Prompt Set QA

Decide pass, revise, quarantine, or reject. Do not quietly repair upstream work.

This skill inherits the ethical floor from skills/ETHICS.md. It enforces anti-hallucination, provenance, blinding, and decay-aware review. Anti-spray and human-send are not applicable.

Inputs

Require:

• prompt_universe.json;
• prompt_architecture.json;
• evidence excerpts and source metadata;
• the versioned contamination register;
• deterministic normalization, lexical scan, exact-hash, and similarity-pair results;
• optional blind human decisions.

Reject inputs that include baseline visibility, current rankings, target performance, answer-derived target pages, or selectors' preferred outcomes.

Run deterministic checks first

Check:

1. schema and provenance completeness;
2. unique IDs and resolved references;
3. target, product, domain, people, slogan, proprietary-category, campaign, flattering-claim, and competitor terms;
4. forbidden answer-derived fields;
5. Unicode normalization, language, length, and one-concept shape;
6. exact normalized hashes;
7. lexical similarity candidate pairs;
8. embedding pairs when a fixed model/version is available;
9. architecture coverage, budget, and aided/lane consistency.

Deterministic target or campaign matches are hard failures in unaided core prompts. B0 target aliases pass only through a declared allowed exception. Embedding similarity may nominate a pair; it may never auto-delete.

Review semantically

For each candidate, judge:

• evidence-to-prompt entailment;
• naturalness and role/locale authenticity;
• whether both variants preserve one canonical intent;
• proximity, journey, act, expected-answer, and aided-status consistency;
• commercial leading or recommendation forcing;
• semantic slogan or flattering-claim leakage;
• whether answer-derived language entered core;
• whether a similar prompt changes a material constraint.