A better prompt explains what you want, gives the facts needed to make it, and says how the result should look. Add limits only when they matter. If the answer misses something, point to that part and explain the change. You do not need a secret phrase or a long role description.
What you need before you start
ChatGPT gives a vague answer, guesses missing facts, uses the wrong tone, or makes something you cannot use.
Bring: The result you want, the reader, the source facts, important limits, and a short example when style or structure matters.
Start with these steps
- Name a finished result: a reply, a comparison table, a lesson, or a plan.
- Supply the facts and label what is still unknown. Keep source material separate from instructions.
- Explain who will use the answer and what they need from it.
- Specify a format and a few checks that match the job.
- Read the first answer, identify a real failure, and ask for a focused correction while preserving the parts that work.
A prompt you can adapt
Make [result] for [reader]. Use these facts: [facts]. The result must [goal]. Keep [limits]. Return [format]. If a fact is missing, [ask / use a placeholder / list it]. Do not invent details.
Replace every bracket with real information. Remove fields that do not affect your task.
What this looks like in practice
Teaching example: 'Write a reminder to a customer. Their appointment is Thursday at 2 p.m. Ask them to reply if they need to change it. Use a friendly text message, under 60 words. Do not add an address or cancellation fee.'
The four parts of a usable brief
What must exist at the end?
Which facts change the answer?
What must remain true?
Where will it be used?
Start with the job, not a topic
A topic tells ChatGPT what to talk about. A job tells it what to do. 'Customer service' is a topic. 'Write a reply to a customer whose parcel arrived damaged' is a job. If you want something you can use, name the result first. That small change often matters more than any fancy wording later.
Use an ordinary action word: write, compare, explain, sort, check, plan, or quiz. Then name the thing you need. For example: 'Compare these two quotes in a table' or 'Explain this paragraph using a household example.' The answer now has a shape you can judge.
Add the outcome when it is not obvious. A reply might need to request a photo, offer an approved replacement, or explain the next update. Without that goal, a polite draft can still leave the reader wondering what happens next. A good prompt makes the job clear enough that you could recognize a successful answer.
A simple prompt structure that works across tasks
Think of your prompt as a short brief you might give a helpful colleague. It needs the task, the relevant facts, and the result you want back. The reader and limits matter when they change the answer. You do not need to fill every category for a simple question.
For example: 'Explain compound interest to someone saving for the first time. Use a small made-up example, show each year's calculation, and explain that the rate in the example is not an offer.' This gives a learning goal, a reader, a useful example, and a boundary. It does not need a paragraph telling ChatGPT it is a world-famous expert.
For a one-line rewrite, an even shorter prompt may be enough: 'Make this sentence easier to understand without changing its meaning: [sentence].' Scale the brief to the task. More complex work needs more context; simple work needs a clear request. The purpose is to reduce guessing, not reach a certain prompt length.
| Part | What to give | Example |
|---|---|---|
| Task | A concrete job | Compare these two delivery options |
| Facts | Material the answer can use | Prices, arrival times, and order size |
| Reader | Who needs the result | A customer choosing shipping |
| Limits | Conditions that must hold | Keep the supplied prices unchanged |
| Format | The useful shape | A table and a short recommendation |
Give context that affects the answer
Context means the background the answer needs. For a meal plan, a kitchen without an oven matters. For a work email, a deadline that has already passed matters. For a learning question, what you already understand matters. Include details that would make you choose a different answer.
Leave out background that does not help. A long company history may bury the two facts a customer reply needs. Instead, provide the problem, the confirmed status, and the next action. If you have a large source, identify which section is relevant and ask the answer to stay within it.
Say which details are facts, estimates, or preferences. 'The budget is $500' sets a limit. 'We hope to spend about $500' may allow discussion of a higher-cost option. 'Make it inexpensive' leaves too much room for guessing. The same words can lead to different results, so use the clearest version of what you mean.
Separate instructions from source material
If you paste an email, document, or notes, label them. A simple line such as 'Source notes begin here' helps distinguish your instructions from the text to be worked on. Follow it with the actual material, then make clear where it ends. This is especially useful when the source itself contains requests or commands.
Tell ChatGPT how to use the source. 'Rewrite this' asks for new wording. 'Extract the agreed deadlines' asks for facts. 'Check whether these two passages contradict each other' asks for comparison. If you only paste text and say 'help,' you leave the whole job undefined.
For a task based on your material, write 'Use only these facts' when outside additions would be harmful. You can still ask for clearer wording or useful organization. The boundary restricts factual claims, not good writing. Ask it to put unknowns in a separate list instead of silently making the draft feel complete.
Use an example when style or format is hard to describe
Words such as friendly, professional, or simple can mean different things to different people. A short example shows what you mean. Give two or three sentences of your own writing, or create a safe sample. Explain what to copy from it: sentence length, level of detail, or tone.
Do not ask it to copy all the sample's facts. If the sample says 'See you Friday,' that does not belong in a message about a Tuesday meeting. Say, 'Use the tone of this sample, but use only the facts in my brief.' This separates the writing style from the information.
A format example can be just as helpful. If you need a table with issue, evidence, action, and owner, show those column names. If you want a short notice, show a heading and two sentences. Examples are most useful when they remove a real uncertainty. You do not need several samples for a straightforward question.
A weak prompt, a better prompt, and a usable result
This teaching example uses invented appointment details. The weak prompt is 'Write a reminder.' It leaves ChatGPT to guess the event, person, date, tone, and requested action. A fluent result might be completely unsuitable.
A better prompt is: 'Write a friendly text-message reminder for a customer. Their repair appointment is Thursday at 2 p.m. Ask them to reply if they need to change it. Keep it under 60 words. Do not add an address, fee, or other policy.' The new information limits the answer to a real job.
A teaching result is: 'Hi! This is a reminder that your repair appointment is Thursday at 2 p.m. Please reply to this message if you need to change the time. Thank you!' The result includes the appointment and one action. It makes no new promise and adds no missing contact details.
Now imagine the result says 'Reply at least 24 hours before your appointment to avoid a fee.' A useful correction is: 'Remove the fee and 24-hour rule. Neither is in my facts. Keep the rest.' The correction names the failure, explains why it is wrong, and preserves the usable part.
Set limits that protect the result
Useful limits prevent a specific mistake. 'Do not invent prices' protects a comparison. 'Keep the quoted sentence unchanged' protects a source. 'Use one question at a time and wait for my reply' protects a tutoring session. Write limits that relate to the job rather than a long list of general good behavior.
Watch for conflicting instructions. 'Explain every detail in three short sentences' may be impossible for a broad topic. 'Make this warm and friendly, but use only five words' may force an awkward result. Choose which need matters more or ask for a short opening followed by a full explanation.
Length should fit the reader's needs. A short message may need a word limit. An article on a broad subject may need enough room to explain the subject well. If you ask for a brief answer and then complain that important details are missing, change the length requirement or divide the work into a short answer and a deeper guide.
Ask for clarification when a gap changes the answer
Sometimes ChatGPT needs an answer from you before it can do the job well. If the venue is unknown, it can still draft an invitation with [venue]. If the entire plan depends on whether you have a car, it should ask about transport first. The best treatment of a gap depends on how much that fact changes the output.
You can say, 'Ask up to three questions only if the answers would change the plan. Otherwise make reasonable assumptions and label them.' This helps avoid both endless questioning and silent guesses. For a quick draft, placeholders may be faster than a long interview.
If you do not know the answer, say so. Ask for two versions under different assumptions, or a list of what you need to find out. A comparison that shows where the decision changes is more useful than a single confident recommendation built on a hidden guess.
Improve the next answer with precise feedback
'Make it better' does not tell ChatGPT what failed. Try feedback linked to the task: 'The opening does not explain why I am writing,' 'The table mixes monthly and yearly prices,' or 'This explanation uses the word fraction before defining it.' Each tells it where the revision should act.
Quote a confusing sentence when the answer is long. Then say how you want it changed: 'Explain this in everyday words and add a small example.' If the problem appears throughout, show one bad passage and one acceptable version. That example can guide the whole revision.
Do not keep polishing a draft that lacks the facts needed to finish it. Get the missing information or mark it. If the conversation has grown tangled with rejected ideas, start a fresh chat with a clean brief and the confirmed source. A reset can help remove confusing context, but it does not replace the missing facts.
Build a small library of prompts you actually use
Save prompt patterns for recurring jobs, such as customer replies, meeting summaries, study questions, or weekly planning. Keep the placeholders obvious. A saved prompt should tell you what to supply, how the answer should look, and what to check.
After using a pattern, improve it only when you learn something. If customer replies repeatedly invent a delivery date, add a clear rule for unknown dates. If your study prompt keeps showing answers too early, ask it to wait for your attempt. These are useful changes because they fix observed problems.
Treat the saved prompt as a starting point. A new task may need current web sources, a different audience, or a file. Replace the old facts every time. The real skill is understanding what information makes the answer useful. Once you can name the job, supply the source, and give clear feedback, you can write good prompts without memorizing a formula.
Check the result before you use it
- Can I tell whether the answer did the task?
- Did I provide the facts needed to do it?
- Are any instructions fighting each other?
- Does the answer preserve the important numbers and promises?
Watch for: Adding more instructions can make a prompt harder to follow. Give the missing fact or a clear example before adding generic commands.