How to Write Prompts for ChatGPT (GPT Style)

ChatGPT is forgiving with casual prompts, which is exactly why people plateau with it. The difference between a vague answer and a great one is almost always three things: a defined role, an ordered set of steps, and a locked output format. Get those right and the model stops guessing what you meant.

Open with a role

Start by telling the model who it is for this task. A role narrows the space of likely answers and sets the vocabulary, depth and tone before you have asked anything. Be specific: a seniority level and a domain work better than a generic title.

You are a senior backend engineer who reviews Node.js code
for security and performance. You are blunt and specific.
You cite the exact line and explain the fix, not just the problem.

Ask for steps, not just an answer

For multi-part work, tell the model to proceed step by step and number the steps. This both improves the reasoning and makes the answer easier to scan and to correct. If you only need the final result, ask it to do the steps internally and return only the conclusion.

Rewrite this product description. Work in this order:
1. List the three benefits a buyer cares about most.
2. Draft a one-line hook for each benefit.
3. Combine them into a 60-word description.
Show your work for steps 1 and 2, then the final description.

Control the format

Do not leave the output shape to chance. State whether you want a table, a JSON object, a bulleted list or prose, and give the exact fields. When you need machine-readable output, describe the schema and say to return only that, with no commentary.

Iterate instead of restarting

When the answer is close but off, do not rewrite the whole prompt. Tell the model what to keep and what to change. This is faster and keeps the parts that were already good.

• Keep the structure, but make the tone more casual.
• Same answer, but cut it to half the length.
• Good list. Now turn the third item into its own paragraph.

The GPT prompt skeleton

1. Role: who the model is for this task.
2. Task: what you want, in one clear sentence.
3. Steps: the order to work in, if it matters.
4. Format: the exact shape of the output.
5. Constraints: length, tone, what to avoid.

Treat the skeleton as a default, not a cage. Drop the steps for a simple question, add a schema for a data task. The habit that matters is being explicit about role, order and format every time it counts.