Short answer
A good work prompt reads like a short task note for a new colleague who knows nothing about your context: role and context, task, material, answer format and, where useful, an example. The answer is a draft, checked against its source. Personal data and passwords stay out. A prompt that works is saved in a shared place.
Tuesday’s meeting is over, your director could not attend, and she wants “a quick summary”. You paste the notes into an AI tool and type “summarise”. The result reads well but is generic: nobody has an action, and one date has drifted to a day you never wrote. This is where the question of how to write prompts for work starts. The tool does not know your job, your reader or what the notes should look like, so unless you write it down, it guesses. This article covers the parts of a prompt and a before and after on an everyday office task. Then it shows how to check the answer, what you should never paste in, and how a team keeps a prompt that works.
A prompt is the whole instruction and context given to a language model. We cover quality control for published text in a separate piece: AI content production. Building an assistant that answers customers is a different job too: how to set up an AI assistant. Here we look at the work you type into a chat window at your own desk.
The five parts of a prompt
Anthropic, OpenAI and Google use different names for the parts but list much the same things. Anthropic’s guide compares the model to a new employee who lacks context on your norms and workflows, and offers a test: show the prompt to a colleague who knows nothing about the topic. If they would be confused, the model will be too. OpenAI’s documentation says a developer message usually has identity, instructions, examples and context sections. Google’s Gemini guide treats clear instructions, constraints, response format, examples and context as separate headings. The table below turns these into office work; that translation is our own summary.
| Part | What you write | What the documentation says |
|---|---|---|
| 1. Role and context | Who the tool acts as, who will read the text, what it is for | Anthropic: even a one-sentence role makes a difference; giving the reason behind an instruction explains the goal |
| 2. Task | One job, a clear verb and limits: what to do, what to leave alone | Google: you can state what to do and what not to do |
| 3. Material | The text to work on: notes, an email, a table | Anthropic: put long documents first and the question last; OpenAI: separate sections with Markdown or XML tags |
| 4. Answer format | Length, headings, table or list, tone | Google: you can ask for a table, bulleted list or paragraph |
| 5. Example | One or two outputs you liked from before | Anthropic: 3 to 5 examples; Google: test the number, too many and the model over-fits to them |
The documents do not fully agree on the number of examples. Anthropic suggests three to five, while Google says to experiment, because too many can make the model over-fit to them. We suggest starting with one in office work and adding more if the output still drifts. Google also asks that all examples share the same format.
Not every prompt needs all five parts. For a routine translation, material and format are enough. Use the parts to write down whatever the tool cannot guess.
Before and after: an email from meeting notes
First, as most of us write it:
Summarise these meeting notes.
Everything the tool does not know is left open: who the summary is for, how long it should be, what shape it should take, and what happens if something is missing from the notes. The version below does the same job with five parts. Note that we have reduced names in the notes to initials; the privacy section explains why.
You are my office assistant, preparing draft emails after meetings.
Context: The notes below are from the marketing team’s weekly meeting. The email will be read by our director, who could not attend. She knows the subject but not the details, and wants to see the decisions and who does what at a glance.
Task: Produce a summary of decisions and actions from the notes. Do not add any decision, date or person that is not in the notes. If a piece of information is not in the notes, write "not in notes" in that cell.
Format: At most 120 words. A two-sentence summary first, then a table with the columns Action | Owner | Date. Write the owner as the initial used in the notes. Tone: short and formal.
Example row: Newsletter draft | A. | Thursday
<notes>
[Paste the meeting notes here]
</notes>
Every line in this version has a job. The role frames the tool. The context sets the level of detail, because it says who the summary goes to. The rule against adding anything and the “not in notes” output apply the method on Anthropic’s page on reducing hallucinations: according to that page, explicitly letting the model say it is unsure can drastically reduce false information. The tags separate the material from the instructions. OpenAI and Google also suggest separating sections with Markdown headings or XML-style tags, and Google asks you to pick one format and use it consistently.
How to check the answer
The answer is a draft. Anthropic’s page on reducing hallucinations describes how models can produce text that is factually incorrect, so the reader before sending is you. This is the order we use:
- Match every item to its source. Find each action in the notes. A line you cannot find is invented. The same page suggests asking the model to find a supporting quote for each claim and to retract any claim it cannot support. You can make this easier by adding “quote the line from the notes for each item” to the prompt.
- Read numbers, dates and names separately. Fluent text does not show they are right. Check dates against a calendar and names against the team list.
- Look for what is missing. Is there a decision in the notes that the summary leaves out? Read the table again, starting from the end of the notes.
- Ask for a self-check too. Anthropic’s prompting guide suggests asking the model to verify its answer against stated criteria before finishing, and says this catches errors, especially in coding and maths. In office text it is a second pair of eyes, not a replacement for yours.
- Have a second person read anything that leaves the building. If you are not the one sending to a customer, a manager or the public, write the approval step down. The same idea applies in automations: human-in-the-loop leaves the decision to a person at the steps where a mistake is costly.
If something is missing, do not start patching the answer by hand. Find which part of the prompt left a gap, fix the prompt and run it again. Google also notes that prompt design can take a few iterations and that rewording the same request can change the response.
What never goes into a public tool
Question 14 of the KVKK guide on generative AI (published in Turkish) collects recommendations on protecting personal data when using these tools. It says to avoid sharing data that can identify a person, such as name, address, telephone number and ID details, and not to share third parties’ personal data either. It recommends anonymised and generalised wording where possible, with a general description in place of names, dates or locations. It also says to avoid sharing health data, financial information and information about legal proceedings. Elsewhere the guide points to the risk that personal data and corporate-level sensitive information shared in prompts could be exposed.
The guide does not list passwords and access keys separately; that rule is ours. The table below shows what to swap before you paste.
| Do not paste | Use instead |
|---|---|
| A customer’s name, phone, address or ID details | Codes such as “Customer A” and “Customer B” |
| Employee or candidate details, salary, health notes | A role name (“sales representative”) and a general description of the event |
| A password, API key, access code or bank details | Nothing: these never go into a prompt |
| An unsigned contract, an unannounced plan, a customer list | A sample text that shows the structure, or a shortened version without names |
| A whole email thread (signatures, phone numbers, extra details) | Only the paragraph you need |
One practical note: read the provider’s privacy policy and settings to see what data the tool collects and how long it keeps it. The KVKK guide also recommends reviewing privacy notices and settings. We would add that you should check whether input is used to train the model. If your company has a business plan, the terms may differ, but the rule on what to paste should still be written down.
Saving a prompt the team will reuse
A prompt that works is no use to the team while it sits in one person’s chat history. These are the steps to save it:
- Mark the parts that change. Put [SQUARE BRACKETS] around what changes each time, such as the meeting notes, the audience or the date, and keep the rest fixed.
- Put it somewhere shared. A shared sheet or document is enough. Columns: job name, prompt, owner, which tool it was tried in, date last tested, and a sample of a good output.
- Write down an owner and a version. One person makes changes and notes the date and the reason. Do not delete the old version.
- Retest when the tool or model changes. OpenAI’s documentation says even different snapshots in the same model family can give different results, and recommends building tests that measure prompt behaviour. In an office, the equivalent is to try each library prompt on one sample job whenever the tool updates.
- Put the privacy rule in the library. Next to every prompt, add the note “no personal data goes into this field”.
Fixed rules such as brand voice are better prepared once and attached than retyped into every prompt; see how to write a brand voice and style guide. If prompts will run inside an automation, this stage becomes a process question. Examples are in business process automation examples, and our approach is on the automation and AI service page.
Common mistakes
- Asking for five jobs in one message. Google suggests splitting complex tasks into parts, making each step its own prompt and feeding one output into the next. Summary first, then the email, then the tone fix.
- Not saying who reads it. The same notes are written differently for a director, a customer and the team. Who will read it should be the first line of the context.
- Having the example copied exactly. Anthropic wants examples to differ from each other so that the model does not pick up an unintended pattern. If you give one example, expect outputs that resemble it closely.
- Sending the first answer. Fluency does not show it is correct. The checklist above takes two minutes.
- Using a ready-made prompt list as it is. A template from the internet does not know your context. Use it as a skeleton and fill in the role, audience and criteria for your own job.
- Burying a secret in a prompt. If a password has been typed in even once, change it.
Common questions
How long should a good prompt be?
There is no fixed length. The test is whether a colleague who knows nothing about the job could do it from the prompt alone, which is what Anthropic suggests. A few lines can be enough for a routine job, and a meeting email might take fifteen. Putting long material (a report, an email thread) at the start and your question at the end is advice from Anthropic, and from Google’s Gemini 3 section.
Is giving the tool a role really necessary?
It is not required, but it is cheap. Anthropic says even a one-sentence role in the system prompt focuses the model’s behaviour and tone. A role is a short way of choosing a point of view and a tone, not a claim to expertise. Writing the reader and the purpose next to it often helps more.
Is pasting company data forbidden?
We are not saying there is a blanket ban; the issue is which data goes to which tool. Where personal data is involved, the KVKK guide’s approach applies: do not share it unless you must, and anonymise it if you do. If your company has no AI-use rule, write a one-page one and attach the table above. Ask your own legal adviser for a legal assessment.
One thing you can do alone today: pick a job you have written at least twice this week, such as a meeting summary, a standard customer reply or a weekly note. Write a five-part prompt and try it on a real example. Then save the version that works in your team’s sheet, with an owner and a date.
Sources
This article is general information, not legal, financial or technical advice. Platforms, their rules and the law change over time, so some details may have changed since 15 November 2026 or contain small errors. Before you make a decision for your business, check the current official sources or write to us.