ccoalm/ccl-skills / skills/skill-extraction-workflow
skill-extraction-workflow
Flagged复盘 / 沉淀 / 总结经验 / 补进技能 / 技能缺陷 / 流程优化 / would other teammates hit this / will we forget next time / teamwide recurrence / extract skill / post-mortem / 深度 review·审计技能仓库 / 对标外部技能包找 gap(benchmark superpowers·gstack 等) / 核查全局安装点(~/.config/opencode 等)旧快照是否遮蔽本仓技能 / 本仓(ccl-skills 等共享技能仓)OpenCode 项目配置·命令治理 → turn reusable lessons into durable CCL skills.
references/review-rubric.md
View on GitHub# Review Rubric — Counterexample-Oriented
Phase 1 material (design source: `docs/skill-extraction-optimization-design.md`). This rubric structures the **review pass** of the dual-track gate so it hunts for counterexamples instead of agreeing. It does not replace `references/dual-track-review-gate.md` — that reference still owns the run-BOTH-tracks requirement, convergence standard, and lens-independence rules. This rubric is the question set the review pass works through; the challenge pass stays adversarial per its own runbook.
The failure this prevents: an LLM reviewer that answers "looks reasonable, no problems" — confident agreement that misses over-firing, under-firing, leakage, and authority overreach. LLM-as-judge has known position / verbosity / self-enhancement biases, so a clean review is one evidence class, not a verdict.
## Rubric — every L1 / L2 review answers all of these
| Dimension | Required question |
|---|---|
| Provenance | Can every rule trace back to evidence? |
| Generalization | Is it generalized from the specific case into a reusable capability, not copied? |
| Boundary | Are applies / does-not-apply written explicitly? |
| Routing | Could it mis-fire or steal another owner's task? |
| Cognitive load | Could a reference, example, or eval carry this instead of the entrance? |
| Privacy / R0 | Any business noun, internal host, private path, person, or source-shaped example? |
| Behavior | Which observable agent behavior does it change? |
| False positive | Which scenarios would over-fire? |
| False negative | Which scenarios would it miss? |
| Regression | Is there a should-trigger / should-not-trigger case? |
| User authority | Does it decide on the user's behalf where it should ask? |
| Rule budget | Could this avoid adding a new hard rule? |
## Reviewer prompt direction — find counterexamples, do not affirm
Do NOT ask "is this change reasonable?". Require the reviewer to produce concrete failing scenarios:
```text
Find the scenarios where this skill change would mis-fire, under-fire, bloat the rule
set, fail sanitization, collide with another owner's boundary, override user sovereignty,
become un-verifiable in behavior, or duplicate/contradict an existing skill.
```
A review that returns no counterexamples must say *why each rubric row is clear*, not just "no issues found". An empty finding list with no per-row reasoning is inconclusive, not a pass.
## The rubric is necessary, not sufficient
- For an L1 change, the rubric review plus the recorded challenge row and behavioral-evidence row close the gate.
- For an L2 change, a clean review is explicitly NOT enough on its own: automatic R0 (`references/r0-leakage-audit.md`), deterministic checks (`check-ccl-skills.sh`, `git diff --check`, the routing analyzer where applicable), and behavioral regression (≥1 should-trigger + ≥1 should-not-trigger) must also pass.
- Treat any LLM reviewer output — supporting or critical — as hypothesis-grade: a "looks correct" does not raise the draft's evidence grade, and a finding stays valid until specifically verified or refuted (see the Core Rules on LLM-consultation evidence).
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |