Content
63%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 richly structured, highly actionable read-only PR-reporting skill with a clear workflow, concrete thresholds, and strong guardrails. Its two weaknesses are repetition that inflates token cost, and a progressive-disclosure design whose referenced detail files are missing from the bundle, breaking navigation.
Suggestions
Deduplicate PR-size benchmarks and the AI-inflated-metrics caveat so each rule lives in one canonical place (Critical Decision Rules), with other sections cross-referencing rather than restating.
Ship the referenced reference/*.md and _common/*.md files in the bundle (or inline the most critical ones, e.g. gh-commands and report-templates) so the Reference Map's links resolve.
Add a couple of literal, copy-pasteable gh command examples inline in Core Contract for the most common fetch patterns, rather than deferring all syntax to the (currently absent) gh-commands reference.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Dense and mostly actionable, but redundancy and some over-explanation remain: PR-size benchmarks (200/400/1000 LOC) recur in Core Contract, Trigger Guidance, and Critical Decision Rules, and the AI-inflated-metrics caveat is restated in both Core Contract and Boundaries→Never; Goodhart's Law and the McKinsey controversy are explained at length. Not a 2 because content is substantive rather than padded, not a 4 because the repetition is more than minor. | 3 / 5 |
Actionability | Highly concrete guidance throughout — "per_page=100" with "gh api --paginate", LOC thresholds (200/400/1000), time benchmarks (<6h/<13h/<26h/<48h), exact output filenames (pr-summary-YYYY-MM-DD.md), and subcommand dispatch. Not a 5 because literal executable gh command examples are deferred to reference/gh-commands.md rather than shown inline, leaving a minor gap for the common cases. | 4 / 5 |
Workflow Clarity | Clear SURVEY→COLLECT→ANALYZE→REPORT→VERIFY sequence in a phase table, with a VERIFY validation step, Ask-First gating for >100-PR batches, and a graceful-degradation feedback rule ("never fabricate; label degraded sections"). Not a 5 because validation checkpoints are stated abstractly rather than as explicit pass/fail checks with retry loops. | 4 / 5 |
Progressive Disclosure | Structure and signaling are excellent — a Reference Map, one-level-deep references, well-organized sections — but the referenced reference/*.md and _common/*.md files are absent from the bundle (only scripts/ exists), so navigation leads to non-existent files and the disclosure is not actually realized. Not a 4 because the missing-file gap is more than a minor organization issue. | 3 / 5 |
Total | 14 / 20 Passed |