Knowledge sharing — version evidence, not tricks
AI operating knowledge should come from evidence: accepted changes, failed runs, review findings, incidents, and measured experiments. Store reusable instructions, skills, tool configurations, fixtures, and decisions in version control with owners and review dates. Conversation channels and demos help discovery, but the durable source must be testable and replaceable when tools change.
Q20 · Organization enablement Max-score evidence: versioned patterns with owners and evidence, updated from reviews and incidents and exercised in recurring team sessions.
Separate durable layers
Section titled “Separate durable layers”- Repository truth: commands, architecture, protected resources, workflow gates, and ownership.
- Reusable capability: tested skills, narrow tool configurations, hooks, and fixtures.
- Decision record: what was tried, baseline, evidence, risks, result, and retirement reason.
- Learning stream: demos, office hours, and discussions that link back to durable artifacts.
- Capture a real signal. Start from a repeated question, failed run, escaped defect, review pattern, onboarding block, or experiment result.
- Choose one canonical artifact. Update repository guidance, a skill, test fixture, runbook, or decision record. Avoid copying the same instruction across tool-specific files.
- Attach evidence. Include a reproducer, accepted example, negative case, source link, or incident reference.
- Review and distribute. Name an owner, expiry or review trigger, and the environments in which the artifact was tested.
- Exercise and retire. Use onboarding and recurring sessions to test discovery; remove superseded patterns and redirects.
Prompts for knowledge maintenance
Section titled “Prompts for knowledge maintenance”Turn this failed run or review finding into the smallest durable artifact. Choose repository instruction, skill, fixture, runbook, or decision record and explain why. Avoid duplicating existing guidance.Audit this pattern for current vendor facts, hidden prerequisites, excessive authority, missing negative cases, absent owner, and unclear retirement trigger.Test whether a fresh agent session and a new engineer can discover and apply this artifact. Record wrong turns, missing links, conflicts, and the evidence produced.Acceptance evidence
Section titled “Acceptance evidence”- Every artifact has an owner, provenance, last verification, and review trigger.
- Current vendor commands and capabilities link to primary sources and are re-checked before material updates.
- Skills and hooks include positive, negative, boundary, and failure fixtures.
- Search and onboarding tests confirm discoverability without private Slack context.
- Superseded instructions are removed or redirect to one canonical location.
Failure pattern: the internal tips archive
Section titled “Failure pattern: the internal tips archive”A large repository of prompts and copied configurations becomes a source of drift when nobody knows which items still work. Prefer fewer owned patterns, test them in fresh sessions, and archive items whose evidence or purpose no longer holds.
Continue with the feedback loop
Section titled “Continue with the feedback loop”Use Maintain to turn incidents into new intent, and Developer onboarding to exercise the knowledge path.