How to Write Better AI Prompts: A Practical Guide With Examples

Most people type a request into an AI assistant the way they would type into a search box: short, vague, and hoping for the best. Then they judge the tool by the mediocre answer. The fix is not clever tricks or magic phrases. It is giving the model the same things you would give a capable new colleague: context, a clear task, constraints and an example of what good looks like. This guide shows you how, with prompts you can copy.

The five-part prompt

Every strong prompt we write contains some version of these five parts. You will not always need all of them, but when an output disappoints, one of them is usually missing.

1. Role and context. Who is the model acting as, and what is the situation? “You are editing a newsletter for small-business owners who are not technical.”
2. Task. The single thing you want done, stated as a verb. “Rewrite the draft below.”
3. Constraints. Length, format, tone, what to avoid. “Under 300 words. Plain English. No bullet points. Do not add new claims.”
4. Example or reference. A sample of the output style, or the source material. “Here is a paragraph in the voice I want: …”
5. Definition of done. How you will judge the result. “A reader should be able to explain the change to a colleague in one sentence.”

Here is the whole thing assembled:

You are editing a weekly newsletter for small-business owners who are not technical.
Rewrite the draft below so it is under 300 words, in plain English, with no bullet points.
Do not add claims that are not in the draft.
Match the voice of this sample paragraph: [paste sample]
Done means a reader could explain the main change to a colleague in one sentence.

Draft:
[paste draft]

That prompt takes 30 seconds longer to write than “make this better” and produces a result you can actually use.

Give context before the task

Models answer the question you asked, not the one you meant. If you ask “is this a good headline?” with no context, you get generic advice. If you say “this is for a LinkedIn post aimed at CFOs, the goal is replies not clicks”, you get a useful critique. Background, audience and goal should come before the instruction, because the model reads top to bottom and uses what it has seen so far to interpret what comes next.

State the format explicitly

If you want a table, ask for a table and name the columns. If you want prose, say “in prose, no lists”. If you want JSON, give the exact shape. Assistants default to headers and bullets because they are safe; you have to opt out. In our Claude review we noted that Claude follows negative format instructions particularly well, but all the major assistants respect them if they are explicit.

Show, do not just tell

One good example beats three paragraphs of description. If you want emails in your voice, paste two emails you actually sent. If you want a report structured a certain way, paste a previous report. This is the single highest-leverage technique in the guide, and the least used.

Ask for the reasoning first on hard problems

For analysis, decisions and anything with numbers, ask the model to work through the problem before giving an answer: “Think through the trade-offs step by step, then give a recommendation.” Modern reasoning models do a version of this automatically, but asking explicitly still improves results on ambiguous tasks and makes the answer easier to check.

Iterate instead of restarting

Your first output is a draft. Instead of rewriting the whole prompt, give feedback the way you would to a person: “Good, but the second section is too long and the tone drifted formal. Tighten it and keep the earlier voice.” Assistants are very good at targeted revision, and a conversation of three short corrections usually beats one enormous prompt.

Set up persistent instructions

Anything you find yourself typing every time belongs in a persistent setting. ChatGPT has custom instructions and Projects; Claude has Projects and style settings. Put your audience, voice rules and standard constraints there once. For a team, a shared Project with a style guide and reference documents is the cheapest productivity win available. See how we use these in our ChatGPT review.

A style instruction that fixes most “AI-sounding” text

Paste this into your custom instructions and adapt it:

Write in a natural, direct voice. Vary sentence length. Prefer prose to lists unless I ask for a list.
No em-dashes. Avoid filler openers ("In today's fast-paced world"), hedging ("it is important to note"),
and summarising closers ("In conclusion"). Make a point rather than presenting balanced options
unless I ask for options. Use concrete examples over abstractions.

Mistakes that produce bad output

• Asking for several things at once. “Summarise this, suggest a title, and draft a tweet” gets three mediocre results. Ask one at a time, or ask for them in a numbered sequence with the format for each.
• No audience. The model cannot pitch the level of detail without knowing who reads it.
• Trusting facts without sources. For anything with a date, price or statistic, turn on web search and check the citation. Our research comparison covers which tools do this best.
• Over-prompting. A two-page “mega prompt” full of rules the model has to juggle often performs worse than a clear five-part prompt plus iteration.
• Not reading the output. Obvious, and still the most common.

Prompts for specific jobs

Summarising a long document:

Summarise the attached report for an executive who has five minutes.
Lead with the three decisions it asks them to make. Then one paragraph of context per decision.
Quote figures exactly as they appear. If something important is unclear in the document, say so.

Debugging code:

Here is a function and the error it produces. Explain the likely cause in two sentences,
then give a corrected version with a comment on the changed line only. Do not refactor anything else.

Getting honest feedback:

Critique this plan as a sceptical investor would. List the three weakest assumptions,
explain why each could fail, and suggest what evidence would change your mind. Do not soften the feedback.

Next steps

Pick one task you do every week, write a five-part prompt for it, and save it as a persistent instruction. That single change will do more for your results than any model upgrade. When you are ready to go further, our guide to AI coding assistants applies the same ideas to software work.