Content
81%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-engineered instruction skill: the readback workflow is unambiguous, sequenced, and rich with validation checkpoints and error-recovery rules, and it consistently delegates math and auditing to sibling skills. The main gaps are minor — placeholder values in the ledger command, an undefined numeric decision bar, some repeated boundary restatements, and references that all point outside the package and could not be verified.
Suggestions
Inline a compact default for the Promote/Rollback decision bar (e.g., a minimum delta-vs-control percentage or significance rule) so step 8 is executable without loading measurement-protocol.md.
Replace the '{"spend": ..., "revenue": ..., "conversions": ...}' placeholder with one complete ledger.py record example using realistic values.
Consolidate the auditor-boundary restatements (intro paragraph, step 7, step 8, Next Best Skill) into a single boundary note to cut repeated tokens.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with operational constraints Claude does not already know (learning-phase resets, "conversion lag (a click today converts days later...)", "Meta 7-day-click vs Google last-click are not comparable") rather than generic explanations, but the auditor/delegation boundary is restated in the intro, Skill Contract, steps 7–8, and Next Best Skill, which could be trimmed to a single note. This fits 'efficient; minor instances of over-explanation' rather than score 5's 'every token earns its place', and is clearly above score 3 since almost all content is skill-specific. | 4 / 5 |
Actionability | Concrete, near-copy-paste commands are given ('python3 "${CLAUDE_PLUGIN_ROOT}/scripts/connectors/ledger.py" record <campaign> --source paid --data ...' plus diff/trend variants) and the Quick Start covers three common cases, but the ledger payload uses '{"spend": ..., "revenue": ..., "conversions": ...}' placeholders and the Promote/Rollback decision bar ('beats control past the bar') is never defined numerically inline. This matches 'mostly executable guidance with minor gaps' rather than score 5's fully copy-paste-ready coverage. | 4 / 5 |
Workflow Clarity | Eight numbered steps in strict order with explicit stop conditions ('If the campaign is still in learning phase, **stop** — do not read or change it'), a validation checkpoint (step 7 signal-integrity check with 'mark the readback **Unproven**'), and a feedback loop ('State the concrete repair before any new readback... then restart the fixed readback window'), plus a 'Done when' checklist. This is the score-5 anchor: clear sequence, explicit validation, error-recovery loops, and a checklist. | 5 / 5 |
Progressive Disclosure | Sections are clear and the Reference Materials list signals each external file with what it contains ('readback windows, required readback fields, the control rule...'), but no bundle exists in this package (no ./references/, ./scripts/, ./assets/) and every detailed reference resolves outside the skill via deep relative paths (../../../references/skill-contract.md, ../../orchestrate/ad-test-designer/references/measurement-control.md) that cannot be verified here. Good structure with minor organization gaps fits score 4; it is not score 5 because navigation to the referenced detail is unverifiable and paths escape the skill's own directory. | 4 / 5 |
Total | 17 / 20 Passed |