Write prompts you can test

Define the job before writing the prompt

Start with a task you can judge. “Help me study” has no obvious stopping point. “Turn this lecture extract into three revision questions, with answers supported by the extract” does. The second version lets you check whether the output did its job.

A useful prompt makes the task, input and desired output clear. Examples can help explain the format or standard you want. But extra words are not automatically extra precision: remove instructions that do not affect the result you need.

Use a small, inspectable prompt

This is an original exercise template. Paste a short public passage into the input section. Read it yourself first so you can spot invented answers. The markers organise the input; they are not a security boundary.

Task: Create revision questions from the passage.

Requirements:
- Write exactly three questions.
- Include a short answer and a supporting quote for each.
- Use only information in the passage.
- If there is not enough information, say so.

Output: Question | Answer | Supporting quote

<passage>
Paste your source passage here.
</passage>

Judge the result with the same checklist

Do not change your definition of a good answer after seeing the answer. Write the checks first, then run the prompt on several passages. Include one with too little information and one containing a distracting instruction inside the passage.

• Are there exactly three questions, or an appropriate explanation of insufficient information?
• Does each answer follow from the supplied passage?
• Is every quoted phrase actually present in the passage?
• Are the questions different enough to be useful revision?

Change one thing and compare

Keep the model and input fixed while you edit one instruction. Save both versions and record which checks improved or regressed. For example, an example row might improve formatting while encouraging repetitive questions. That is a trade-off you can see, rather than a feeling that the prompt is “better.”

Repeat a few runs when output variability matters. One successful answer does not establish reliability. If a particular passage keeps failing, preserve it as a future test case rather than quietly removing it from your examples.

Know when the prompt is not the problem

A model cannot quote a paragraph it never received. An application cannot parse a table if its parser expects JSON. Some failures belong to retrieval, input handling or validation rather than the wording of the prompt.

For your learning log, save the prompt, the test inputs, the checklist and one failure analysis. This exercise needs no elaborate framework. The valuable habit is turning “looks good” into a result another person could check.