Agent Self-Improvement Log

Problem

Coding agents often repeat the same failures: broken commands, user corrections, tool integration issues, and outdated assumptions. If feedback stays only in one session, the next run may miss it. This skill turns those signals into local, searchable notes, routing errors, corrections, and feature requests into ERRORS.md, LEARNINGS.md, and FEATURE_REQUESTS.md, and creating a path from session feedback to project memory.

How It Works

The core loop is capture, deduplicate, distill, and promote:
- Capture: when a command exits non-zero, an exception appears, or the user says “actually” / “no”, append an entry with an ID like LRN-YYYYMMDD-XXX.
- Structure: entries carry status, priority, area tags, related files, and reproduction steps, making them easy to search with grep.
- Deduplicate: search .learnings/ for similar entries first, then link them with See Also instead of creating duplicates.
- Promote: durable rules that apply across files can be distilled into CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; OpenClaw workspaces can also map them to SOUL.md or TOOLS.md.
- Close the loop: move entries from pending to resolved after fixes, or to promoted after they become project guidance.

Boundaries

This works best in coding-agent workflows with a stable project root, not in throwaway chats. Logs may contain sensitive command output, so avoid recording secrets, tokens, full source, or unredacted configs. Cross-session sharing should happen only when the user asks and the environment is trusted. Teams that want shared knowledge should explicitly version .learnings/ and keep entries concise.