stbenjam/skillsaw

Agent instructions behave like code, but most teams still review them like prose. skillsaw gives them a linter. It validates structure across every major AI coding ecosytem, guards against many supply-chain attacks, and applies content and context rules backed by research and frontier lab guidance.

It understands Agent Skills, Agent Plugins v1, Claude Code plugins, OpenAI Codex plugins and marketplaces, CLAUDE.md, AGENTS.md, GEMINI.md, QWEN.md, Cursor, Copilot, Cline, Devin, Kiro, OpenCode, Muse Code, Grok Build, Google Antigravity, hooks, agent configuration, MCP Registry server.json publisher metadata, Vercel skills CLI lockfiles, and eval formats. Safe structural fixes can be applied automatically; everything else comes with precise, agent-friendly guidance.

Get started | Browse the rules | Read the documentation

See it work

Watch an AI agent grade, fix, and configure a repository from scratch.

Try it

Paste this into your coding agent to onboard skillsaw now:

Read and follow the instructions at
https://raw.githubusercontent.com/stbenjam/skillsaw/refs/heads/main/skills/skillsaw-onboard/SKILL.md
to onboard this repo to skillsaw.

Or install the onboarding skill for regular use (recommended). Run this from the repository you want to lint:

npx skills add stbenjam/skillsaw

Select skillsaw-onboard and your coding agent, then ask the agent to use that skill to onboard the repo. See Onboard with AI for plugin installation alternatives.

Or run it yourself. No installation is required with uvx:

uvx skillsaw tree      # See what skillsaw detects
uvx skillsaw           # Lint the current repository
uvx skillsaw fix       # Apply safe, deterministic fixes
uvx skillsaw baseline  # Accept existing findings and fail only on new ones

For lint --fail-on info, use baseline --include-info to accept existing INFO findings too. A configured fail-on: info includes them automatically.

What it catches

• Multi-ecosystem structure & compatibility: schema, frontmatter, and manifest validation for Agent Skills (SKILL.md), Claude Code, OpenAI Codex (project config, plugins & marketplaces), Grok Build (project config, plugins & marketplaces), Google Antigravity (configuration in any customization root — .agents/, .agent/, _agents/, _agent/ — its rules/ and agents/ prose, plugins, hooks, MCP servers and registries), Agent Plugins v1 (plugin.json, mcp.json), GitHub Copilot & VS Code custom agents (.github/agents/), OpenCode configuration, APM packages, MCP server maps, and MCP Registry metadata.
• Content quality & token economy: research-backed rules detecting instruction drift across duplicate files, lost-in-the-middle attention dead zones, cognitive overload, section length violations, weak language, contradictions, and repetitive inline tool-call examples.
• Discovery & repository integrity: unreferenced bundled files, broken internal file references, inconsistent terminology, missing stop conditions, and stale baselines.
• Security & supply chain:
  • Dangerous lifecycle hooks: blocks arbitrary remote code execution, download-and-execute (curl | sh, wget | bash), and script obfuscation (eval) in hooks.json and settings.
  • Prohibited & unvetted MCP servers: enforces strict MCP allowlists across root, plugin, and custom agent configurations.
  • Prompt injection & stealth payloads: detects invisible Unicode (ASCII smuggling, zero-width tags, bidi overrides), high-entropy encoded payloads (base64/hex), and hidden instructions in comments and code fences.
  • Environment & context security: flags dangerous environment overrides (LD_PRELOAD, NODE_OPTIONS, PYTHONPATH), unallowlisted dynamic context injection, and embedded credentials. Deterministic autofixes: safe, instant automated fixes for invalid frontmatter, broken headings, missing manifests, unclosed code fences, and schema keys via skillsaw fix. skillsaw detects repository types automatically and lints multiple formats in the same project. See supported repository types and the complete rule reference for details.

Built for real workflows

skillsaw works locally, in CI, and inside coding-agent workflows. It provides line-level findings, explanations for every rule, deterministic autofixes, baselines for gradual adoption, GitHub and GitLab integration, and text, JSON, SARIF, HTML, and Code Climate output. Rules are configurable, and projects can add local rules or install rule plugins. Typo’d or wrong-typed rule options in .skillsaw.yaml are reported with did-you-mean suggestions instead of being silently ignored.

Codex portable plugins can place OpenAI metadata and hooks in plugin.json under extensions.com.openai. Skillsaw follows that overlay’s precedence over .codex-plugin/plugin.json, validates its declared resources, and checks portable skills/ and mcp.json through the Agent Plugins rules.

Measure the result

Every run produces a letter grade based on weighted violation density. The same data can be rendered as a self-contained report card for a README or project dashboard.

skillsaw’s own report card, generated with skillsaw badge --large.

Learn how to generate a grade badge and report card for your project.

Contributing

Contributions are welcome. See CONTRIBUTING.md for the project guidelines and DEVELOPMENT.md for the local setup.

Questions and bug reports belong in GitHub Issues. For a shareable diagnostic bundle, run skillsaw feedback in the affected repository and review the ZIP before attaching it to an issue. Files selected with --include or --config are copied verbatim, including BOMs and line endings. Selections default to 4 MiB per file and 16 MiB total across distinct ZIP members; use positive byte values with --max-file-bytes and --max-total-bytes to override these limits. An oversized selection stops before diagnostic lint and creates no bundle. These limits cover selected file bytes only, not diagnostic output or total process memory. skillsaw is licensed under the Apache License 2.0.

Thank you to our contributors

skillsaw is better because people contribute code, bug reports, and ideas. Thank you!