163 Email

Problem

When coding agents repeatedly hit the same issue, or users point out a mistake or missing capability, that feedback often stays trapped in chat. This skill turns such signals into searchable, actionable project records.

How it works

It works around three files: ERRORS.md for command failures, exceptions, and external tool breaks; LEARNINGS.md for corrections, knowledge gaps, and best practices; and FEATURE_REQUESTS.md for missing capabilities. Each entry carries Status, priority, See Also, and a stable Pattern-Key, making dedupe and merge easier with grep. If an entry proves useful across tasks, it can be distilled into a short rule and promoted to CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; in OpenClaw workspaces, it may also enter SOUL.md or TOOLS.md. It also supports ingesting simplify-and-harden candidates and promoting a pattern into prevention rules when it appears in at least two tasks within 30 days and has Recurrence-Count >= 3.

Boundaries

It is not an auto-fix engine; it converts agent experience into Markdown memory. It fits long-lived repositories, multi-agent workflows, and teams that want to codify engineering conventions. For one-off questions, the maintenance cost may outweigh the benefit. Error entries should include reproducible output, and learning entries should include concrete fixes, otherwise promoted notes risk becoming vague guidance.