TESTING METHODOLOGY
How a prompt earns “Tested.”
Our verification badge means a human actually ran the prompt and recorded the result. Here is exactly what that involves — and what it does not promise.
Last reviewed: 2026-07-27
Two states: Draft and Tested
- Draft — a reusable template we believe is sound, but which has not yet been formally run and recorded. Most library templates start here.
- Tested — the prompt was run on a named model on a named date by a named reviewer, and the result matched the page’s description.
We never show a “Tested” badge on a Draft, and we never invent an example output. If the “Example output” section is empty, it is because we have not captured a real one yet.
What a test run involves
- Run the exact published prompt with the documented example input.
- Use the recommended model, with grounding/tools set as the page describes.
- Check that the output matches the promised structure and that the “why it works” claims actually hold.
- Try at least one documented failure case to confirm the fix works.
What we record
Each tested prompt’s verification record shows the model tested, the test date, and the reviewer. You can see this record in the sidebar of every prompt page.
What “Tested” does not promise
- It is not a guarantee. Generative models are non-deterministic; your run may differ.
- It does not certify factual accuracy of the model’s output — you must still verify sources and claims yourself.
- It reflects one point in time. Models change; we re-test and update the date.
Independent testing standard, not affiliated with or endorsed by Google. Gemini is a trademark of Google LLC.