A good prompt is a recipe card. The kitchen changes. ChatGPT, Claude, and Gemini each have their own habits. The card stays the same, so the dish comes out closer to what you asked for.
You do not need a long prompt. You need the same seven parts, filled in plainly. Copy the lines below, replace the brackets, and send it.
1. Role: who should answer
The same question sounds different from a marketer, an engineer, or a teacher. Say which one you want.
You are a [ROLE] with expertise in [DOMAIN].
A few roles that change the writing immediately:
- Act as a senior software architect
- Act as a financial analyst
- Act as a marketing copywriter
- Act as a teacher explaining this to a beginner
2. Task: the one job
Use a verb, and name the thing. Create, analyze, summarize, compare, generate, design, explain, or improve.
Your task is to [SPECIFIC ACTION].
Filled in:
Your task is to create a product description for a new smartwatch.
3. Context: the situation
This is the background the model cannot guess. Audience, field, goal, and any limit that would change the answer.
Context:
- Audience: [WHO]
- Industry: [FIELD]
- Goal: [OBJECTIVE]
- Constraints: [LIMITS]
Filled in:
Context:
- Audience: Busy professionals
- Industry: Health technology
- Goal: Increase product sales
- Tone: Professional and trustworthy
4. Input: the material
Paste the thing the model should work from. A paragraph, a piece of code, a product list, a set of requirements. If you leave this out, it will invent the facts.
Input:
[DATA HERE]
5. Output format: what the answer should look like
This part does more work than people expect. Ask for the shape of the reply before you ask for the words.
Useful shapes: bullet points, a table, JSON, a step-by-step guide, an email, a short report, a checklist.
Output format:
- Use bullet points
- Maximum 200 words
- Include headings
6. Constraints: the rules
Say what to leave out, and how long it should be.
Constraints:
- Keep it under 150 words
- Use simple English
- No technical jargon
Other rules worth writing down: the tone, how much detail, and words you do not want.
7. Example: optional, and worth it
One sample of a good answer teaches the shape faster than another paragraph of rules. Skip it when you are in a hurry. Add it when the first reply keeps missing.
Example:
[A short sample of the answer you want]
The card, ready to copy
Put the seven parts together. Delete any section you do not need. Do not delete the task, the output format, or the constraints.
Role:
You are a [ROLE].
Task:
Your task is to [ACTION].
Context:
[BACKGROUND]
Input:
[DATA]
Output format:
[HOW THE ANSWER SHOULD LOOK]
Constraints:
[RULES]
Example:
[A SHORT SAMPLE]
Follow one example all the way through
Here is the same card filled in for a product. You can change the product and keep the shape.
Role:
You are a senior marketing copywriter.
Task:
Write a product description.
Context:
The product is a portable blender for travelers.
Input:
- USB rechargeable
- Lightweight
- Blends in 15 seconds
- 500ml capacity
Output format:
- A headline
- 3 bullet benefits
- A call to action
Constraints:
- Under 100 words
- Persuasive tone
Send that. If the reply is close but not right, change one line, usually the output format or the constraints, and send it again. Do not rewrite the whole card.
The five-line version, when you are in a hurry
Act as: [ROLE]
Do: [TASK]
For: [AUDIENCE]
Using: [INPUT]
Format: [OUTPUT]
A workout plan, written this way:
Act as: Fitness coach
Do: Create a workout plan
For: Beginners
Using: Bodyweight exercises
Format: Weekly schedule
Four lines you can add later
Use these only when the basic card is not enough.
- Ask it to show the steps: “Explain your reasoning step by step.”
- Ask it to slow down: “Think carefully before answering.”
- Ask for a second pass: “Give a first draft, then an improved version, then the final version.”
- Name the voice: “Write in a friendly, professional tone.”
The one rule that matters
Clarity beats cleverness. A short instruction that says what finished looks like will beat a long instruction that only sounds thorough.
Tell the model what a good answer is. Then get out of the way.