Lean and efficient: each step pairs executable code with one terse, load-bearing note (e.g. 'cat_columns matters — Giskard treats categoricals differently for slicing + drift detection') and never explains basic concepts Claude already knows, fitting the every-token-earns-its-place anchor.

Every step provides complete, copy-paste-ready Python (Dataset/Model wrapping, scan, suite generation, catalog tests) plus a concrete CI YAML snippet — fully executable with specific parameters rather than pseudocode.

A clearly numbered 7-step sequence with an explicit pass/fail gate (Step 5 raises SystemExit when suite_results.passed is false) and an Anti-patterns table serving as a pitfall checklist; this is not a destructive/batch workflow so no cap applies.

A single well-organized SKILL.md with clear section headers and one one-level-deep external reference (the Giskard tabular quickstart link, used consistently), with no nested/bundle references to chase.

Lists multiple concrete actions — 'scan() vulnerability detector + test catalog', 'Wrap a prediction function in giskard.Model + a DataFrame in giskard.Dataset', 'emit test suites that pass/fail in CI' — matching the multiple-specific-actions anchor.

Explicitly answers both what (scan + catalog, wrap model/dataset, emit CI suites) and when ('Use when a trained tabular or NLP model is about to ship with no test suite of its own, or when a feature-engineering or hyperparameter change needs a pre-merge scan').

Covers natural terms a user would say — 'ML models', 'test suite', 'fairness', 'pre-merge scan', 'feature-engineering or hyperparameter change' — beyond library jargon, with good variation coverage.

Occupies a clear niche (Giskard-specific ML vulnerability scanning) with distinct triggers tied to model shipping and pre-merge scans, making overlap with other skills unlikely.