Content
75%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-built instruction-only skill: dense, jargon-free of padding, with concrete formats, thresholds, and enums, a clear sectioned workflow with genuine validation gates, and clean one-level-deep referencing from the entry file. The main deficits are moderate redundancy between the policy sections and the Anti-Patterns/Red Flags sections, and bundle hygiene — orphaned reference files, an orphaned assets directory, and broken cross-bundle links.
Suggestions
Link the four orphaned reference files (brainstorm-evaluation.md, direction-quality-evaluation.md, explore-mode.md, research-spike.md) and the cost-record assets from SKILL.md's References section so they are discoverable from the entry point.
Fix or remove the broken cross-bundle links in decision-evidence-fidelity.md pointing at ../../common-tdd/references/quality-contract.md and ../../common-code-review/references/orchestration.md, which do not resolve in this bundle.
Consolidate the Anti-Patterns and Red Flags sections, which restate rules already stated in sections 1-6, to remove redundancy and recover token budget.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is extremely dense with zero concept over-explanation and every line carrying a rule, but the Anti-Patterns and Red Flags sections largely restate constraints already given in sections 1-6 ("No unlabeled claims" restates the evidence-tagging rule; "No pending approval leakage" restates section 6), which is trimmable redundancy rather than the every-token-earns-its-place of anchor 5. | 4 / 5 |
Actionability | For an instruction-only skill the guidance is mostly executable: a concrete SNC tier table, exact evidence tags ("confirmed(<path>)", "assumed", "unknown"), an approval enum ("approval: pending | approved(<who>, <YYYY-MM-DD>) | assumed-autonomous"), hard limits ("Review card ≤250 words", "batch max 3"), an exact reply phrase, and file pointers. It falls short of anchor 5 only because much of the guidance is policy constraints whose full copy-ready formats live in reference files rather than inline. | 4 / 5 |
Workflow Clarity | Sections 1-7 form a coherent sequence (size → shared understanding → evidence → questions → options → approval → publication) with real checkpoints (pre-approval placeholder/contradiction/ambiguity check, "NOT RUN/not assessed is never PASS", "Deep requires independent check"). It sits below anchor 5 because the numbered sections read as policy areas more than an explicitly ordered procedure, and the correction loop after "Reply ok or corrections" is implicit. | 4 / 5 |
Progressive Disclosure | The body links six one-level-deep reference files at their point of use plus a References section, and SKILL.md itself is a lean overview. However, scoring against the actual bundle reveals gaps beyond "minor": four of ten reference files (brainstorm-evaluation.md, direction-quality-evaluation.md, explore-mode.md, research-spike.md) are never linked from SKILL.md, the assets/ directory is reachable only through an unlinked file, and decision-evidence-fidelity.md contains broken cross-bundle links (../../common-tdd/…, ../../common-code-review/…). | 4 / 5 |
Total | 16 / 20 Passed |