Self-Improvement Log

Problem

AI coding agents often repeat the same mistakes: forgetting the package manager, skipping type checks, misreading environment setup, or losing context after a session ends. Manually maintaining CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md tends to lag behind real work, and useful rules end up buried in chat history.

How It Works

This skill turns runtime experience into traceable Markdown logs. During first use, it initializes .learnings/ at the project or workspace root with LEARNINGS.md, ERRORS.md, and FEATURE_REQUESTS.md. When a command fails, a user corrects the agent, knowledge becomes outdated, or a feature request appears, the skill appends a categorized entry and assigns a stable ID such as LRN-20250115-001 or ERR-20250115-A3F using the TYPE-YYYYMMDD-XXX pattern.

Key steps include:
- Log immediately: prefer short summaries, reproduction steps, and relevant file paths.
- Track status: move entries from pending to in_progress, resolved, wont_fix, or promoted.
- Detect recurrence: search existing logs with grep, add See Also links, raise priority, and decide whether the issue needs docs, automation, or an architecture fix.
- Promote durable rules: write broadly applicable guidance into CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; in OpenClaw workspaces, also consider SOUL.md and TOOLS.md.
- Use optional hooks: trigger reminders or error detection via UserPromptSubmit and PostToolUse hooks.

Boundaries

It fits long-lived coding-agent workflows, not one-off scripts or memoryless tasks. Logs can capture sensitive context, so avoid recording secrets, tokens, full source files, or raw command output by default. Before sharing with a team, decide whether to keep logs local, commit them as shared knowledge, or track templates while ignoring entries.