Validates prompt sets for contamination, evidence entailment, duplicates, and architecture consistency.
npx skills add https://github.com/elvisun/newsjack --skill prompt-set-qa
Please help me install this Agent Skill. Skill: prompt-set-qa Source: https://github.com/elvisun/newsjack/tree/main/skills/prompt-set-qa Command: npx skills add https://github.com/elvisun/newsjack --skill prompt-set-qa
Prompt sets built for AI visibility measurement can be contaminated by target brand terms, unsupported claims, semantic duplicates, or answer leakage, which invalidates downstream panel results. This Skill gates a generated prompt universe before human panel selection, assigning each candidate a pass, revise, quarantine, or reject decision with evidence.
prompt_qa.json file with per-candidate decisions, rule results, duplicate actions, reconciled counts, and a gate status for human review.Ask the agent to run prompt-set-qa on prompt_universe.json and prompt_architecture.json with the contamination register and evidence excerpts, then produce the QA report and prompt_qa.json.
High-intent search queries and answers about installing and using this skill.
Run deterministic checks first for schema, provenance, contamination terms, hashes, and similarity pairs, then review each candidate semantically for entailment and naturalness. Assign each prompt a pass, revise, quarantine, or reject decision and emit a prompt_qa.json ledger.
Contamination is the presence of target brand, product, domain, slogan, campaign, or flattering-claim terms inside prompts that are supposed to be unaided. Deterministic matches in unaided core prompts are hard failures unless a declared allowed exception applies.
Duplicates are found through normalized exact hashes, lexical similarity pairs, and optional embedding pairs from a fixed model. Merges happen only when job, journey, constraints, and answer kind are materially the same; locale, persona, and aided-status differences are protected.
Quarantine prompts that are potentially useful but grade D, answer-derived, semantically suspicious, locale-unreviewed, or awaiting evidence or a human decision. Reject is reserved for contaminated, unsupported, materially leading, or irreparable cases.
No. The QA step must not quietly repair upstream work, invent sources, or change the architecture to rescue a prompt. Wording fixes are routed back to realistic-prompt-generation as revise decisions.
High-stakes prompts must stay inside the declared navigation or information boundary. Prompts asking the model to diagnose, prescribe, or make professional judgments are quarantined unless that judgment is explicitly in scope with qualified review.
shashwatgtmGenerate a ranked 25-prompt AEO pack for European B2B buyers.
supermalangScaffold deterministic Playwright screenshot baselines with human-only approval.
zilliztechDetect exact and near-duplicate records using hash checks and semantic similarity.
sajalsuhaneRun the CEO, design, and Eng review pipeline by reading skills from disk.