Agent Self-Improvement Log

Problem

This skill addresses a common failure mode in AI-assisted engineering: agents encounter errors, user corrections, stale knowledge, tool quirks, and missing capabilities, but those insights often remain trapped in one conversation and are re-encountered later. It turns transient session experience into searchable, trackable project notes by logging them to structured Markdown files.

How It Works

• Initialise a .learnings/ directory with ERRORS.md, LEARNINGS.md, and FEATURE_REQUESTS.md, without overwriting existing files.
• Capture events at the moment they occur: command failures, user corrections, API or tool failures, knowledge gaps, and feature requests, using IDs like TYPE-YYYYMMDD-XXX.
• Track entries with Status, Priority, Area, and See Also fields so related issues can be linked instead of duplicated.
• Detect recurring patterns by searching existing entries, then update Recurrence-Count, Last-Seen, or create a new entry with a stable Pattern-Key.
• Promote durable rules into project memory files such as CLAUDE.md, AGENTS.md, and .github/copilot-instructions.md; OpenClaw workspaces can also use SOUL.md and TOOLS.md.

Boundaries and Notes

It fits engineering workspaces where agent collaboration is iterative and the team wants to keep the loop of “incident — fix — rule” in the repository. It is not a substitute for tests, CI, or formal postmortems, and it should not log secrets, tokens, environment variables, or full sensitive configuration unless explicitly requested. Prefer short summaries, redacted excerpts, and file paths. Cross-session sharing should only be used in trusted environments and should send concise summaries rather than raw transcripts.