How to write prompts that actually work: 7 principles

Most people talk to ChatGPT or Claude like it's a search box: a short phrase and then they wait for magic. The result is usually generic. A good prompt isn't longer for the sake of it; it's structured to give the model the context, role and format it needs to deliver something usable.

1. Give it a clear role

Start by telling it who it is. "Act as a senior copywriter specialized in product launches" triggers a different register and vocabulary than saying nothing. The role sets the bar for the answer.

2. Context before the task

The model knows nothing about your business. Give it the minimum context: who you're addressing, what tone you use, what you tried before and why it failed. Three lines of context beat ten lines of instructions.

3. Specify the output format

Do you want a table? A list of 5 items? A 120-word email? Ask for it explicitly. If you don't define the format, the model picks one for you — and rarely gets it right.

4. Show an example (few-shot)

A single example of the ideal output is worth more than any adjective. "I want it to sound like this: [example]" removes most of the iterations.

5. Set limits and constraints

Tell it what NOT to do: "no jargon", "max 3 sentences", "don't make up data". Constraints cut noise and raise precision.

6. Ask it to think before answering

For reasoning tasks, add "think step by step before giving me the final answer". The model exposes its logic and makes fewer mistakes.

7. Iterate, don't rewrite

If the first answer is on track but not there yet, don't start over. Refine what you have: "good, now make it more direct and drop the second paragraph".

A well-built prompt is the difference between using AI as a toy and using it as a work tool.

These seven principles get you 80% of the way. The other 20% is having battle-tested prompts ready for every recurring task — which is exactly what you'll find in our collections.