AutoHarness
AutoHarness turns your real coding sessions into reusable skills. It improves them when your workflow changes, merges overlapping skills, and archives the ones that stop helping.
No separate training loop. No benchmark to maintain. No background daemon.
A skill layer that improves itself
Claude Code can use skills, but those skills usually have to be written and maintained by hand.
AutoHarness learns from the work you are already doing. Over time, it builds a focused skill layer around your actual projects, tools, and decisions.
| Learns from real work | Turns sessions you are already having into reusable skills. There is no separate data collection or replay step. |
|---|---|
| Keeps related skills together | Checks existing skills before creating a new one. If the scenario already exists, it updates or merges the skill instead of adding a duplicate. |
| Keeps skills that get used | Tracks whether Claude Code follows a skill in later requests. Skills survive based on usage, without a benchmark, oracle, or separate evaluation run. |
| Leaves your skills alone | Only manages skills created by AutoHarness. Anything you wrote or installed stays untouched. |
| Keeps a paper trail | Records why each skill was created or changed, along with the supporting evidence. You can review that history or use it to build a benchmark later. |
How it works
A learning pipeline runs beside Claude Code and stays off its recall path — learned skills are plain native skills, recalled by the host's own name-and-description mechanism as if a human had written them.
| Component | Role |
|---|---|
| Capture | Grabs each turn — user input, agent output, tool I/O — and redacts at egress. |
| Reflect | At an episode boundary, reads the session window and the existing skill index, then decides add / merge / patch / delete. Proposes only; no write tools. |
| Promote | The only writer. Lints each proposal for safety and structure, then lands it atomically into the live skill directory. |
| Lifecycle | Ranks skills by usage rate. New skills sit in probation until they've had a fair sample of requests; a skill never used across probation is archived. Archives, never deletes — moving a folder back revives it. |
| Ledger | Per-skill append-only record: why each skill was born or changed, with a real, redacted slice of the session that taught it. |
Get started
Requires python3 on your PATH. Type these in the Claude Code input box:
/plugin marketplace add tigerless-labs/autoharness/plugin install autoharness@autoharnessThen run /reload-plugins (or restart Claude Code).
Zero config. It now watches your sessions and lands learned skills into .claude/skills/ in the background. Cadence and lifecycle thresholds are tunable through AUTOHARNESS_* environment variables.
More in the works — built in the open, shipped fast.
Explore Tigerless Labs