Start with a Clear Job
Good prompts usually make four things explicit:
<ul>
<li>the task</li>
<li>the audience</li>
<li>the output format</li>
<li>important constraints</li>
</ul>
Prompting gets stronger when the model understands what success should look like.

Examples can improve structured output quality by showing the model the format or style you want.

Use Constraints and Format Deliberately
Useful prompt techniques include:
<ul>
<li>delimiters around source material</li>
<li>exact output schema</li>
<li>step ordering when needed</li>
<li>word or length constraints</li>
<li>explicit do-not-do rules</li>
</ul>
This is especially helpful when the output needs to fit into a product workflow.

What is the strongest way to request structured JSON from a model?

Longer prompts are always better than shorter prompts.

Handle Missing Information and Edge Cases
A robust prompt often says what to do when:
<ul>
<li>required context is missing</li>
<li>the request is ambiguous</li>
<li>the model should refuse or escalate</li>
<li>the answer should include uncertainty</li>
</ul>
This leads to better product behavior than hoping the model guesses correctly.

If a user's request is ambiguous and important context is missing, what is usually the best model behavior?

Iterating prompts is often necessary because prompt design is empirical, not purely theoretical.

Evaluate, Do Not Just Admire
Prompting is not a one-time copywriting task. It is an iterative product activity.
Strong teams:
<ul>
<li>test prompts on representative examples</li>
<li>compare versions</li>
<li>look for regressions</li>
<li>combine prompt work with analytics and manual review</li>
</ul>

Which PM habit is strongest when improving prompts in a production AI feature?

Prompt engineering should be paired with product analytics and manual review, not treated as a one-time copywriting task.