Training

A prompt framework for work: five steps to useful answers

A simple prompt framework for writing prompts that work in ChatGPT, Claude, Gemini and Copilot. It includes a guide to what to fix when an answer goes wrong and a full worked example from an accounting firm.

What a prompt is and why it decides the quality of the answer

When ChatGPT returns something bland, many people conclude that AI is no use for their job. More often the assistant has done what it could with a request that was too thin. An office manager at an accountancy practice who asks for “an email reminding a client about missing documents” gets back something polite and empty, because the request says nothing about which documents are missing or when they are due. A simple prompt framework turns that request into a draft you can send after two small edits.

A prompt is the instruction you type into an AI assistant such as ChatGPT, Claude, Gemini or Copilot. The assistant knows nothing about your company, your clients or your deadlines. It only knows what the prompt tells it, and it fills the rest with whatever is most likely, which is usually the most generic option.

Use is already common. Gallup reported in May 2026 that 30% of US employees use AI at work several times a week or more, and 15% use it daily. Using it often does not mean asking well, and asking well can be learned quickly with any of the usual assistants.

The prompt framework in 5 steps

Think of a prompt as the brief you would give a capable colleague in their first week. Each step answers a question that person would ask you, and together they give your prompt its structure.

  1. 01

    Context and role

    Say who you are, who the work is for and what role the assistant should take. The role is the position it should write from, for example “act as a sales negotiator at a Manchester estate agency used to first-time buyers”. The role sets the tone, and with the context in place the assistant no longer has to fill the gaps with guesses.

  2. 02

    Task and goal

    State what it should do with a clear verb, such as draft, summarise, compare or suggest three options. Then say what the result is for, for example booking a viewing, reassuring a customer or preparing a decision. One task per prompt works better than five at once.

  3. 03

    Material and examples

    Paste in what it should work from, such as the customer's email, your meeting notes or a few rows of a spreadsheet. If you have an example of a good result, include it. A past email that worked shows your house style better than any description. Remove personal or confidential details first.

  4. 04

    Format and constraints

    Describe the shape of the result and its limits. Maximum length, bullets or paragraphs, language, level of formality, words to avoid and facts it must not make up. A line such as “if you are missing information, ask me before you answer” prevents a lot of errors.

  5. 05

    Iterate and check

    The first answer is a draft. Ask for specific changes in the same conversation, such as “shorter”, “warmer tone” or “add the deadline”. Then check figures, names, dates and any legal references before you use the text. You remain responsible for the result.

Bad answer, missing ingredient: how to fix a prompt

When an answer is no use, one of the method's ingredients is almost always missing. Spot the symptom and fix only that part of the prompt. You don't have to start again.

  • Generic, could be for any company

    Context is missing. Add your sector, your town, the type of client and your relationship with them. Three concrete details change the answer more than three paragraphs of instructions.

  • Correct, but useless for what you wanted

    The goal is missing. Say what should happen once the reader has read it, for example that they send the documents or agree to a Saturday viewing.

  • Too long, or the wrong tone

    The format is missing. Set the length, the structure and the register. “Five lines at most, friendly but professional, no long sign-off” is usually enough.

  • Contains made-up facts

    Material or a constraint is missing. Paste in the real data and say it must not invent figures, names or references. Ask it to put anything it doesn't know in square brackets.

  • Doesn't sound like your company

    An example is missing. Paste in one of your own texts that worked and ask it to match the style. It is the quickest fix and the one people forget most.

Example

An accounting firm in Leeds chases its clients for VAT documents

Take a typical case: an accounting firm in Leeds with four account managers. Every quarter they send dozens of emails asking clients for documents missing from their VAT return. Some managers write them by hand. Others ask AI for “an email asking a client for some missing documents” and then rewrite most of it, because the assistant has no idea which documents are missing, when they are due or how well the firm knows the client.

The firm rewrites the prompt using the method. “I work at an accounting firm that keeps the books for small businesses. Act as the account manager for this client, a furniture shop we have worked with for four years. Draft an email asking for the supplier invoices for September and the bank statement for August, which we need by 14 October to file their July to September VAT return. We want them to send everything this week without the email sounding like a telling-off. Friendly but professional, under 120 words, with the documents in a bullet list. Do not mention any other documents or amounts. Here is one of our previous emails as a style reference.”

The second prompt covers the first four steps. The fifth comes afterwards. The manager reads the reply, asks for a tweak if needed and checks the dates before sending. Then the prompt is saved with variable fields such as [client], [documents] and [deadline], so colleagues can reuse it next quarter.

To see whether it helps, each manager notes for a month how many edits each draft needs before it goes out, and whether the client replies first time. If the edits go down, the prompt stays in the library. If not, the team looks at which symptom appears and fixes the missing ingredient.

Common prompt-writing mistakes

  • Asking for everything at once

    Summarising a contract, flagging risks, drafting an email and building a table in one prompt mixes four tasks. Split them into separate messages in the same conversation and check each result.

  • Taking the first answer

    Copying and pasting without reading is the quickest way to send a mistake to a customer. An assistant can state a figure or a legal reference with full confidence when it does not exist.

  • Pasting data that should stay inside

    Customer details, payroll or confidential contracts only go into tools your company has approved. If in doubt, anonymise the text or replace names with placeholders such as [client].

  • Hunting for magic phrases

    Prompt lists found online can help you start, though they know nothing about your business. Your own prompt, tested on your own cases, usually does much better than any trendy formula.

A shared prompt library and saved instructions for the team

When a prompt works, it usually stays in one person's chat history and everyone else starts from scratch. A shared library fixes that. A shared document is enough, organised by task or by department. For each prompt, note the task it handles, the assistant it was tested on, the text with variable fields in square brackets and a sample result.

Many assistants now let you save standing instructions or set up a project space with reference documents. Use that for the context that rarely changes, such as who your company is, how you address clients and which words you avoid. The prompt itself is then left for what changes each time, which is the task, the material and the goal.

Make one person responsible for reviewing the library every few months. They remove prompts nobody uses, add improvements from colleagues and check that none contains confidential data. New starters then begin with examples that already work in your company.

From one good prompt to a team habit

Pick a text your team writes every week, such as a reply to a quote request or a meeting summary. Write the prompt with the five steps, compare the result with what you would have written by hand and save the version that works. Repeat with a second task the following week.

The method sticks when people practise on their own cases. That is why Obinova's training includes a section on prompting, with exercises on emails, documents and reports similar to the ones your team handles. It covers ChatGPT, Claude, Gemini and Microsoft Copilot, and runs on a dedicated video platform or in person at your company. The prompts that come out of the exercises make a good start for your library.

Frequently asked questions

What is a prompt?

A prompt is the instruction you give an AI assistant such as ChatGPT, Claude, Gemini or Copilot. It can be a short question or a detailed brief with context, material and a format. The quality of the answer depends largely on what the prompt explains, because the assistant only knows about the task what you tell it.

How long should a good prompt be?

There is no set length. A simple question fits on one line. A piece of work, such as an email to a client or a report summary, usually needs five to ten sentences covering context, task, material and format. As a guide, include whatever you would tell a new colleague so they could get the job right first time.

Does the same prompt work in ChatGPT, Claude, Gemini and Copilot?

Generally, yes. The five steps work with all the usual assistants, because they all need context, a clear task and a format. Answers can vary from one tool to another. Test your important prompts in the assistant your team actually uses, and note in the library which one they were checked on.

Which prompt framework is best: this one, RTF or CO-STAR?

They cover much the same ground under different labels. RTF stands for role, task and format, and CO-STAR adds items such as audience and style. The five steps here add two things teams often skip: pasting in real material with an example of good output, and checking the answer. Pick one framework, use it on your own tasks for a few weeks and keep it if the drafts need fewer edits.

Is prompt training worth it for a team?

It usually is once several people use AI every day, each in their own way. A Harvard Business School and BCG study of 758 consultants found higher-quality, faster work with AI on suitable tasks, and worse results outside them. Training helps people tell the two apart and ask for what they need.

What should you never put in a prompt?

To be safe, no personal data about customers or staff, confidential contracts, internal figures or passwords, except in tools your company has approved under business terms. Before pasting a document, remove names and identifying details or replace them with placeholders such as [client]. Put these rules in a short usage policy that everyone on the team knows.

Train your team in prompting on its own cases

Book a free 30-minute video call. We will go through the texts your team writes every week and suggest how to practise prompting on them.

Free 30-minute diagnosis →
Free diagnosisWhatsApp

WhatsApp · Reply within 2 hours during business hours