Content
72%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.
The body is an exceptionally concrete internal engineering contract: every rule names its owning file, function, or regression test, and it wastes no tokens on concepts Claude already knows. Its weaknesses are structural — monolithic multi-rule bullets instead of a tightened, split-out layout — and the absence of any sequenced workflow, though as a constraints/ownership skill it has less need for one.
Suggestions
Split the dense multi-clause bullets into shorter single-rule statements (or a bullet list per rule) so each invariant is individually scannable.
Move the long regression-test inventories and the test.evidence config-key ownership rules into a separate reference file (e.g., references/evidence-config.md) and link to it from SKILL.md, improving progressive disclosure.
For the few genuinely procedural flows (e.g., publishing evidence to the orphan branch, media attach with token failure), present a short numbered sequence with an explicit validation checkpoint so workflow clarity does not depend on implicit checkpoints.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with repo-specific invariants, owners, and test names that Claude could not know, with no padding or explanations of generic concepts, matching 'Efficient; minor instances of over-explanation that could be trimmed'. It is not a 5 because many bullets are extremely long run-on sentences packing multiple rules together, which could be tightened into shorter scannable statements. | 4 / 5 |
Actionability | Guidance is fully concrete: exact functions and files (publishRunEvidence in internal/pipeline/steps/evidence_publish.go, paths.ResolveEvidenceRoot, RunManager.pinRunEvidenceDir), exact commands and rules (os.Remove, never RemoveAll; no runtime.GOOS branch; fail-closed upload errors), and named regression tests per behavior. Per the rubric's code-vs-instruction note, the absence of code is not penalized because the instruction-only guidance is directly actionable. | 5 / 5 |
Workflow Clarity | The content presents invariants and ownership rules rather than a sequenced procedure; sections like 'Regressions:' and fail-closed error handling act as implicit checkpoints, but no explicit step sequence or validate-fix-retry loop is given. It is not a 4 because the anchor requires a clear sequence with most checkpoints present, and not a 2 because the mechanisms described (collect, publish, attach, render, cleanup) are each coherently defined with their failure modes. | 3 / 5 |
Progressive Disclosure | There are two clear section headers ("Test Evidence Stays Out of the Code Branch", "no-mistakes Owns Its Own Scratch") and no nested references, but roughly 60 lines of dense monolithic bullets inline config-key ownership rules, rendering-surface rules, and long regression-test lists that could live in separate reference files. It is not a 4 because content that clearly belongs in separate files is inlined and the huge single-paragraph bullets make navigation hard; it is not a 2 because real section structure and consistent per-topic organization do exist. | 3 / 5 |
Total | 15 / 20 Passed |