Give your prompt a job to do.

A useful prompt is a clear brief: a task, relevant context, a usable output and a way to judge the result.

Name the deliverable

Replace a broad request such as “help with marketing” with a specific artifact: “Draft three subject lines for this approved newsletter.” Say who will use it, what decision or action it supports and what a good result must include. Define success before polishing the wording.

Supply the right context

Include the audience, relevant source material and important definitions. Label reference material separately from your instructions so it is easier to see what should be analyzed. Explain gaps rather than asking the model to guess. Use material you have permission to share and keep credentials out of the prompt.

Specify a format and an example

Ask for an output you can actually use: a table with defined columns, a brief with named sections or a short list of options. Where style matters, give a small original example and explain what makes it work. Avoid constraints that conflict, such as asking for an exhaustive analysis in a single sentence.

Make uncertainty and review explicit

Ask for source references when the input supports them, and have the model identify missing evidence. State whether you want a draft, analysis or an action. An agent brief should identify allowed tools, approval points and stop conditions; also configure permissions in the product itself. Prompt wording alone is not an access control.

A brief you can adapt

Task: summarize these five approved interview excerpts for a product team. Output: a table of themes, source IDs, counterexamples and open questions. Boundaries: use only the supplied excerpts; do not invent quotes or claim that this sample represents all customers. Review: flag every claim that needs the researcher’s judgment. This gives you something concrete to check against the original material.

Improve with examples of real work

Try the prompt on a few representative inputs, including an incomplete or awkward one. Check accuracy, completeness, format and the time needed to repair the output. Change one important instruction, then compare again. Record the model and date if you share results. Sometimes the missing ingredient is better source material or a different workflow, rather than a longer prompt.

Make one yours

Choose a template, fill in its fields and add your context. The preview is assembled in your browser; it does not call an AI model. Favorites save the template only, while a text download keeps your finished version. Open the prompt library ↗

Keep exploring

Updated October 7, 2026. Provider documentation can change.