skill-loop

Self-improving skills for AI coding tools.

Skills are static prompt files. Codebases are not. skill-loop closes the feedback loop so skills improve automatically when they degrade.

SKILL --> RUN --> OBSERVE --> INSPECT --> FIX
  ^                                       |
  +---------------------------------------+

1. Observe -- Automatically logs every skill invocation: what ran, whether it succeeded, and how the user reacted
2. Inspect -- Detects failure patterns, staleness (dead file references), routing errors, and usage trends
3. Amend -- Proposes targeted SKILL.md patches grounded in evidence from past runs
4. Evaluate -- Tests amendments against recent failure cases on a git branch before any human sees a PR
5. Update/Rollback -- Merges improvements via PR; rolls back if post-merge monitoring shows regression

| Package | Description | |---------|-------------| | @stylusnexus/skill-loop | Core engine (parser, registry, telemetry, inspector, amender) | | @stylusnexus/skill-loop-cli | CLI (npx skill-loop <command>) + MCP server (serve) | | @stylusnexus/skill-loop-mcp | MCP server (works with any MCP-compatible AI tool) | | @stylusnexus/skill-loop-claude | Claude Code adapter (PreToolUse/PostToolUse hooks) | | @stylusnexus/skill-loop-codex | OpenAI Codex adapter | | @stylusnexus/skill-loop-copilot | GitHub Copilot adapter |

Add one block to your MCP config. That's the only setup.

Claude Code (.mcp.json in your project root):

{
  "mcpServers": {
    "skill-loop": {
      "command": "npx",
      "args": ["@stylusnexus/skill-loop-cli", "serve"]
    }
  }
}

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "skill-loop": {
      "command": "npx",
      "args": ["@stylusnexus/skill-loop-cli", "serve"]
    }
  }
}

Then ask your AI tool: **"Initial