Self-Improvement Learning Log

Problem Being Solved

In AI-agent-assisted development, useful context is easily lost: a failed pnpm install, a correction like “regenerate the OpenAPI client first,” or a CI permission gotcha can recur if it is not captured. self-improvement-v2 focuses on turning errors, corrections, feature requests, and knowledge gaps into searchable, trackable, promotable project memory rather than fixing code directly.

How It Works

The skill expects a .learnings/ directory at the project or workspace root, usually with Markdown logs such as:

• LEARNINGS.md: user corrections, knowledge gaps, best practices
• ERRORS.md: command failures, exceptions, external tool issues
• FEATURE_REQUESTS.md: capabilities the user wants added

When triggers appear—non-zero exits, stack traces, “that is outdated,” “can you also support this,” or a better approach discovered—the agent should append a structured entry. Entries use stable IDs such as LRN-20250115-001, ERR-20250115-A3F, or FEAT-20250115-002, plus Status, priority, area tags, related files, and See Also links.

The promotion step matters most: if a lesson is not one-off but applies across files, tasks, or team conventions, distill it into a concise rule and write it to project memory such as CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md. In OpenClaw workspaces, it may also go to SOUL.md or TOOLS.md. Recurring patterns can be tracked with Pattern-Key and Recurrence-Count before becoming system-prompt guidance.

Boundaries

This skill fits long-lived agent workspaces where tools read project memory, including Claude Code, Codex, Copilot, and OpenClaw. It does not automatically implement fixes, and its value depends on timely, specific, actionable logging. Avoid recording secrets, tokens, full source files, or long raw output; prefer redacted summaries. Hooks are optional and may inject reminders or inspect command output, so review their setup before enabling.