Prompting for business teams: the six habits that produce usable output, and templates to copy.

By Zain Masroor · 13 October 2026 · 11 min read

Most poor AI output is a poor request. Six habits fix it: say who the assistant is writing as and who will read it; give it the actual material rather than describing it; state the format, length and tone; say what to leave out; ask it to list what it was not given and what it assumed; and refine the draft in the same conversation rather than starting again. The tasks UK firms use most, a client email, a document summary, a proposal section, a spreadsheet question and a policy rewrite, each have a template below that can be pasted as it stands. The step that makes the habits stick is putting the templates into shared projects or custom GPTs on a business plan, so the tool already knows the firm’s tone and nobody starts from a blank box.

Most errorsA poor request, not a poor model

Six habitsRole, material, format, exclusions, uncertainties, iterate

Make it stickA shared template library

The six habits

None of these needs training in prompt engineering. They are the same things you would tell a capable new colleague.

01Role and reader. “You are writing as our practice manager to a client who is not technical.” The model writes differently for different readers, and it cannot guess.

02Material, not description. Paste the email thread, the document, the figures. “Summarise our contract with X” without the contract produces a summary of contracts in general, which is the most common way people get confident nonsense.

03Format, length, tone. “Three short paragraphs, plain English, no bullet points, no exclamation marks, British spelling.” Say it every time or put it in the shared instructions.

04Exclusions. “Do not mention pricing. Do not speculate about their reasons.” What to leave out is as important as what to include, and the model will not infer it.

05Uncertainties. End with “List anything you were not given, anything you assumed, and any statement I should check.” The answer is usually honest and often the most useful part of the output.

06Iterate. Reply to the draft with what is wrong rather than starting a new conversation. The second draft is where the quality is, and the model keeps the context.

How to check AI output before you rely on it →Prompting training for business teams, from £1,200 →

Templates for the five common tasks

Paste as they stand, replacing the bracketed items. Each ends with the uncertainties request, because that is the habit most people drop.

01Client email. “You are writing as [name, role] at [firm] to [client], who is [context]. Here is the thread: [paste]. Reply to confirm [what], propose [what], and ask [what]. Three short paragraphs, plain English, warm but not casual, British spelling, no bullet points. Do not mention [exclusions]. Then list anything you assumed.”

02Document summary. “Here is [document type]: [paste]. Summarise it for [reader] in [length], covering [the points that matter to us]. Quote the exact wording for any obligation, date or figure. Where the document does not say, write ‘not stated’. Then list anything you should check against the original.”

03Proposal section. “You are writing the [section] of a proposal from [firm] to [prospect] for [service]. Our position: [paste notes]. Their brief: [paste]. Write [length] in our tone: [describe or attach examples]. Include [what], exclude [what]. Do not invent figures, references or client names. Then list any claims that need a source.”

04Spreadsheet question. “Here is a table: [paste or attach]. Columns mean [definitions]. Answer: [question]. Show the calculation. If the data cannot answer it, say so and say what is missing. Do not estimate figures that are not in the table.”

05Policy rewrite. “Here is our current [policy]: [paste]. Rewrite it for [reader] in plain English at [length], keeping every obligation and deadline exactly, in our house style: [describe]. Flag any clause that seems out of date or contradicts another. Do not add obligations that are not in the original.”

From templates to a shared library

The habits stick when the templates live in the tool rather than in a document nobody opens. On Claude Team, a project holds the firm’s instructions, tone examples and templates, and every conversation started in it inherits them. On ChatGPT Business, a custom GPT does the same. Build two or three for the teams with the most to gain, with the firm’s real proposal, letter and policy as examples, and the first thing a new user sees is the tool already writing like the firm. This is the single change that moves adoption most, and it is the first thing we build in a rollout.

Claude Team plan for a UK business →Claude or ChatGPT for a UK business →AI change management: getting the tools used →

What not to paste

The templates assume a business plan, where content is not used for training and a data processing agreement applies. On a personal account, the same paste puts a client’s material into a third party’s training data. The policy rule is simple: business plan, and never special-category data, card or bank details, or anything a client marked confidential without their written agreement. Our GDPR guides and policy template set out the position; the training session for each team shows three real examples of what not to paste, which is what makes the rule stick.

AI acceptable-use policy template →Is Claude GDPR compliant? A UK guide →Is ChatGPT GDPR compliant? A UK guide →

When the request is not the problem

Sometimes the output is poor because the task needs something the assistant cannot do from a prompt: answer from your own documents reliably, which is a knowledge assistant; act across your systems, which is an agent; or handle volume unattended, which is a build. The tell is that no amount of rewording fixes it. Our guides on knowledge assistants and agents set out when that is the case and what each costs.

A knowledge assistant on your own documents →What an AI agent is, and when an SME needs one →AI training for staff: UK cost →

Common questions

How do I write a good prompt for business use?

Give the role and the reader, paste the actual material, state the format, length and tone, say what to exclude, ask for what was assumed or not given, and refine the draft in the same conversation.

Why does the AI give generic answers?

Usually because it was given a description instead of the material. “Summarise our contract” without the contract produces a summary of contracts in general. Paste the document.

Do staff need prompt engineering training?

They need the six habits, shown on their own work in a ninety-minute session, and a shared template library so they do not start from a blank box. Not a course.

What is a shared project or custom GPT?

A workspace inside Claude Team or ChatGPT Business that holds the firm’s instructions, tone examples and templates, so every conversation started in it inherits them. It is the single change that moves adoption most.

How do I stop the AI inventing figures or references?

Tell it not to, give it the material, ask it to mark anything not stated as “not stated”, and ask for a list of claims to check. Then check them; the checks are structural, not optional.

Is it safe to paste client documents into an AI?

On a business plan, with a policy, for ordinary business data, yes. Never on a personal account, and never special-category data, payment details or material a client marked confidential without their agreement.

Want the library built on your own documents?

We build the shared projects on your real templates and tone and run a session per team on the six habits, at a fixed price, delivered within three weeks.

Start a conversation →