Conciseness

The body is dense and operational — command blocks, threshold tables, and checklists with almost no explanation of concepts Claude already knows (MCAR/MAR/MNAR theory is delegated to the reference file). Minor trimmable padding remains, e.g. the persona opener 'You are an expert data quality engineer... You move fast, think in impact' and some feature bullets that restate what the command examples show. Not anchor 5 because those bits don't fully earn their tokens; not anchor 3 since verbosity is isolated, not systematic.

Actionability

All three scripts get copy-paste bash examples with real flags (--columns, --threshold 0.05, --method zscore --threshold 2.5, --monitor, --format json), and the remediation playbook uses concrete thresholds. However, Mode 1 step 4 'Cross-column checks — Inspect referential integrity, duplicate rows, and logical constraints' provides no executable command or method, matching anchor 4's 'minor gaps' rather than anchor 5's fully covered common cases.

Workflow Clarity

Each mode is a clearly numbered sequence (Profile → Missing Values → Outliers → Cross-column → Score & Report), and risky operations carry checkpoints: 'Confirm uniqueness key with data owner before deduplication' and 'Never auto-remediate 🔴 findings without human confirmation'. Anchor 5 is not met because there are no error-recovery feedback loops (e.g., what to do when a script fails on mixed-type columns) and cross-column validation is implicit rather than an explicit checkpoint.

Progressive Disclosure

Well-sectioned overview with all operational content inline and theory pushed one level deep to a clearly signaled, real file ('references/data-quality-concepts.md — MCAR/MAR/MNAR theory, DQS methodology, outlier detection methods'). All script paths referenced in the body exist in the bundle, and navigation via headers and tables is easy — matches anchor 5.

Specificity

Lists multiple concrete actions — 'Audit datasets for completeness, consistency, accuracy, and validity', 'Profile data distributions, detect anomalies and outliers, surface structural issues, and produce an actionable remediation plan' — covering the full assess-to-remediate workflow. Anchor 4 (minor gaps in coverage) fits worse since no capability area of the workflow is missing.

Completeness

Explicitly answers both parts: what ('Audit... Profile... detect... produce an actionable remediation plan') and when ('Use when the user asks to check data quality, profile a dataset, hunt outliers or missing values, or validate data before analysis or model training') with concrete trigger phrases. Matches anchor 5 exactly; anchor 4 would require the 'when' to be less explicit.

Trigger Term Quality

Good natural-phrases coverage: 'check data quality', 'profile a dataset', 'hunt outliers or missing values', 'validate data'. Falls short of anchor 5's 'comprehensive coverage... including synonyms and file extensions' — no data-file formats (CSV, Parquet) and common variations like 'clean data', 'data issues', or 'data health' are absent.

Distinctiveness Conflict Risk

The data-quality-audit niche is mostly distinct with quality-specific triggers, but broad phrases like 'profile a dataset' and 'validate data before analysis' carry minor overlap risk with general data-analysis or EDA skills. Anchor 5 requires minimal conflict risk with distinct triggers, which these broader phrases don't fully meet.