tigerless-labs/autoharness

What it solves

AutoHarness is a self-learning skill layer for Claude Code that allows the AI agent to automatically acquire, refine, and manage its own capabilities without manual intervention. It prevents the "bloat" of duplicate skills by consolidating similar scenarios into single, updated skills rather than stacking near-duplicates.

How it works

The project implements a learning pipeline that runs alongside the host agent:

• Capture (CAP): Records user inputs, agent outputs, and tool I/O from sessions.
• Reflection (REF): Periodically analyzes session windows and decides whether to add, merge, patch, or delete skills based on the current context and existing skill index.
• Promoter: Validates and stores the proposed skills as plain files in the .claude/skills/ directory, ensuring only self-authored skills are modified.
• Lifecycle Management (MNG): Tracks usage rates of skills. New skills enter a probation period; if they are used, they graduate to a mature pool. If the mature pool exceeds capacity, the least-used skills are archived.
• Ledger (LED): Maintains an append-only record of why each skill was created or changed, linked to redacted session evidence.

Who it’s for

Users of Claude Code who want their agent to become more proficient at specific project-specific workflows or general techniques through experience, without having to manually write skill files.

Highlights

• Self-Maintaining: Automatically prunes unused skills and merges duplicates to keep the skill layer clean.
• Zero-Config: Installs as a plugin with no mandatory configuration required to start learning.
• Non-Destructive: Only touches skills it generated itself; user-authored skills are left untouched.
• Evidence-Based: Keeps a detailed paper trail of the session transcripts that justified every skill modification.