Writing prompts that work: a practical guide for business teams

Two people in the same team use the same AI assistant. One gets a reply that’s ready to send after a light edit. The other gets three paragraphs of polite filler and decides AI is overrated. The difference is rarely the tool. It’s what they asked for, and what they gave it to work with.

Writing good prompts isn’t a technical skill. It’s the same skill as briefing a capable new colleague who knows nothing about your company. This guide covers the parts of a good prompt, gives you templates for everyday tasks, and explains when a prompt stops being enough.

Think of it as briefing a new colleague

Imagine a sharp temp starts on Monday. They write well and learn fast, but they’ve never seen your clients, your products or your tone of voice. Say “reply to this customer” and you’ll get something generic. Tell them who the customer is, what happened, what you can offer and how you usually write, and you’ll get something you can send.

A language model is in the same position every time you start a new chat. It knows a lot about the world and nothing about your situation, except what you put in the prompt.

The parts of a good prompt

Not every prompt needs all eight parts. A one-line question is fine for a one-line answer. But when the result disappoints, the missing piece is almost always on this list.

Anatomy of a good prompt

• Context Who you are, who it’s for, what happened
• Role Which kind of expert to write as
• Task One clear job, starting with a verb
• Source material The email, document or data to work from
• Format Length, structure, language, tone
• Constraints What it must not do or promise
• Examples One or two samples of a good result
• Questions first Let it ask what’s missing

Context and role

Say who you are, who the output is for, and why. “I run customer support at a building-materials distributor. Our customers are small construction firms.” A role (“write as an experienced accounts receivable clerk”) can set the vocabulary and tone, but context does most of the work. A role without context is just a costume.

Task

One clear job, starting with a verb: summarise, draft, extract, compare, classify, rewrite. If you need three things, list them as numbered steps or use three separate prompts.

Source material

Paste or attach what the model should work from: the customer’s email, the contract clause, the meeting transcript, the price list. Tell it to use only that material and to say so when something isn’t there. This one habit removes most invented answers, and our article on why AI makes things up explains why. Mark clearly where the material starts, for example with a line that says “EMAIL:” above it.

Format

Describe what the result should look like: “Three bullet points, under 80 words, in Slovak, formal.” Or: “A table with columns for supplier, invoice number, amount and due date.” If the output goes into another system, describe exactly the format that system needs.

Constraints

Say what it must not do: no discounts, no delivery dates we haven’t confirmed, no legal advice. Where you can, say what to do instead, because positive instructions work better than a list of don’ts: “If the customer asks for a discount, say our sales manager will call them.”

Examples

One or two examples of a good result teach tone and format faster than any description. Paste a reply your best support agent wrote and add “write in this style”. Remove personal details first.

Let it ask questions

End with: “Before you start, ask me up to three questions if anything important is missing.” It’s the cheapest way to find the gap in your brief.

Weak and strong: the same request twice

Here is the same job, a reply to a complaint about a late delivery, written two ways.

Two prompts for the same job

Weak prompt

• “Reply to this customer complaint.”
• No context about the order or the customer
• No rules on what you can offer
• Result: polite, generic, promises a refund you don’t give

→

Strong prompt

• Context: a 3-year customer, delivery 4 days late
• Source: the customer’s email and the delivery facts
• Rules: apologise once, give the new date, no compensation
• Format: formal, under 120 words, signed by Maria
• Result: a reply you can send after a quick read

The strong version in full:

I handle customer support at a building-materials distributor.
Our customers are small construction firms who order by email.

Task: draft a reply to the customer email below.

Facts:
- The order arrived 4 days late because our supplier delayed it.
- The customer has ordered from us for 3 years.
- The missing items will arrive on Thursday.

Rules:
- Apologise once, briefly. Don't name the supplier.
- Give the new delivery date.
- Don't offer a discount or compensation. If they ask,
  say our sales manager will call them.

Format: formal English, under 120 words, signed "Maria, Customer Support".

EMAIL:
[paste the customer's email here]

It takes two minutes longer to write. It saves the ten minutes you would otherwise spend fixing the answer, and the answer is safer.

What the strong prompt produces

Prompt

Draft a reply to this complaint. Facts: delivered 4 days late, missing items arrive Thursday. Rules: apologise once, no compensation; if asked, our sales manager will call. Formal, under 120 words, signed by Maria.

Answer

Dear Mr Horváth, I’m sorry your order arrived four days late. The remaining items will be delivered on Thursday, and we’ll confirm the time window the day before. If you’d like to discuss the delay, our sales manager will gladly call you. Thank you for your patience. Maria, Customer Support

Templates for everyday tasks

Copy these, fill in the brackets, and adjust them to your work.

Replying to an email

Context: [who you are, who the customer is, relevant history]
Task: draft a reply to the email below.
Facts you can use: [order status, dates, what you can offer]
Rules: [what not to promise; when to hand over to a person]
Format: [language], [formal or friendly], under [N] words, signed [name].
If anything important is missing, ask me before you write.

EMAIL:
[paste]

Summarising a document or a thread

Summarise the document below for [who], who needs to [decide what].
Give: 1) the three most important points, 2) all dates, amounts
and deadlines, 3) open questions or risks.
Use only the document. If something isn't stated, write "not stated".
Bullet points, under [N] words, in [language].

DOCUMENT:
[paste]

Extracting data into a table

From each invoice below, extract: supplier name, invoice number,
issue date, due date, total amount, currency, VAT amount.
Return a table with one row per invoice and exactly these columns.
If a field is not on the invoice, leave the cell empty. Do not guess.
Dates as YYYY-MM-DD. Amounts as numbers without currency symbols.

INVOICES:
[paste]

Meeting notes

Below is the transcript of a [type of meeting] with [participants and roles].
Write meeting notes with: decisions made, action items (owner and
deadline), open questions.
Include only what was actually said. Mark anything unclear as [unclear].
Under one page, in [language].

TRANSCRIPT:
[paste]

When a template works, save it in a shared prompt library so the whole team benefits, not just the person who wrote it. Our guide to getting your team to actually use AI covers how to run one.

Examples from different teams

The same parts work in any department. A few prompts we’d expect to see in an ordinary company, and why each one works:

• Accounting firm: “From the bank statement lines and the invoice list below, list every payment that has no matching invoice. Use only these two lists. Return a table with date, amount, counterparty and the reason it didn’t match.” It works because the model gets both sources, a narrow job and a fixed output. A person still reviews the table, and the totals are checked in the spreadsheet, not by the model.
• Recruitment agency: “Rewrite the job ad below for experienced warehouse supervisors. Keep every requirement and the salary range exactly as written. Remove internal jargon. Under 250 words.” The constraint protects the facts the client signed off on, while the model improves the wording.
• Private clinic: “Turn our preparation instructions for a fasting blood test into a friendly text message under 300 characters. Don’t add anything that isn’t in the instructions.” The last sentence matters most: patients must never receive advice nobody approved.
• Sales: “Here are my notes from today’s call with a new customer. Draft a follow-up email that confirms what we agreed, lists the next steps with dates, and asks about expected monthly volume. Friendly, under 150 words.” The rep’s notes are the source, so the email can’t promise what wasn’t discussed.

How to improve a prompt

Good prompts are rarely written in one go. Treat the first version as a draft and improve it in short rounds.

Improve it in rounds

1. 1

   Round 1

   Write a first version

   Use the parts above. Don’t aim for perfect.

2. 2

   Round 2

   Test it on 5–10 real cases

   Include an awkward one and one where information is missing.

3. 3

   Round 3

   Fix the pattern, not the case

   If answers keep running long, change the format rule instead of trimming one answer.

4. 4

   Round 4

   Save and share

   Put the version that works in the team’s library, with an example of a good result.

When an answer is wrong, ask why. Usually something was missing from the brief: a fact, a rule, an example. Add it to the prompt instead of fixing the answer by hand, and every later answer improves too. You can also ask the model directly: “What information would have helped you do this better?”

Common mistakes

• Too short. “Write a LinkedIn post about our new service” gets you the average of every LinkedIn post ever written.
• Too many jobs at once. Summarise, translate, draft a reply and assess the risk, all in one prompt, and each job gets done halfway.
• Asking for facts instead of giving them. “What’s our return policy?” without pasting the policy invites a guess.
• One chat for everything. A whole day of unrelated tasks in one conversation means earlier instructions fade and contexts get mixed. Start a fresh chat for each task.
• Pasting what shouldn’t go in. Personal data, client contracts or health information in a tool your company hasn’t approved. Follow your AI usage policy, or write one.
• Skipping the check. A good prompt reduces errors. It doesn’t remove them.

When a prompt isn’t enough

Prompts are great for one person doing varied work. They hit their limits when:

• someone pastes the same prompt and the same reference documents every day;
• results are copied from the chat into another system by hand;
• the output has to be consistent across many people, not just good for one;
• the volume is large, say hundreds of emails or documents a week;
• the model needs information that lives in your systems: stock, orders, customer history.

That’s the point to turn the prompt into a workflow. Your software fetches the data, sends the prompt, checks the result and puts it where it belongs, with a person approving where it matters. Our post on workflows, AI workflows and agents explains the options, and prompting, RAG or fine-tuning covers what to do when better instructions stop being the answer.

What to do next

1. Pick one task you do every week and write a prompt for it using the eight parts.
2. Test it on five real examples, and fix problems in the prompt, not in the answers.
3. Share the version that works with your team.
4. Note which tasks keep needing the same prompt and the same copy and paste. Those are your first candidates for a workflow.

When you reach that point, we can help turn a prompt that works for one person into a tool that works for the whole team. Get in touch.