How to write AI prompts that actually work

12 September 2025 · 6 min read

If you have ever asked an AI for help and got back something vague, generic or subtly wrong, the model is usually not the problem. The prompt is. A large language model can only work with what you give it, and most prompts give it far too little to go on.

The good news is that better prompting is a skill, not a talent. A handful of habits will lift the quality of almost everything you get back. Here is the structure we use, and the one our Prompt Optimizer builds for you automatically.

Give it a role

Start by telling the model who it should be. 'You are a senior financial-services compliance analyst' produces a very different answer to no framing at all. A role sets the vocabulary, the depth and the assumptions the model brings to the task.

Supply the context

The model cannot see what you can see. Tell it who the output is for, what has happened so far, and any constraints that matter. A paragraph of context is often worth more than a page of instructions.

State the task plainly

Be specific about the one thing you want. 'Summarise this in five bullet points a busy director can scan' beats 'help me with this'. If there are several things, number them, and consider running them as separate prompts.

Set the constraints

Length, tone, what to avoid, the audience, the reading level. Constraints are not limits on the model, they are how you get the specific answer you actually need instead of a plausible average of everything.

Describe the output format

Ask for the exact shape you want to receive: a table, a list, a short email, JSON, headings. Telling the model how to lay the answer out saves you the reformatting and makes the result usable straight away.

Show an example

If you have a sample of the style or format you want, paste it in and say more like this. One good example teaches the model faster than three paragraphs of description, especially for tone and structure.

Iterate

Your first prompt is a draft. Read the response, notice where it missed, and add the missing context or constraint. Two or three quick rounds will get you further than trying to write the perfect prompt in one go.

Put those habits together and you have a prompt that gives a model a real chance of being useful. If you would rather not assemble it by hand every time, our Prompt Optimizer does the work for you, and goes a step further by tailoring the result to the specific AI model you are using.