Content
81%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A well-structured, highly operational workflow skill: clear sequencing, fail-closed validation at every decision point, and concrete tool-level guidance. Main weaknesses are minor redundancy in the provenance rules, missing executable grounding commands for the provider adapter, and all detail living inline in one long file.
Suggestions
In Step 1, include the actual provider-adapter command for the GitHub case (e.g., gh pr view owner/repo#NNN --json author,state,headRefOmit) so the fact fields are directly executable.
Deduplicate the reverse-provenance rules — state the 'related ≠ origin' rule once and reference it from the Fail-closed list and Common Mistakes table.
Consider moving the custody decision table and evidence-grading detail into a reference file to keep SKILL.md as a leaner overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and rule-driven with no padding explaining concepts Claude already knows, but there is minor repetition — '外部对象触发内部 thread 只能算 related' appears in both the 证据分级 table and the Fail-closed 规则, and the proposal-source-as-origin prohibition is stated twice — that could be trimmed to reach the lean anchor. | 4 / 5 |
Actionability | Concrete tool calls (gh pr comment, gh auth status, cat_cafe_search_evidence, github_record_external_review_verdict), typed field tables, a copy-paste JSON metadata example, and scenario-to-custody decision tables make the guidance mostly executable; the gap is that Step 1 grounding says '例如 GitHub 场景用 gh' without ever giving the actual command (e.g., gh pr view --json) that populates the fact fields. | 4 / 5 |
Workflow Clarity | Steps 1-6 are explicitly sequenced, each with fail-closed validation branches ('停住，输出 unknown', '升级到 operator') that serve as error-recovery feedback loops, plus entry conditions and a Common Mistakes checklist; merge operations are explicitly gated (maintainer-only, no self-review/self-merge), so the destructive-operation cap does not apply. | 5 / 5 |
Progressive Disclosure | No bundle files exist or are referenced, so the skill is a single self-contained file with clear section headers, tables, and a sibling-skill relationship map for navigation; however at ~190 lines, detail like the custody decision tables and provenance evidence-grading is inlined in one file where it could be split, matching 'good structure; minor organization gaps' rather than the fully split anchor. | 4 / 5 |
Total | 17 / 20 Passed |