Content
82%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, high-signal operational contract for a read-only audit skill: efficient prose, concrete classification rules, and clear sequencing with validation gates. The main gaps are unnamed artifacts — the shared references and the "common JSON contracts" are invoked but never identified.
Suggestions
Name or link the specific shared references in step 3 (e.g., which measurement, benchmark, or policy files to read) so the instruction is executable without guessing.
Identify the "common JSON contracts" by name or path in the Output section, since the conductor schema is central to steps 7-8.
Add a brief recovery loop for invalid output (e.g., re-check against the schema and re-classify when findings fail validation) to strengthen workflow checkpoints.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational throughout — procedure steps, boundary rules, and evidence requirements with no explanations of concepts Claude already knows. Sharp, non-obvious rules like "A new account does not prove a new Pixel" and "Raw events do not prove conversion maturity" earn their tokens; only the path-resolution parenthetical in step 3 is slightly wordy, which is not enough to drop below anchor 5. | 5 / 5 |
Actionability | Concrete instruction-only guidance: an 8-step procedure, named reference path (`ads/references/meta-audit.md`), and explicit classification rules with enum labels (`account_cold_start`, `pixel_cold_start`, `conversion_cold_start`). It falls short of anchor 5 because "the common JSON contracts" and the "shared measurement, benchmark, creative, automation, policy, and scoring references" are referenced without being named or located, leaving specific execution gaps. | 4 / 5 |
Workflow Clarity | The 8-step procedure is clearly sequenced (contract → inputs → references → normalize → evaluate → classify findings → return → render) with validation gates ("schema-valid findings", "only from the validated JSON run bundle", mutation gate for drafts). It is not anchor 5 because no explicit validate-fix-retry feedback loop is described (e.g., what to do when findings fail schema validation). | 4 / 5 |
Progressive Disclosure | No bundle files ship with this SKILL.md, and the ~60-line body is well-organized into Procedure, Boundaries, Cold-start evidence contract, and Output sections with a clear one-level reference to `ads/references/meta-audit.md`. Minor gap: the "shared ... references" are mentioned generically without names or paths, keeping it below anchor 5's well-signaled navigation. | 4 / 5 |
Total | 17 / 20 Passed |