Content
88%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 unusually disciplined process skill: fully sequenced workflow with genuine validation checkpoints, concrete paths/commands/verdict labels instead of abstraction, and bundle content properly delegated to real reference files. The residual costs are repetition of the posting-authorization rules across four sections and the two-hop category index.
Suggestions
State the posting-authorization and "requests are never optional/deferred" rules once in a single section and reference it from sections 5, 6, and Tone, cutting several hundred repeated tokens.
Consider listing the most common category pages directly in the References section (or inlining the category index into the References list) so navigation to them is one hop instead of going through references/categories/README.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's intelligence throughout — no explaining what a PR, diff, or blame is — and nearly every sentence carries a directive ("Do not run `git log` before writing the pre-diff model—the current checkout may already be the PR branch"). It is not a 5 because the posting-authorization and "findings are change requests, never optional" rules are restated in sections 5, 6, Tone, and Judge criteria; consolidating those would trim real tokens. | 4 / 5 |
Actionability | For an instruction-only skill the guidance is maximally concrete: an exact review-record path (`.mastracode/scratch/reviews/<owner>-<repo>-<pr>.md`), specific commands (`git log -S`, blame), an explicit ordering constraint before opening the diff, named verdict labels (approve / request changes / needs discussion), and a per-claim verification menu (bug fix, public API, refactor, performance, schema, packaging). The scoring note says absence of code is not penalized when guidance is this actionable. | 5 / 5 |
Workflow Clarity | The sequence is explicit and numbered (build the model → understand what changed → review → recover history → verify → decide and brief), with validation checkpoints throughout: the pre-diff model must be written before the diff is opened, conclusions must be scrutinized before finalizing, broken-control tests are required where detection ability is uncertain, and re-review reloads only what changed. Feedback loops (drop a request on contrary evidence, fix-and-revalidate) are stated, and the outward-facing posting steps are gated by explicit approval checks. | 5 / 5 |
Progressive Disclosure | The three bundle references (archaeology.md, templates.md, categories/README.md) are all real files, well-signaled both inline and in a References section, and the dense category knowledge is properly pushed out of SKILL.md. It is not a 5 because navigation goes two hops for categories (SKILL.md → references/categories/README.md → the individual category pages), which is an index pattern rather than strictly one-level-deep references. | 4 / 5 |
Total | 18 / 20 Passed |