How a prompt is tested before it gets the label

Some prompt pages on this site carry a label such as "Run on ChatGPT, Claude and Gemini on 13 Nov 2026". This post explains exactly what is behind that label. It is a machine test. Nobody on the team sits down and tries each prompt by hand, and no person reads every answer.

Step 1: an invented work example

A prompt has gaps for your own material, such as [PASTE YOUR NOTES] or [WHO THE EMAIL IS FOR]. A script asks a small AI model to invent one short, believable work situation and to fill every gap with it. For example: notes from a project meeting at a made-up company, or a customer complaint about a late order.

The example is invented on purpose. It contains no reader's data and no company's real documents.

Step 2: the same prompt on three assistants

The filled prompt goes to three current AI assistants, one each from OpenAI, Anthropic and Google: ChatGPT, Claude and Gemini. Each one gets it the way you would use it: pasted into a new chat. A prompt with several steps is sent step by step in the same chat.

Some prompts tell the AI to ask you questions first and then wait. When an assistant does that, the script answers once: "That is everything I have. Use your best judgement and go ahead."

Step 3: a separate model checks each answer

A fourth AI model, which is not one of the three being tested, reads each answer. It scores it from 0 to 1 against what the prompt page itself promises: "What it does", "Use it when", "Watch out for", the shape of the output the prompt asks for, and the sample on the page. An answer passes at 0.7 or higher.

Step 4: the label

The prompt passes when at least two of the three assistants pass. Only then does the page get the label, with the date of the run and the names of the models in the "How we test" note. A prompt without the label has not been through this test, and its page does not suggest that it has.

When a new prompt fails, the problems the checking model found are used to revise it once, and it is tested again. If it still fails, it is not published as tested.

What the test does not prove

• The checker is a model too. It can be wrong, in both directions.
• One example per run. A prompt that works on one invented situation can still fail on your very different material.
• Your answer will differ. AI tools give a different answer each time, and they change from month to month.

So treat the label as "this prompt worked in a fair trial on three tools", not as a promise. Read what you get back before you use it.

Use the same idea at work

You do not need a script to get most of the value. Before you share a prompt with your team:

1. Run it on a messy, real example, with private details removed.
2. Try it in a second AI tool, if your team uses more than one.
3. Read the answer and ask: did it add anything that was not in my input?
4. Write down one case where it fails, so your team knows what to watch for.

Browse the prompts in the prompt library, or see past editions in the issue archive.