitter                   Disable AST parsing, use regex
  --llm                              Enhance with Claude/GPT-4o
  --llm-provider <anthropic|openai>  Choose LLM provider

Community:
  --enrich                           Find matching community skills
  --enrich --apply                   Download and install them
  --enrich --apply --pick 1,3        Cherry-pick by number
  --trust <official|community|all>   Filter by trust tier
  --no-cache                         Force re-fetch of skill index

Contributing

See docs/CONTRIBUTING.md for development setup, testing, and PR guidelines.

License

MIT