Conciseness

The body is mostly efficient — dense tables (EU AI Act tiers, stage-to-role map) carry judgment Claude doesn't reliably hold. But the ~60-item Keywords section duplicates the description's trigger list with no operational value, and time-stamped specifics ("Frontier models 2026: Claude Sonnet 4.6 ~$3/$15 per M tokens", dated regulation status) will go stale without being quarantined in a dated/old-patterns section. Not 4: the keyword dump and perishable pricing are real trimming targets, not minor instances.

Actionability

Quick Start and Workflows give concrete runnable commands ("python scripts/model_buildvsbuy_calculator.py path/to/use_case.json"), but the input format is underspecified — "Define use_case.json (volume, latency, accuracy, team size, budget)" is a field hint, not a schema or example — and none of the three referenced scripts exist in the bundle, so the commands cannot actually be executed as shipped. This is concrete-but-incomplete guidance with missing key details, matching anchor 3 rather than 4.

Workflow Clarity

All four workflows have goals, durations, numbered steps, concrete commands, cross-checks with other advisor skills, and decision-logging steps — a clear sequence. Not 5: checkpoints are advisory cross-checks rather than explicit validation of tool output (e.g., no sanity-check step on the TCO or breakeven numbers the calculators emit). Not 3: these are non-destructive advisory workflows, so the validation cap doesn't apply, and most checkpoints are present.

Progressive Disclosure

On paper the structure is good — clearly signaled, one-level-deep references ("See references/model_buildvsbuy_strategy.md for full decision tree") — but scoring against the actual bundle, no references/, scripts/, or assets/ directory exists: all four referenced .md files and all three referenced .py scripts are dangling paths. Navigation dead-ends at every pointer, so the promised deep content simply isn't there and the skill is effectively an unbacked monolith. This matches anchor 2's broken-structure level rather than 3, since the disclosure chain, not just signaling, is the problem.

Specificity

The description lists four concrete, distinct capabilities: "model build-vs-buy decisions (API vs fine-tune vs in-house)", "AI risk classification under EU AI Act + US state patchwork", "AI cost economics (API-to-self-hosted breakeven)", and "AI team org evolution" — comprehensive coverage of the skill's four domains, not vague fluff. Third-person voice is used correctly ("Chief AI Officer advisory for startups"), so no voice penalty applies.

Completeness

Both questions are explicitly answered: the "what" is the four named advisory domains with parenthetical specifics, and the "when" is an explicit "Use when..." clause with concrete decision triggers and keyword triggers. This matches the anchor-5 example structure (what + explicit use-when with concrete trigger phrases).

Trigger Term Quality

Triggers combine action phrases ("deciding whether to call an API or fine-tune", "calculating when self-hosting pays off", "sequencing AI hires") with a natural keyword list ("CAIO, AI strategy, model selection, foundation model, fine-tuning, EU AI Act, NIST AI RMF, AI governance, model risk, or AI economics") — these are the terms a founder or CAIO would actually say. Not below 5: no significant natural synonyms are missing for this domain.

Distinctiveness Conflict Risk

The closing "Strategic only — does not duplicate engineering AI/ML skills" explicitly carves out a niche and defuses overlap with tactical engineering skills. Not 5 because trigger terms like "model selection", "fine-tuning", and "foundation model" would also naturally fire for the referenced engineering skills (rag-architect, prompt-governance, llm-cost-optimizer) and adjacent C-level advisors (CFO/CDO on build-vs-buy TCO), leaving minor overlap risk that the single boundary sentence only partly resolves.