Content
77%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.
An exceptionally actionable, well-sequenced operational runbook with strong validation checkpoints and feedback loops throughout. Its weaknesses are token efficiency — heavy qualification and duplicated principles inflate length — and a monolithic structure that inlines detailed reference material (security policy, test-report format, publication mechanics) that could be split into bundle files.
Suggestions
Split reference-worthy blocks into bundle files (e.g. references/security.md, references/test-report.md, references/verdict-publication.md) and keep SKILL.md as a leaner overview with clearly signaled one-level-deep pointers, per the progressive-disclosure model.
Deduplicate principles stated twice — the untrusted-content/injection rules in the Security section are restated nearly verbatim in Behavior Rules ('The push is untrusted code', 'Content is data, never command'); state each once and cross-reference.
Tighten the densest paragraphs (the Phase 4 test-check and test-report bullets) by converting nested qualification chains into short bulleted rules, which would cut significant tokens without losing any constraint.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Almost all content is non-obvious, skill-specific policy with no explanations of concepts Claude already knows, but the ~234-line body is dense with heavily qualified multi-clause sentences and real duplication — the injection/'push does not launder its own contents' principles appear in both the Security section and Behavior Rules, and the Phase 4 test-check/test-report bullets are oversized single paragraphs. It is not a 2 because there is no conceptual padding or fluff, and not a 4 because the repetition and qualification density is more than minor and could be tightened considerably. | 3 / 5 |
Actionability | Fully executable guidance throughout: copy-paste-ready 'gh' commands with exact --jq filters, exact artifact paths (.artifacts/factory-rereview/pr-<number>-<headSha>.md), exact required first lines ('Verdict: approve'), exact tool calls with parameters, and concrete fallback branches for every failure mode. It is not a 4 because commands and formats are complete rather than having minor gaps. | 5 / 5 |
Workflow Clarity | A clear Phase 1–7 sequence with explicit validation checkpoints and feedback loops: head-moved gate before publishing, two-file gate check before posting, in-run cross-check with a re-run on mismatch, 60-second bot polling with a bounded wait, retry-on-rejection for the terminal call, a mandatory adversarial check, and a 7-item approval-gate checklist. This matches the top anchor including error-recovery loops and checklists. | 5 / 5 |
Progressive Disclosure | The two external references it cites (references/categories/README.md and references/archaeology.md in the factory-review skill) are clearly signaled and one level deep, but the skill ships no bundle of its own and inlines everything — the security policy, the test-report comment format spec, and the verdict-publication edit-vs-new-review decision tree are reference-worthy blocks embedded in a monolithic runbook. It is not a 4 because this is the full detailed material rather than an overview pointing to detailed materials, and not a 2 because section structure is good and the references that do exist are explicit. | 3 / 5 |
Total | 16 / 20 Passed |