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 highly operational skill body: two clearly sequenced 7-step workflows with per-step template anchors that all resolve to real bundle content, explicit fail-closed validation and decision gates, and clean one-level-deep progressive disclosure. The main cost is token weight from systematic policy repetition — the same scoring and rights gating rules are restated across at least five sections where a single canonical statement plus cross-references would do.
Suggestions
Consolidate the NOT_SCORED/NEEDS_INPUT scoring rule (currently restated at lines 47, 58, 64, 87, 112, 126, 131) into one canonical section — e.g. a 'Scoring gate' block near the Instructions — and reference it from the other locations instead of re-stating it.
Merge the repeated rights fail-closed statements (paid step 7, repurpose step 7, Decision Gates, line 131) into the single most load-bearing location and cross-link the rest.
Trim the overlap between Skill Contract 'Writes'/'Promotes', the Handoff Summary block, and the Save Results section, which currently re-cover the same WARM/HOT authorization and identity-hygiene rules three times.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's competence — it teaches no concepts (no 'what is whitelisting' tutorial) — but the fail-closed scoring rule ('NOT_SCORED/NEEDS_INPUT with no /25, rank, tier, or spend') is restated at lines 47, 58, 64, 87, 112, 126, and 131, and the rights fail-closed rule is repeated about five times (lines 92, 104, 112, 131), with heavy overlap between Skill Contract, Handoff Summary, Decision Gates, Example, and Save Results. This is 'mostly efficient but could be tightened' rather than the 4 anchor's 'minor instances', since the redundancy is systematic rather than incidental — though it is deliberate policy emphasis, not concept over-explanation, so it stays above 2. | 3 / 5 |
Actionability | Both modes carry 7 numbered steps, each linked to a verified fill-in template anchor in references/templates.md (all 14 step headers plus both worked examples exist), with exact field names (`creator_ref`, `status_observed_at`, `status_evidence_ref`), an exact naming convention (`[campaign]_[creator_ref]_[platform]_[type]_[variation]_[date]`), named KPIs (CPM, CTR, CPC, CVR, ROAS), exact save paths, and a concrete markdown output table in the Example section. For an instruction-only skill this is copy-paste-ready guidance covering the common cases, matching the top anchor. | 5 / 5 |
Workflow Clarity | Each mode has a clear 7-step sequence with explicit validation checkpoints and feedback loops: 'Only when all five are present may the asset receive /25', missing evidence returns NEEDS_INPUT with 'the exact gaps' listed for re-request, rights states fail closed, and the Decision Gates section separates 'Stop and ask' from 'Continue silently' conditions. Per-mode 'Done when' checklists and the termination/visited-set rule round out error recovery, matching the anchor for explicit validation steps and checklists. | 5 / 5 |
Progressive Disclosure | The body is a genuine overview with one-level-deep, well-signaled references: templates.md (646 lines holding all step templates, platform setup guides, and both worked examples) and atom-extraction.md are real files whose anchor links all resolve, each step names its specific template section, and the Reference Materials section summarizes every file's contents. Content is appropriately split — the 646 lines of templates live outside SKILL.md — making navigation easy. The links pointing into the parent repo (sibling SKILL.md files, ../../../references/platforms/) are clearly signaled and one level deep, so they do not break the structure. | 5 / 5 |
Total | 18 / 20 Passed |