Self-Improvement Learning Log

Problem

Coding agents often repeat avoidable mistakes across sessions: using the wrong package manager, forgetting generated clients, or tripping over the same API quirk. Conversational context disappears, so the same failure can reappear weeks later. This skill turns errors, corrections, and feature requests into durable text artifacts instead of transient chat history.

How It Works

It maintains .learnings/ at the project or workspace root:
- ERRORS.md: command failures, exceptions, and timeouts
- LEARNINGS.md: corrections, knowledge gaps, and best practices
- FEATURE_REQUESTS.md: requested capabilities

Entries use IDs like LRN-20250115-001, with metadata such as Status, Priority, Area, and See Also. When a lesson is broad or recurring, it can be distilled into concise rules and promoted to CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; in OpenClaw workspaces, guidance may also land in SOUL.md and TOOLS.md.

Boundaries

This is useful for iterated codebases and shared agent conventions. Logs can capture sensitive output, so prefer summaries and redactions over raw command transcripts. Hook reminders are optional and should not be mistaken for automatic repair. If local notes should stay private, keep entries out of version control.