Ten QA-agent prompts for checking work before it ships. Each prompt gives your assistant a role, numbered steps, a strict JSON output format, and a test plan that tells you how to verify the prompt itself: run it on a clean input and a broken input, and check the verdict flips. A QA prompt that always passes checks nothing.

What you get:

1. Data QA (3 prompts): dataset validation gate, dedup audit (exact, key collision, fuzzy), schema-drift diff across file versions.
2. Document and package QA (3 prompts): PDF extraction QA with page-tree and orphan check, zip package integrity with path-traversal scan, broken-reference sweep for README and manifest files.
3. API and output QA (4 prompts): fail-closed test matrix design for upload chains, output-format compliance grader, cover QA gate (baked text, busy ratio, contrast), honest-claims audit that checks copy against the actual artifact.

Every prompt ran once as an internal scenario during packaging, and the results ship in the pack. Eight of the ten checks were executed against real files in a working product pipeline: a 52-row state dataset, a 4-page PDF, a shipped 4-entry zip, a live README, and a published 1200x630 cover.

The test plans use negative tests on purpose. Package QA gets a copy with an injected traversal entry. The grader gets a reply wrapped in polite prose. The claims audit runs against a real artifact with counted rows. If the verdict does not flip on the broken input, the prompt is not doing its job.

Paste a prompt into ChatGPT, Claude, Gemini, or any capable assistant, fill the bracketed placeholders, and require the exact JSON format. The pack also includes a spreadsheet copy of all 10 prompts for Notion, Sheets, or Excel.

License: personal and commercial use. No resale or redistribution of the pack itself.