Content
85%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 lean, well-sequenced instruction body with excellent validation and error-recovery structure and a concrete canonical result schema. Its weaknesses are unresolvable references to the operating contract, capability manifests, and finding/worker schemas, and the absence of a bundle that would let a reader actually load those materials.
Suggestions
Add explicit file paths for the artifacts the procedure depends on (e.g., 'Read ads/contract.md' and 'Load references/manifests/<platform>.json') or include them in a references/ directory so the operating contract, capability manifests, control registry, and audit-worker-result v1 schema are resolvable.
Show the exact invocation for dispatching platform workers (e.g., the agent prompt template or command for 'audit-google') so the dispatch step is copy-paste executable rather than implied.
State where the common Finding v1 schema lives and how severity is looked up from the control registry, since validation in step 7 currently depends on schemas the body only names.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense operational instruction with no explanation of concepts Claude already knows ('Run deterministic scoring. Do not calculate or repair scores in the prompt.'), and the long JSON block is a canonical schema spec rather than padding, matching the 'every token earns its place' anchor. | 5 / 5 |
Actionability | Provides exact worker names, a complete canonical JSON result example, numeric coverage thresholds (60-79%, <60%), and named output files, but key execution inputs are referenced without resolvable locations ('Read the main ads operating contract', 'Load each selected platform capability manifest, control registry', 'the installed audit-worker-result v1 schema' — no paths and no bundle files exist). | 4 / 5 |
Workflow Clarity | The 11-step procedure is clearly sequenced with explicit validation checkpoints and feedback loops ('Validate every result against the common finding schema. Retry one transient failure; record all other failures and recovery hints', 'Verify bundle completeness, citations, privacy, and render integrity') plus a worked failure/weighting example. | 5 / 5 |
Progressive Disclosure | Sections are well-organized, but the skill depends on external materials (the 'ads operating contract', 'platform capability manifest', 'control registry', 'audit-worker-result v1 schema') that are neither given file paths nor present in a references/ directory, leaving navigation to the core detail material unclear. | 3 / 5 |
Total | 17 / 20 Passed |