Hi Product Hunt 👋 We built Auto-Harness around a simple question: what if Claude Code could learn from the work you are already doing, instead of requiring you to manually write and maintain every skill? Auto-Harness runs alongside your normal Claude Code sessions and turns useful patterns from real work into reusable skills. When a similar scenario appears again, it updates or merges the existing skill rather than creating another near-duplicate. Skills that prove useful remain active, while unused ones are eventually archived, keeping the skill layer focused instead of letting it grow indefinitely. A few principles shaped the project: • Learn from real sessions — no separate training or replay workflow • Consolidate instead of accumulate — related experiences improve one skill • Validate through actual use — no benchmark or oracle required in the active loop • Stay out of the way — no resident daemon and no additional recall system • Respect user-owned skills — Auto-Harness only manages the skills it created itself • Keep an evidence trail — every creation and update records why it happened The result is a self-maintaining skill layer that gradually adapts to your projects, workflows, and corrections while remaining transparent: everything is stored locally as ordinary Claude Code skill files. Auto-Harness is open source, MIT licensed, and available as a Claude Code plugin. We would especially appreciate feedback on the learning behavior, lifecycle design, and the kinds of workflows you would want it to capture. Thanks for checking it out — we’d love to hear what you think.

Auto-Harness

Self-Learning Skills for Claude Code

Self-Learning Skills for Claude Code