Conciseness

Mostly efficient and dense with non-obvious domain knowledge (the replace-vs-merge table, per-surface authentication, verify-on-the-visible-surface), but several sections could be tightened: 'The request is the authorisation' restates its table in a follow-up narrative paragraph ('So: a user who says... gets the plan first...'), the MCP-vs-local-install guardrail over-explains runtime variations, and rhetorical flourishes ('it is a failure, and it is the most common one in this domain') pad the message. It is not 2 because none of the prose explains concepts Claude already knows — it is all domain-specific earned tokens, not generic padding.

Actionability

Concrete executable guidance is present: three copy-ready probe commands ('aws sts get-caller-identity', 'aws datazone list-domains --region <region>', 'aws glue list-form-types --region <region> --max-results 5'), a specific snapshot pattern ('/tmp/<asset>-before.json'), named verification commands ('glue get-table', 'glue get-table-versions'), and a router table pointing to real files. It is not 5 because the body is a router — the actual write-command blocks users would copy-paste live in the reference files rather than here — and not 3 because what is inline is fully executable rather than pseudocode.

Workflow Clarity

Despite destructive and batch operations, validation is explicit and central: the numbered 'snapshot, write, verify, restore' loop with a restore-on-mismatch error-recovery step, verification 'on the surface where the damage is visible' with a concrete counter-example (list-iterable-forms masking a destroyed StorageDescriptor), a mandatory platform-detection step before anything, and a read-back-before-reporting-success rule. This clears the destructive/batch cap with feedback loops and checklists matching the anchor-5 example, so it is not 4.

Progressive Disclosure

A router table maps twelve distinct user intents to twelve real reference files (all verified present in references/, one level deep, no nesting), each labeled with applicable platforms, while cross-cutting invariants are kept in the overview with an explicit justification for why they outrank chapters. It is not 4 because navigation is unambiguous and every reference is well-signaled with no content that clearly belongs in a separate file left inline — the overview holds only globally applicable rules.

Specificity

The description lists a dozen-plus concrete actions ('add a description to a table, document a dataset, write a README, set an owner, add column comments, tag with a glossary term, create a business glossary, create a metadata form, publish an asset, create a data product, run the data source crawl, delete a catalog asset') plus diagnostic question-forms, giving comprehensive coverage of the write surface. It is not 4 because no meaningful capability is left as a gap; even platform detection ('Detects which catalog platform an account uses before writing') is stated.

Completeness

Both questions are answered explicitly: what ('Writes and curates business metadata in an AWS data catalog: the Amazon SageMaker Catalog (Amazon DataZone), or the AWS Glue Data Catalog. Detects which catalog platform an account uses before writing') and when ('Triggers on:' followed by a concrete trigger list, plus a 'Not for:' boundary list). It is not 4 because the 'when' is not merely present but enumerated with concrete trigger phrases, matching the anchor-5 example structure exactly.

Trigger Term Quality

Triggers are natural user phrasings including verbs and question-forms users would actually say: 'write a README', 'set an owner', 'tag with a glossary term', 'which assets in my project are missing descriptions or owners and what to fix first', 'which catalog does this account use and what can I do with it'. It is not 4 because both imperative-form and question-form synonyms are covered comprehensively for a service-domain skill (no file extensions apply here).

Distinctiveness Conflict Risk

The 'Not for:' clause explicitly delegates four adjacent capabilities to named sibling skills ('use exploring-data-catalog', 'use finding-data-lake-assets', 'use querying-aws-sagemaker-catalog', 'use managing-data-asset-access'), and the write-path niche is clearly claimed. It is not 4 because the explicit routing of near-miss intents makes overlap risk minimal rather than minor.