Scope
Date window: 2026-05-18 00:00 to 20:33, Asia/Shanghai.
Rules used: DOCUMENT_REVIEW_LEARNING_LOOP.md and team-document-review-learning-loop/SKILL.md.
Recovery archive refreshed at /Users/jack/code/099-github/notebooklm-mcp/output/codex-thread-archives/snapshots/2026-05-18/203310.
Material was sufficient. This is not a no-material review.
Main Work Clusters
| Cluster | Document type | Review stance |
|---|---|---|
| Perplexity Agent Skills method intake | LLM wiki, skill quality gate, review page | Good candidate; external method is C-level and local adaptation remains D-level until reused. |
| Agent skill design quality gate | Project skill and compiled method | Strong structure; lightweight validation substituted because the full validator lacked yaml. |
| Agent Control Room upgrade | Registry, specialist runbooks, task bus, templates | Useful architecture upgrade; runtime automation is not proven yet. |
| Task index updates | TSV task evidence | Good traceability; keep 13-column validation. |
| Touched historical evidence | Manual, video, API and test docs | No positive learning inferred from mtime alone. |
Work Style By Role
| Role | Observed style | Learning boundary |
|---|---|---|
| Jack / reviewer | Converts teaching and article intake into project rules with review evidence. | Preserve C/D labels until local proof exists. |
| Skill author | Moves toward eval-first, route-description-first, gotcha-driven skill design. | Do not claim all existing skills are audited. |
| Orchestrator | Adds named agents, queues, templates, and handoff contracts. | Manual pilot required before automation. |
| Specialist agents | Separate code review, PM scheduling, WeDrive knowledge, product strategy, and delivery assets. | No specialist writes or approvals without the existing gates. |
Accepted Pattern Candidates
- Skill quality should start with routing and eval examples.
- External method adoption should carry evidence levels and local validation status.
- Skill bodies should keep judgment, gotchas, and stop rules rather than generic command lists.
- Agent control room should map skills to named roles and queues.
- File-based task bus is a low-risk bridge before runtime automation.
- Result templates should force validation, risk notes, and next actions.
Anti-Patterns And Quality Gates
External method as local law
Keep Perplexity as C-level and local adaptation as D-level until repeated local use validates it.
Premature automation
Do not claim L4 maturity before a manual task-bus pilot succeeds and is archived.
Casual description changes
Add positive, negative, adjacent-domain, and forbidden-load examples before changing skill descriptions.
Mtime-only learning
Only promote files with explicit new content, validation, review decision, or user feedback.
Next Actions
- Run one manual task-bus pilot before adding runtime automation.
- Create lightweight eval examples for
agent-skill-design-quality. - Add a registry validator for agent docs, queues, and referenced skill paths.
- Keep daily review separating accepted candidates, anti-patterns, and pending questions.