Teaching Claude Code to Remember What You Actually Do

What it does

AutoHarness is a Claude Code plugin that sits beside your normal workflow and distills reusable skills from the turns you were already typing. It captures episodes of real work, reflects on them in background child sessions, and lands the results as plain SKILL.md files in .claude/skills/. A built-in lifecycle manager tracks whether each skill is actually being invoked or read; unused ones graduate to an archive folder, while frequently used ones get patched and merged instead of duplicated.

The interesting bit

The project treats skill maintenance as a garbage-collection problem rather than a training pipeline. There is no daemon, no benchmark suite, and no hot-reload on update—just a lazy lifecycle recomputation at session start that ranks skills by opportunity-relative usage rate and moves the losers out of recall. It also keeps an append-only ledger per skill so you can audit exactly which session fragment justified every create, patch, or archive decision.

Key highlights

• Learns from live sessions, not replay loops or held-out evals: skills are born from the work you were already doing.
• Merges same-scenario skills instead of accreting near-duplicates; the reflector folds related episodes into one symbol.
• Zero third-party dependencies; runs as plain Python hooks and an MCP server.
• Only touches skills it authored—your hand-written or installed skills are left completely alone.
• The README cites a jump from 42% to 78% on CORE-Bench (per the HAL paper) by improving the harness around the same model.

Caveats

• Size caps on captured windows and staged skill bodies are fixed constants, not configurable environment variables.
• Uninstalling the plugin leaves its generated skills and state on disk; manual cleanup of marked folders is required to fully remove traces.
• Maturity thresholds and capacity caps are described as deliberate placeholders pending empirical calibration.

Verdict

Worth a look if you treat Claude Code as a daily driver and want your repetitive rituals to persist without curating JSON by hand. Skip it if you expect a polished, fully calibrated system—this is explicitly a bet that one slice of the harness can maintain itself, rough edges and all.