Content
73%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-architected orchestration skill: the five-phase workflow with gates, mode detection, and the Definition of Done checklist are exemplary, and the thin-index intent with per-phase on-demand loading is exactly right. The main costs are heavy restatement of the same invariants across principles, anti-patterns, and DoD, and references that either cannot be verified in the provided bundle or point outside it.
Suggestions
State each invariant once: keep either the per-phase prose or the workflow table (not both), and merge Core Principles / Anti-patterns / Definition of Done — 'null is never confirmation' and 'never assume a global install' each currently appear 5-6 times.
Inline a minimal receipt example (command, raw output, verdict token) in the V4 section so the core output contract is usable before loading rules/receipt.md.
Fix the reference that escapes the skill directory (../../../agents/shared/rules/verification-receipt.md) by moving that content inside the bundle, e.g. rules/consumer-wiring.md, so every referenced path resolves portably.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes Claude's competence — it never explains known concepts — but the same rules are restated many times: the 'null is never confirmation' invariant appears in the boundary section, V4 gate, V4 prose, Core Principle 6, an anti-pattern, and the DoD (~6 times), 'never assume a global install' appears 5 times, and the V1–V5 workflow is narrated three times over (phase table, per-phase prose, Required Reading table). Not 2 since every line carries real information; not 4 because the structural repetition goes beyond 'minor instances that could be trimmed'. | 3 / 5 |
Actionability | Gives concrete, executable guidance: specific tool names per tier ('grep, ast-grep', 'tsc --noEmit', 'go build/go vet/staticcheck', 'cargo check/clippy', 'pyright/mypy'), a mechanical mode-detection rule ('Parse the first token of $ARGUMENTS'), and tier-selection heuristics ('A claim about symbol absence ... is usually a Tier 1 grep'). Not 5 because the core output contract — an example receipt — is deferred to rules/receipt.md rather than shown inline, leaving a gap if that file is unavailable. | 4 / 5 |
Workflow Clarity | Five explicitly sequenced phases (V1–V5) each with a named gate, a mode-detection table, a fallback chain in toolchain discovery (checks.yaml → inspector → manifest scripts), verdict disposition rules ('null or non-reproducing result drops or contradicts a finding'), and a Definition of Done checklist. The stop-early ladder rule and DoD together act as explicit validation checkpoints, matching the 5 anchor rather than 4. | 5 / 5 |
Progressive Disclosure | Explicitly a 'thin index' with one-level-deep, well-signaled per-phase references and a 'load on demand — do not preload' rule — textbook structure. Not 5 because the referenced rules/*.md files are not present in the provided bundle (the references cannot be verified to resolve), and one reference escapes the skill directory entirely ('../../../agents/shared/rules/verification-receipt.md'), which will break in a portable skill layout. Still clearly above 3 since references are prominently signaled, not buried, and content is appropriately split. | 4 / 5 |
Total | 16 / 20 Passed |