How to Write AI Prompts: Six Parts, Real Examples

Most bad AI answers are answers to bad questions. Type “write something about our new service” and you’ve given the model one sentence and asked it to guess the rest: audience, length, tone, purpose, the service itself. It guesses. The result reads like a brochure nobody asked for.

“Prompt engineering” sounds like a secret craft. Mostly, it’s what you’d tell a new hire on their first day. Anthropic puts it as a “golden rule” in its prompting best practices: show your prompt to a colleague who knows little about the task. If they’d be confused, the model will be too.

The six building blocks of a good prompt

Not every prompt needs all six. The more the output matters, the more of them you should supply.

1. MaterialLong documents and data first. Anthropic recommends putting them at the top, above the instructions.
2. TaskOne clear verb: write, shorten, compare, find, rate.
3. Purpose and contextWhat it's for, who it's for, what they already know.
4. FormatLength, structure, table or prose, tone.
5. ExamplesOne to three samples of the output you want.
6. Fallback ruleWhat to do when information is missing or the model is unsure.

The block people skip most is purpose. “Summarise this report” produces something different from “Summarise this report so I can decide at Monday’s board meeting whether to extend the budget.” Once the model knows the purpose, it knows what to keep and what to drop. Anthropic’s guide says explicitly that explaining the motivation behind an instruction improves results.

Before and after: three rewrites

Theory only goes so far. Here are three everyday prompts, each in a weak and a strong version.

First, a service description. The weak version: “Write a blurb for our new maintenance package.”

The strong version:

Write one paragraph for our website about a new IT maintenance package.
Audience: owners of small dental practices, no technical background.
Facts: £79 a month plus VAT; includes updates, backups and a support line
on weekdays 8am to 6pm; 4-hour response time.
Tone: plain, reassuring, no jargon.
Length: 80 to 100 words. No superlatives, no exclamation marks.

Second, a decision. Weak: “Should I hire a second employee?” Strong: give revenue, workload, the cost of the role and what worries you. Then ask outright for pushback: “Play devil’s advocate. Give me the three strongest reasons against.” Language models tend to agree with whoever is asking. If you want disagreement, you have to order it.

Third, shortening. Weak: “Make this shorter.” Strong: “Cut to 120 words maximum. Keep every number and deadline. Remove the introduction and repetition first.” With a priority order, the model doesn’t decide at random what goes.

Examples are the strongest lever

Anthropic calls examples one of the most reliable ways to steer format, tone and structure, and suggests three to five when consistency matters. For everyday work, one or two usually do the job.

Here’s how that looks when you need product descriptions in a consistent house style:

Write a product description in the style of the examples.

<example>
Linen apron, natural. Heavy half-linen, adjustable neck strap,
two deep pockets. Machine washable at 40°C. 68 x 85 cm.
</example>
<example>
Oak bread basket. Solid wood, oiled, with linen cloth.
24 cm diameter, 9 cm high.
</example>

Product: chef's knife, 20 cm Damascus steel blade, olive wood handle,
hand-sharpened, not dishwasher safe.

Tags such as <example> aren’t compulsory, but they help. Anthropic recommends this kind of markup so the model doesn’t mix up instructions, material and examples. It works with other assistants too.

Say what to do, not what to avoid

Long lists of bans are a common mistake: no clichés, not too long, no jargon, not too formal. The guide advises describing what you want instead. “Reply in short, direct sentences and address the reader as ‘you’” works more reliably than “don’t be so stiff.”

Bans still have a place for very specific phrases. “Don’t use ‘I hope this email finds you well’” is unambiguous. “Don’t sound like AI” isn’t.

Break big jobs into steps

Prompts that ask for everything at once rarely deliver. “Research the market, write a strategy and turn it into slides” doesn’t overwhelm the model so much as the structure. Chaining works better: research first, you check it, then a strategy based on the checked research, then the slides.

Anthropic lists prompt chaining as a technique of its own. The practical benefit is obvious. You see after each step whether the direction is right and correct course early. An error in step one that only surfaces in step four costs you everything in between.

Here’s how that looks for a small agency:

1. “Summarise the five client emails attached: which requests come up more than once?”
2. You strike two points that rest on misunderstandings.
3. “Based on the remaining three requests, outline a proposal.”
4. “Write out section 2 of the outline, 150 words, tone as in the example.”

Each step takes a minute. Together they produce a draft you can stand behind.

Mistakes almost everyone makes

Read colleagues’ prompts and the same patterns keep appearing:

• The prompt states the task but no facts, so the model fills the gap with generic filler.
• Several tasks are crammed into one sentence with no order.
• No length is given, and models then write too much rather than too little.
• The format is unclear: table, bullet points or a letter?
• There’s no instruction for what to do when information is missing.
• After a poor answer, the same prompt gets sent again instead of being improved.

That last one deserves a sentence more. When an answer misses, say exactly what’s wrong: “Too formal, and the deadline in paragraph two is wrong; it’s the 15th.” That gets you further than starting over.

Let the model ask first

Here’s a trick that rarely appears in guides and saves a lot of back-and-forth. Ask the model to put questions to you before doing the work.

I need a plan for a half-day client workshop.
Before you start, ask me up to five questions you'd need answered
to do this well. Wait for my answers.

You find out quickly what context is missing and skip three rounds of corrections. It’s especially useful when you’re not yet sure what you want yourself.

Allow “I don’t know”

Language models would rather answer wrongly than not at all. Researchers at OpenAI have argued that this comes from how models are trained and evaluated: guessing gets rewarded, admitting uncertainty doesn’t. Our piece on AI hallucinations covers the details.

One line in the prompt helps: “If the information isn’t in the text or you’re unsure, say so instead of guessing.” With documents, add: “Cite the location in the document for every claim.” How that works with long files is shown in our guide to summarising PDFs with AI.

Five templates for everyday use

• Email: Write an email to [recipient]. Situation: [facts]. Goal: [outcome]. Tone: [x]. Max [n] words.
• Summary: Summarise so I can decide [decision]. Cite the source for each point. Mark gaps as "not found".
• Challenge: Give the three strongest objections to my plan. Be specific, not polite.
• Shorten: Cut to [n] words. Keep all numbers. Drop intro and repetition first.
• Questions first: Before you start, ask me up to five questions.

If you use templates like these daily, stop pasting them in. Custom assistants with saved instructions handle that; see our walkthrough on building your own AI assistant. For inbox-specific prompts, our AI email assistant guide goes further.

When the prompt isn’t the problem

Sometimes wording isn’t the issue. If the model needs information that only exists in your internal systems, no prompt will fix that. The assistant needs secure access to that data, with permissions handled properly. That’s a development job, the sort a studio like OnlineBuilders takes on.

For everything else: take the prompt you typed most often this week and add the two building blocks it’s missing. Usually that’s purpose and an example. Other tool guides live under tools.

Questions people ask

Should I tell the AI it's an expert in something?

Giving it a role helps set tone and technical depth, and Anthropic lists role prompting as a technique. Context and purpose matter more, though. A role with no facts behind it doesn't get you far.

How long should a prompt be?

As long as it needs to be. Two hundred words of context often beat ten words of instruction. What you can cut is repetition and pleasantries.

Are there prompts that always work?

No. Templates are a starting point, but every model responds a little differently and models change with updates. Re-test your most important prompts after major releases.

Does saying please make a difference?

There's no solid evidence that politeness changes output quality in a meaningful way. Clarity does. Write the way you'd brief a capable colleague.