A ChatGPT prompt is just an instruction, but the wording decides whether you get a vague paragraph or exactly what you needed. This guide shows the moves that reliably improve answers, then gives you six prompts you can copy, paste, and adapt.
Give it a role
Telling ChatGPT who it is sets the vocabulary, depth, and assumptions it uses. Compare "Explain APIs" with "You are a backend engineer explaining REST APIs to a junior developer." The second one narrows the audience and the level, so the answer stops hedging.
A role is one sentence at the start. Name the expertise and, if it matters, the audience:
You are a nutritionist writing for busy parents with no time to cook. Keep advice practical and cheap.
Do not overdo it. "World-class, award-winning, genius" adjectives add nothing. A plain, accurate role beats a flashy one.
Be specific about the task and the output
Most weak answers come from vague requests. "Write about our product" could mean a tweet, a landing page, or a manual. Spell out the task, the length, and what the finished thing should look like.
Say what you want in concrete terms:
- Task: what to produce (an email, a summary, a function).
- Length: word count, number of bullets, or paragraphs.
- Audience: who reads it and how much they already know.
- Constraints: tone, things to avoid, must-include points.
Weak: "Summarize this article." Better: "Summarize this article in five bullets for a busy executive. Each bullet under 15 words. End with one sentence on why it matters."
Add context
ChatGPT only knows what you tell it in the conversation. If the answer depends on your situation, paste the situation. For a reply to a customer, include the customer's message. For a code fix, include the code and the error. For a study plan, include how many hours a week you actually have.
You can label context so the instruction stays separate from the material:
Rewrite the email below to sound warmer but stay professional. Keep it under 120 words. EMAIL: Hi, your invoice is 14 days overdue. Pay now to avoid a late fee.The more the task depends on facts only you have, the more context you should paste. Do not make ChatGPT guess your product name, your deadline, or your audience.
Show examples (few-shot)
When you want a specific style or format, show one or two examples instead of describing them. This is called few-shot prompting, and it works because a good example carries more information than an adjective.
Say you want product taglines in a particular voice:
Write taglines in this style: Input: a note-taking app Output: Your brain, backed up. Input: a budgeting app Output: Know where it went. Now do this one. Input: a habit tracker Output:Two examples are usually enough to lock in a pattern. Use this whenever the shape of the output matters more than you can easily put into words: classification labels, a consistent JSON structure, a brand tone.
Ask for step-by-step reasoning when it helps
For math, logic, planning, or anything with multiple steps, asking ChatGPT to work through it out loud reduces mistakes. A phrase like "think step by step" or "show your reasoning before the final answer" pushes it to check itself.
A shop sells pens at 3 for $2. I have $10. How many pens can I buy? Work through it step by step, then give the number on its own line.
Skip this for simple lookups or short creative tasks where reasoning just adds noise. It earns its keep on problems where a wrong intermediate step ruins the result.
Control the format
If you do not say how you want the output shaped, you get prose. Ask for the exact format and you save yourself cleanup. You can request bullets, a numbered list, a table, JSON, a fixed word count, or a specific set of headings.
- Table: "Return a table with columns Name, Price, Pros, Cons."
- JSON: "Return only valid JSON with keys title, tags (array), summary. No text outside the JSON."
- Length: "Answer in exactly three sentences."
- Structure: "Use these headings: Problem, Options, Recommendation."
When you need data you will paste elsewhere, JSON or a table is the difference between usable output and a paragraph you have to retype.
Iterate with follow-ups
You rarely need the perfect prompt on the first try. Because ChatGPT remembers the conversation, you can steer the result with short follow-ups: "shorter," "more formal," "add a code comment on each line," "redo the second option, the first is fine." Treat the first answer as a draft and refine it. This is often faster than writing one enormous prompt.
System and custom instructions
Beyond a single message, ChatGPT lets you set standing instructions that apply to every chat, sometimes called custom instructions. This is where you put things you would otherwise repeat: your name, your job, your preferred tone, the fact that you want answers in metric units or in Spanish. In the API, the same idea is the "system" message. Set it once and every prompt inherits it, so your individual prompts can stay short.
Six prompts you can copy
Adapt the parts in brackets. These combine a role, a clear task, context slots, and a format.
1. Welcome email
You are an email copywriter. Write a welcome email for new users of [product, a habit-tracking app]. Warm, plain language, under 150 words. Include one clear next step and a friendly sign-off. No exclamation marks.
2. Summarize a text
Summarize the text below in five bullets, each under 15 words, for someone who will not read the original. Add one final line: "Bottom line: ...". TEXT: [paste text]
3. Rewrite in a tone
Rewrite the message below to sound [calm and reassuring]. Keep the facts, keep it under [80] words, remove jargon. Return only the rewritten message. MESSAGE: [paste message]
4. Write a Python function
You are a Python developer. Write a function named slugify(text) that lowercases the input, replaces spaces with hyphens, and removes non-alphanumeric characters. Include a docstring and three example calls with their outputs as comments. Return only the code.
5. Make a study plan
Build a four-week study plan to learn [basic SQL]. I can study [five hours a week]. Return a table with columns Week, Focus, Practice Task, Hours. Assume no prior knowledge and keep each cell short.
6. Extract data to a table
From the text below, extract every person mentioned. Return a table with columns Name, Role, Company. If a field is missing, write "unknown". No text outside the table. TEXT: [paste text]
Put it together
The pattern behind all six is the same: a role, a specific task, any context the model needs, and the exact output format. Start there, then iterate. If you want the underlying principles in more depth, read prompt engineering basics, and for a broader walkthrough across AI tools see how to write AI prompts. When you want a prompt built for you, try the prompt generator.