About this skill

The Problem: Session Learnings Are Ephemeral

Many coding-agent mistakes are not exotic: the wrong package manager, a missed API convention, or a user correction treated as temporary context. The issue is that these signals stay trapped in one transcript and are invisible to later agents, so the same failure can recur. This skill turns immediate feedback into searchable Markdown logs in the project, instead of relying on a one-line reminder in the prompt.

How It Works

Categorized logging: command failures go to .learnings/ERRORS.md; corrections, knowledge gaps, and best practices go to .learnings/LEARNINGS.md; missing capabilities go to .learnings/FEATURE_REQUESTS.md.
Structured entries: each entry has an ID, Status, Priority, Area, and See Also link, making it easier to search with grep and connect recurring issues.
Promotion to project memory: when a learning is reusable across files or sessions, it is distilled into a short rule and written to CLAUDE.md, AGENTS.md, or .github/copilot-instructions.md; OpenClaw workspaces can also use SOUL.md or TOOLS.md.
Recurring-pattern escalation: entries sharing a Pattern-Key accumulate Recurrence-Count values, and only after meeting thresholds such as repeated occurrence, multiple tasks, and a time window are they promoted into more stable prompt guidance.
Optional skill extraction: if a learning is recurring, verified, and not tied to a one-off debugging session, it can later be extracted into a standalone skill.

Boundaries

This skill addresses whether feedback is visible and reusable; it does not automatically patch code or validate the promoted rules. Its value depends on logging immediately, keeping entries specific, and reviewing them regularly. If a team shares .learnings/, it also needs decisions about Git tracking, sensitive error output, and avoiding the promotion of one-off debug notes into long-term conventions.

Screenshots

Use Cases

When Claude Code keeps failing on an API fix, record the command output and related files in `ERRORS.md`.
After the user corrects the package manager, log the `npm` mistake and the `pnpm` rule as a learning entry.
When Git push auth fails across sessions, track recurrence and promote the rule to `AGENTS.md` once qualified.
After debugging a test timeout, decide if the fix is reusable and extract it as a standalone skill candidate.

Best For

Engineers using Claude Code or Codex who want errors and user corrections captured as searchable rules for later sessions
OpenClaw workspace maintainers who need tool gotchas and behavioral constraints written into `AGENTS.md` or `SOUL.md`
Developer leads responsible for AI-agent project conventions who want one-off debugging notes turned into shared memory
SREs repeatedly handling Git, CI/CD, or dependency-install failures who need failure context logged and recurrence tracked