addyosmani/agent-skills: Production-grade engineering skills for AI coding agents.
Production-grade engineering skills for AI coding agents. - addyosmani/agent-skills
AI Agent Plugin
Summary
Autoharness is a self-learning skill layer plugin for Claude Code that distills reusable skills from active sessions, merges same-scenario duplicates, and prunes unused entries. It runs as a background hook with zero third-party dependencies and only modifies skills it generated itself.
Highlights
- Distills skills from real sessions automatically; no separate data-collection or replay loop needed
- Merges same-scenario skills into one instead of stacking near-duplicates, with fold provenance tracking
- Validates skills by adherence in later turns (loads over availability), not held-out benchmark scores
- Claims 42% to 78% improvement on CORE-Bench (HAL) by letting the harness maintain its own skill layer
- Zero third-party dependencies; runs entirely in Python 3.11+ as hooks and an MCP server
auto-generated
Context
Audience
Developers using Claude Code who want their agent to accumulate and maintain reusable skills from real work without manual curation
DomainAI Agent Orchestration
Formatopen-source software project
Accessopen source on GitHub; install via Claude Code plugin system
Source Overview
Autoharness is a Python 3.11+ Claude Code plugin that distills reusable skills from active sessions, merges same-scenario duplicates, and prunes unused entries on configurable cadence.
It runs as a background hook with zero third-party dependencies and modifies only skills it generated itself. Install via the Claude Code plugin system; no daemon or separate config required.
Topics
Related
Claude CodeMCP (Model Context Protocol)CORE-Bench (HAL)swyx's Big Model vs Big HarnessAI agent skill management
