AutoHarness: let Claude Code turn each session’s lessons into skills
TL;DR
AutoHarness is an MIT-licensed Claude Code plugin that distils reusable approaches from real sessions into skills. What sets it apart is lifecycle management: it updates skills with new lessons, merges same-scenario skills, and prunes ones that go unused.
Every Claude Code session contains experience: how to build this project, recurring mistakes to avoid, naming conventions, the release process. Usually that experience disappears when the session ends, or you have to write it into CLAUDE.md yourself. AutoHarness is a Claude Code plugin (MIT licence) that automatically extracts reusable practices from real sessions, saves them as skills, and manages their whole lifecycle.
The skill lifecycle
| Stage | What happens |
|---|---|
| Create | Automatically once a session has done enough work, or call /learn at any time |
| Update | Skills are refined as they're reused in later turns |
| Merge | Skills for the same situation are merged into one, with a merge history recorded |
| Prune | Skills unused for a long time are removed |
AutoHarness only edits skills it created itself; skills you wrote or installed elsewhere are untouched. That's an important safety boundary.
Why the idea is interesting
- Knowledge no longer lives in one person's head: lessons from one session become usable in the next.
- Less repetition: no need to re-explain how to run tests or the commit conventions every session.
- Self-cleaning: the weakness of accumulating skills is that more means noisier; the merge and prune mechanisms target exactly that.
Installation
/plugin marketplace add tigerless-labs/autoharness
/plugin install autoharness@autoharness
# requires Python 3.11+; then /reload-pluginsRisks to manage
- Broad hooks: the plugin has hooks matching every tool, running Bash and editing files. Read it carefully as you would any software with broad permissions, especially on machines holding company source code.
- Wrong lessons: automatically extracted lessons can be wrong or too specific to one situation. A wrong skill reloaded every session is worse than no skill.
- Information leaks: generated skills may contain internal paths, customer names or sensitive details from the session. Review them before sharing.
- Cost: extracting lessons also uses tokens; watch usage after enabling it.
A safe way to use it
- Try it first on a personal project or non-sensitive repo.
- After a few sessions, open the skills folder and read each generated skill.
- Fix or delete wrong skills; keep the genuinely useful ones.
- For good skills the whole team can use, move them into a shared repo after review instead of leaving them on one person's machine.
How to apply it
- Dev: use
/learnafter a hard debugging session to save the diagnostic process. - Tech Lead: treat generated skills as drafts of process documentation; edit them before standardising for the team.
- PM: this is a way to "grow" process knowledge from real work, but someone needs to own reviewing it so quality doesn't drift.