How to write prompts that produce useful results

Start with the work you need done

“Help with my business” is too broad to evaluate. “Turn these notes into a customer reply of under 120 words, preserving the quoted price and completion date” gives the assistant a concrete target. OpenAI's getting-started guidance similarly centers the goal and the context supplied to a conversation. OpenAI: Use ChatGPT — Official Learning Resources

Use an audience description only when it changes the result. “For someone who has never used a spreadsheet” is useful. A long list of prestigious roles can add noise without adding facts. If a professional decision matters, assigning a professional persona does not supply the credentials or accountability of a real professional.

A reusable prompt structure

Start with the task. Add the approved facts or source material. State boundaries such as word count, tone, and details that must not change. Then specify what a good result looks like. A sample answer can communicate your preferred style more effectively than a dozen adjectives.

Turn the notes below into a clear email for a first-time customer. Use only the supplied facts. Keep prices, dates, exclusions, and promises unchanged. Use short paragraphs. Finish with the one question I need answered. List any missing facts separately from the email.

For a comparison, name the criteria: cost, included work, exclusions, timing, and unknowns. For a tutorial, specify the starting skill level and the device. For a summary, tell the assistant whether uncertainty, dissent, and exceptions must be retained.

Repair an answer precisely

Instead of saying “make it better,” identify the defect: “The tone is too formal,” “You omitted the cancellation condition,” or “The table combines a monthly cost with an annual cost.” Ask for a targeted correction and preserve the parts that already work.

When the task is complicated, separate stages. First confirm the facts and desired structure; then draft; then check. Do not require a long planning ceremony for a simple sentence edit. Extra conversation should remove uncertainty, not become a substitute for finishing the work.

Check the result against a small rubric

Use three questions: Does it preserve the facts? Does it satisfy the format? Can the intended reader act on it? Read a proposed customer message as if you were the customer. Look especially for invented commitments and polite language that hides the actual answer.

Keep a short prompt notebook with the task, your successful wording, and a finished example. Record why a prompt worked. This makes your own library useful even when product names, menus, and model versions change. Avoid collecting hundreds of generic prompts that you never use.