Skip to content

AutoHarness

Claude Code plugin that learns reusable skills from your sessions and prunes unused ones.

Stars
10.6k
Forks
645
Licence
MIT
Language
Python
Last update
7 days ago
Type
Dev tool

What it is

AutoHarness is a plugin for Claude Code, Anthropic's coding assistant. While you work, it turns useful episodes from your sessions into reusable skill files under .claude/skills/, merges similar skills, updates them as they are used, and removes ones that stop being used. It only manages skills it generated itself and leaves your own or installed skills alone.

  • Let skills accumulate from a debugging or coding session without writing them by hand
  • Use /learn to save a lesson from the session you are in right now
  • Review the grouped skill index that is added at the start of each session
  • Clean up the self-authored skill library by uninstalling and deleting the state directory

Great for

  • Developers who use Claude Code heavily and repeat the same kinds of tasks
  • Teams that want project-level skills to build up without manual upkeep
  • People who want to see and control what gets learned, since skills are plain files with a ledger marker

Not ideal for

  • Users who do not use Claude Code, since the plugin only runs inside it
  • Users on Python older than 3.11 or without python3 on PATH, because hooks silently stay off
  • People who want a measured benchmark score for skill quality, since the README says validation comes from use rather than a benchmark
Step-by-step guide: how to use AutoHarnessInstall, first run, examples, tips and FAQs

How to get it

  1. 1

    See the README

    You'll need: Claude Code · Python 3.11+ available as python3 on PATH

    Add the marketplace (type inside Claude Code)

    /plugin marketplace add tigerless-labs/autoharness

    Install the plugin (type inside Claude Code)

    /plugin install autoharness@autoharness
  2. 2

    Run it

    • Install the plugin with the two commands above inside Claude Code
    • Run /reload-plugins or restart Claude Code
    • Work normally; skills are learned in the background after enough tool calls
    • Optionally run /learn to distill the current session immediately
    • Check .claude/skills/ in your project for self-authored skills
  3. 3

    Get the most out of it

    • Tune AUTOHARNESS_REFLECT_EVERY_N to control how often learning runs (default 50 tool calls)
    • Set AUTOHARNESS_DISABLE_GLOBAL=1 for project-only learning
    • Use AUTOHARNESS_INDEX_SUSPENDED=1 if you want to stop the index from using context
    • Enable auto-update for the marketplace so future releases arrive when the version is bumped

Downloads

Strengths

  • Installs as a Claude Code plugin with zero third-party Python dependencies, according to the README
  • Only touches skills it wrote itself, so your own skills are not modified
  • Tunable cadence and scope through AUTOHARNESS_* environment variables

Trade-offs

  • Requires Python 3.11+ as python3 on PATH; the README warns that older interpreters silently disable the hooks
  • The README lists Linux and macOS as the platforms; Windows is not mentioned
  • Effectiveness is described as validated by use, not a benchmark, so quality claims are hard to verify independently

How it compares

ProjectStarsTypeLevelUpdatedBest for
autoharness (this)10.6kDev toolbeginner7 days agoDevelopers who use Claude Code heavily and repeat the same kinds of tasks
ponytail182kAI model / agentbeginner8 days agoDevelopers who use Claude Code or similar AI agents and find the generated code too long
marketingskills57.9kLearning resourcebeginner6 days agoTechnical marketers who already use AI coding agents
agentic-awesome-skills49kDev toolintermediate4 days agoDevelopers using Codex or Claude who want agent-chosen instructions for a project
agents42.1kDev toolbeginner4 days agoDevelopers who already use one of the listed AI coding tools and want ready-made specialist agents
planning-with-files27.8kDev toolbeginner6 days agoDevelopers running long multi-step tasks in Claude Code or Codex CLI

Stats synced from GitHub. The assessment above was drafted with AI from the project's README on 2026-10-29 and may miss recent changes — check the README before relying on a command. MIT licence: free to use, modify and distribute, including commercially, as long as you keep the copyright notice.