How to Write Better AI Prompts

Most poor AI output comes from vague instructions rather than model limitations. A request that a new colleague could not act on without asking questions will not produce good results either.

These are the patterns that reliably help.

1. Say who it is for and what it is for

Compare “write about climate change” with “write three paragraphs explaining why sea levels rise, for a curious 14 year old, avoiding jargon”.

The second specifies audience, length, scope and tone. Each one removes a decision the model would otherwise guess at.

2. Provide the source material

This is the highest value habit and the most underused.

A model asked to recall facts from memory may invent them. The same model asked to summarise, rewrite or extract from text you paste in is far more reliable, because the information is in front of it rather than reconstructed.

If you have the document, the notes or the data, include them.

3. Show an example of what good looks like

Describing a format is weaker than demonstrating it. If you want output in a particular shape, paste one example and say “match this structure”.

Two or three examples work better than a paragraph of description, particularly for tone, which is difficult to specify in words.

4. Set constraints

Constraints are easier for a model to follow than preferences.

• “Under 150 words” rather than “keep it brief”
• “Use only the information I provided” rather than “be accurate”
• “No bullet points” rather than “write naturally”
• “Return JSON with keys name and date” rather than “give me structured data”

Negative constraints work too, and are worth stating when you keep getting something you do not want.

5. Ask for reasoning on hard problems

For multi step problems, asking the model to work through its reasoning before giving an answer improves accuracy, because each step conditions the next.

For simple requests this adds length without benefit.

6. Iterate instead of restarting

Treat the first output as a draft. Saying “shorter, and cut the introduction” is faster and usually better than rewriting the prompt from scratch, because the model keeps the useful parts.

Restarting is the right move only when the first attempt misunderstood the task entirely.

7. Split large tasks

A request to research, outline, write and edit in one go produces mediocre results at every stage. Doing them as separate steps, checking each, produces better output and makes errors easier to spot.

Habits that do not help

• Politeness formulas. Please and thank you are fine, and they do not improve quality.
• Threats or rewards. Offering a tip or claiming someone’s job depends on it has no reliable effect.
• “You are a world class expert.” Specifying a role can help set register, but superlatives add nothing.
• Very long prompts. Length is not the goal. Relevant detail is.
• Asking it to confirm accuracy. A model saying it is confident is not evidence that it is correct.

Always verify facts

No prompt technique removes the need to check claims, numbers, names, quotes and citations. Prompting changes how useful the draft is, not whether it can be trusted without review.

Why that limitation exists is explained in our guide to how generative AI works.

A template that works

Context: what the situation is
Task: what you want produced
Audience: who will read it
Format: length, structure, style
Material: any text or data to use
Constraints: what to avoid

Not every request needs all six. Most poor results are missing at least three.

Frequently asked questions

Does prompt wording really matter that much?

For simple questions, little. For anything with a specific audience, format or source material, it matters a great deal.

Should I use one long conversation or start fresh?

Keep related work in one conversation so context carries over. Start fresh for an unrelated task, since stale context can pull answers in odd directions.

Why does the same prompt give different answers?

Generation involves randomness. Asking for several options at once is more efficient than asking repeatedly.

Can I ask it to write in my style?

Yes, and pasting a few samples of your own writing works far better than describing the style.

Is there a best prompt format?

No universal one. Clear structure helps, and the specific layout matters less than including the right information.

How do I stop it being so wordy?

Set a hard limit and say what to leave out. “Under 100 words, no introduction or summary” is more effective than “be concise”.