Autoharness: a self-learning skill layer for Claude Code

Tigerless Labs released autoharness, a plug-in for Claude Code that learns skills directly from your active sessions. Each episode converts into a skill without manual data collection, and similar scenarios merge into one to prevent the accumulation of near-duplicates.

On the CORE-Bench (HAL) benchmark, the same model with autoharness scores 78% compared to 42% without it. A skill remains in the active layer if it is used in subsequent turns and is archived if it stops being called. The plug-in only modifies skills it created itself.

autoharness is written in Python with no external dependencies and runs without a resident daemon. Install it via /plugin marketplace add tigerless-labs/autoharness. It is licensed under MIT.