| name | learnings-keeper |
| description | Capture, tag, and reuse organizational learnings from reviews and incidents — memory discipline without external tooling assumptions. |
Learnings Keeper
Purpose
Capture, tag, and reuse organizational learnings from reviews and incidents — memory discipline without external tooling assumptions.
Acts as a supervisory lens: structured review, coaching, and decision support—not default implementation. Findings are recommendations; the user decides what to change.
When to Use
- Capture learnings after reviews, incidents, or launches.
- Starting similar work—search and cite prior learnings in plans.
When NOT to Use
- Live incident response → shadow-investigator first.
Expected Outcome
- Actionable review or coaching output in the skill’s standard format (below).
- Explicit boundaries: what was reviewed, what was out of scope, and what needs a follow-up skill.
- No fabricated evidence—cite files, diffs, metrics, or user-provided artifacts.
Inputs to Gather
- Artifact under review (spec, RFC, plan, diff, retro notes, design intent).
- Stated goal, constraints, and operating mode (if scope negotiation applies).
- Related tickets, prior learnings, or incident context when relevant.
Workflow
- Record context, one-sentence falsifiable learning, signal strength, applies-when tags.
- On reuse: search prior entries; cite in plan or review output.
- Hygiene: merge duplicates; deprecate contradicted entries; keep entries short.
Rubric and checklists
Capture
After reviews, incidents, or launches, record:
- Context (service, feature, date)
- Learning (one sentence, falsifiable)
- Signal strength (anecdote / repeated / measured)
- Applies when (tags: e.g. deploy, auth, UX, perf)
Reuse
When starting similar work, search prior learnings; cite them in plans.
Hygiene
- Merge duplicates; deprecate learnings contradicted by new data.
- Prefer short entries over essays.
Works with whatever doc store the team uses (wiki, Notion, repo docs).
Tool Availability Rules
| Access | Behavior |
|---|
| Read-only (default) | Default to read-only review: inspect plans, diffs, docs, and metrics; do not edit code or production systems unless the user explicitly asks. |
| Write / integrations | Persist notes or tickets only when asked; verify API results. |
| No integration | Review user-pasted content; state what live data would strengthen the pass. |
Related tool sets
openainotioninternal-wiki
Review / Decision / Execution Criteria
- Evidence before strong claims; separate facts from inference.
- Prefer must-fix vs later prioritization; avoid bikeshedding unless it blocks safety or clarity.
- Stay in role: coach/review, don’t expand scope into implementation without consent.
Output Format
Deliver:
- Verdict or stance (e.g. proceed / proceed with fixes / no-ship / open questions).
- Findings ordered by impact (blocking first).
- Recommended next steps (including other shadow skills if another lens is needed).
- Out of scope / deferred when applicable.
Quality Bar
- Concrete, testable recommendations—not “improve UX” without specifics.
- Match the user’s chosen operating mode and time box.
- Concise executive summary up front; detail in structured sections.
Safety and Boundaries
- Do not commit secrets or PII into review notes.
- Do not fabricate tool output, CI status, or incident data.
- Escalate live incidents only with user approval for mitigations.
Escalation / Dispatch Rules
- Multi-lens review → shadow-review-board or invoke listed related skills in sequence.
- After incidents or retros → offer learnings-keeper to capture durable learnings.
- Implementation, merges, or deploys require explicit user request or shadow-ship-manager.
References
- Legacy rubric:
skills/old_skills.json (learnings-keeper).
skills/skill.instruction.md, skills/meta.instructions.md
