A good AI prompt states the task, supplies relevant information, and describes the result you need. Add important limits and explain what the assistant should do when information is missing. Write the request so you can check whether the answer actually fulfilled it, rather than relying on impressive wording.
This guide shows how to write a good AI prompt for an ordinary drafting task. You will turn a fictional book-swap brief into a clear request, inspect a sample answer, and revise a specific weakness. The method can improve communication with an assistant, but it cannot guarantee accuracy or replace your review.

Key Takeaways
- Start with one concrete deliverable and the person who will use it.
- Supply the facts the task needs, while leaving out irrelevant or sensitive material.
- Specify format, length, and the few constraints that genuinely matter.
- Revise a failed requirement explicitly and verify the result before using it.
What should you ask the AI to produce?
Replace a broad topic with an action and deliverable. ‘Help with an event’ could mean a planning checklist, invitation, budget, or social post. ‘Draft an invitation for a community book swap’ gives the assistant a more useful starting task.
Name the audience and purpose. An invitation for neighbors unfamiliar with the event needs different detail from a reminder for registered volunteers. Without that context, the assistant may choose a plausible audience that is not yours.
Google Cloud’s prompt-writing guidance recommends explaining what you are trying to accomplish and providing relevant detail. You can apply that principle in plain language without learning special syntax or adopting an elaborate prompt formula.
Choose a success condition before requesting the draft. For an invitation, that might be: the reader understands what the event is, when it happens, what to bring, and whether they can come without bringing a book. The goal is a useful message, not simply an enthusiastic one.
Which context should you include, and what should you leave out?
Provide the facts that change the answer. Those may include the event details, the reader’s experience level, a passage to summarize, or the requirements of a particular document. Background that does not affect the task can make the request harder to inspect.
Separate supplied information from instructions. Use simple labels such as ‘Task,’ ‘Confirmed details,’ and ‘Output.’ If you paste reference text, identify it as material to use rather than mixing it into the sentence that tells the assistant what to do.
Do not include private customer records, confidential business information, or other sensitive material merely to make a prompt more detailed. Use information you are authorized to share with the service and check its applicable data settings. A fictional or anonymized example is often enough for learning the method.
If something important is unknown, say so. ‘The street address is not yet confirmed; list it as missing and do not invent one’ is clearer than leaving the assistant to guess whether an address should appear. More detail is useful only when it is relevant and reliable.
How do output format and constraints improve the request?
Specify the shape of the answer. A table can help compare options; a short paragraph can work for an invitation; numbered steps suit a procedure. Asking for everything in every format usually creates material you will have to cut later.
Give a realistic length limit if the destination needs it. ‘Under one hundred words’ is more checkable than ‘quite short.’ Leave room for the essential information and count the final text yourself if the limit matters.
Anthropic’s prompting guidance emphasizes clear instructions, output requirements, and useful examples. A small original example can demonstrate your preferred style, but identify it as a style example so its facts are not accidentally imported into the new answer.
| Vague requirement | More usable instruction | How you can check it |
|---|---|---|
| Make it professional | Use plain, friendly language for local residents | Read for jargon and exaggerated claims |
| Keep it short | Keep the invitation under 100 words | Count the invitation only |
| Include everything | Include date, time, venue name, cost, and book limit | Check each supplied fact |
| Fill in any gaps | List missing practical details separately | Look for invented addresses or arrangements |
Keep the constraints compatible. A very short message cannot also include a detailed policy, several examples, and a complete FAQ. Decide what belongs in the deliverable and what should remain a separate note.
What does a complete prompt and reviewed answer look like?
Start with an explicit fictional brief
The following example is invented for practice. It does not announce a real event. Its confirmed details are a book swap at Riverside Reading Room on 18 April, from 2 to 4 p.m.; entry is free; participants may bring up to three books in good condition, and people without books may also attend.
Here is a complete prompt you can adapt:
- Task: Draft an invitation for local adults who have not attended our community book swap before.
- Confirmed details: Riverside Reading Room; 18 April; 2–4 p.m.; free entry; up to three books in good condition; visitors without books are welcome.
- Output: First give one invitation under 100 words. Then give a separate list of practical details we still need to confirm.
- Style: Friendly, clear English. Describe the event without exaggerated promises.
- Limits: Use only the supplied facts. Do not invent an address, refreshments, registration requirement, or accessibility arrangements. Do not publish or send the invitation.
Inspect a possible answer against the brief
A human-written example of an acceptable invitation is: ‘Join our community book swap at Riverside Reading Room on 18 April, from 2 to 4 p.m. Entry is free. Bring up to three books in good condition to swap, or come along without books. It is a chance to browse and meet other local readers.’
The separate editorial note could ask for the venue’s street address, access information, and whether any arrival instructions are needed. Those questions do not belong inside the list of confirmed arrangements.
The invitation stays within the requested limit and preserves the essential details. It does not promise a café, describe an accessible entrance without evidence, or require registration. The useful outcome is a draft ready for someone to complete and review, not a claim that an AI-generated response was measured or guaranteed.
How should you revise an answer that misses the point?
Identify the failed requirement. ‘Make it better’ leaves the problem unclear. ‘The draft says registration is required, but that was not provided. Remove that statement and keep the other confirmed details unchanged’ gives a specific correction.
If the tone is too promotional, name the phrases you want removed and the effect you want instead. Ask for a plain statement of what happens rather than adding more adjectives to ‘professional.’ Keep any facts that already passed your check.
Google Cloud’s prompt-engineering overview includes iteration as part of the process. Treat revision as a targeted experiment: change the instruction that addresses the problem, then compare the next answer against the same brief.
Do not assume the revision preserved everything else. Check the time, quantity limit, and attendance conditions again. A response can fix one issue while introducing another. For source-dependent claims, use our AI fact-checking guide rather than asking the assistant to certify its own correctness.
When should you split a task or save a prompt as a template?
Split the work when the deliverables depend on decisions you have not made. For example, confirm the event brief before generating an invitation, poster text, and volunteer instructions. Otherwise every later output may inherit the same unresolved assumption.
Save a template once its structure repeatedly helps. Replace changing facts with clearly marked fields and review them each time. A template is a reusable request, not a permanent guarantee that a new tool or model will behave identically.
For learning tasks, our AI study-tutor guide shows how to keep your own attempts and feedback central. Your AI Learning Coach is a related catalog resource for building a broader practice process.
Keep a final human checkpoint appropriate to the task. Routine wording still needs factual review; consequential decisions may require qualified judgment or another workflow entirely. A clear prompt improves the starting instructions, while responsibility for using the result remains with the person doing the work.
Conclusion
Choose one small drafting task and write its deliverable, confirmed information, output format, and limits. Check the first answer against those requirements. Revise one concrete failure, then review the whole result again before sharing or using it.
Frequently Asked Questions
Do I need to tell the AI to act as an expert?
A role can help communicate the kind of writing or audience you want, but it does not create expertise or verify the answer. ‘Explain this to a beginner gardener’ may be more useful than an elaborate professional persona. Keep the task, evidence, and review criteria explicit whichever role you choose.
Should every prompt contain an example?
No. A straightforward task may need only clear instructions and the relevant material. Add an example when it resolves ambiguity about style, structure, or level of detail. Make sure it is yours to share and label what the assistant should imitate, so it does not reuse unrelated facts from the example.
