Self-Improvement Learning Log

Problem

Coding agents often make mistakes, receive corrections, miss requested capabilities, or hit tooling gotchas, then lose that context between sessions. test002984 addresses how to turn one-off debugging, corrections, and feature signals into searchable, trackable project memory instead of leaving them in chat history.

How It Works

It splits experience into three Markdown logs:
- .learnings/ERRORS.md: command failures, exceptions, API timeouts, and external tool integration errors.
- .learnings/LEARNINGS.md: user corrections, knowledge gaps, best practices, and simplify-and-harden patterns.
- .learnings/FEATURE_REQUESTS.md: user-requested missing capabilities.

Key steps include:
1. Initialize the .learnings/ directory without overwriting existing files.
2. Generate stable IDs such as LRN-YYYYMMDD-XXX, ERR-..., and FEAT-....
3. Search for duplicates first, then link entries with See Also and raise priority when needed.
4. Distill broadly applicable learnings into CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md.
5. Track lifecycle with statuses like pending, in_progress, resolved, wont_fix, and promoted.

Boundaries

It is best for projects with long-term maintenance value: agent conventions, tooling gotchas, workflows, and team preferences. Avoid writing secrets, tokens, environment variables, or full transcripts directly into the log files; prefer short, redacted summaries. It is less useful for one-off tasks, highly private environments, or workspaces without stable state. Cross-session sharing should be enabled only in trusted environments and only when explicitly requested.