Agent Self-Improvement Log

Problem

Coding agents often remember a mistake, correction, or tool quirk within one session, then repeat the same issue in the next. If lessons stay only in chat history, future sessions cannot read them quickly, and durable project rules are hard to form.

How It Works

This skill writes reusable lessons into local Markdown files instead of relying on model memory. Core flow:
- Log immediately: command failures go to .learnings/ERRORS.md, user corrections or knowledge gaps go to .learnings/LEARNINGS.md, and missing capabilities go to .learnings/FEATURE_REQUESTS.md.
- Structure entries: use IDs like LRN-YYYYMMDD-XXX, ERR-..., or FEAT-..., with Status, Priority, Area, and See Also fields.
- Promote to project memory: when a lesson applies across files or persists over time, distill it into a short rule in CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md.
- Detect recurrence: search similar entries with grep -r "keyword" .learnings/, increment Recurrence-Count, and promote repeated patterns into system prompt guidance when thresholds are met.

Boundaries

It fits agent workflows that maintain persistent project files and review regularly; one-off Q&A benefits less. Logging adds small context overhead, so keep entries reproducible, actionable, and prioritized rather than pasting long incident notes into memory files.