Self-Improvement Learning Log

Problem

Coding agents often repeat the same failures: broken commands, stale APIs, user corrections, and missing capabilities. That knowledge usually stays trapped in a transcript, so the next session can make the same mistake and the team cannot reuse the rule.

How It Works

• Log immediately: record errors, corrections, knowledge gaps, and feature requests in .learnings/ERRORS.md, LEARNINGS.md, or FEATURE_REQUESTS.md, using IDs like ERR-YYYYMMDD-XXX for lookup.
• Track lifecycle: mark entries as pending, in_progress, resolved, or promoted, and connect similar entries with See Also links instead of duplicating them.
• Promote durable rules: distill cross-cutting lessons into short rules in CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; OpenClaw can also target SOUL.md and TOOLS.md.
• Detect recurrence: ingest simplify-and-harden candidates, dedupe with Pattern-Key, count occurrences, tasks, and a 30-day window, then promote repeated patterns into system-prompt feedback.

Fit And Limits

This skill focuses on structured logging and promotion, not automatic bug fixing. It fits Claude Code, Codex, Copilot, and other agent workflows, with OpenClaw as a recommended host. Teams can commit .learnings/ for shared knowledge or gitignore it for local use. Define categories, priorities, and promotion targets first, otherwise logs can become an unreviewed stream.