How to Write a Good Prompt at Work: One Method and 10 Templates

In the survey we run before every workshop, the most common complaint about AI at work is the same: "the answers are too general". When we look at the prompts, almost all of them are one line long.

An AI assistant is not a search engine. It is closer to a capable new colleague on their first day: fast, well read and completely unaware of your clients, your tone and what happened last week. A good prompt is simply a good briefing.

What makes a prompt good?

We teach six parts, under the name ROCSRS (in Estonian, REKSPS). The full ROCSRS article explains each part in depth; here is the short version.

Part	The question it answers	Example (finance)
Role	Who should the AI act as?	"You are a controller preparing the monthly management report."
Objective	What must the result achieve?	"Explain the three largest deviations so the CEO can decide whether to act."
Context	What does it need to know?	"Attached: budget versus actual for September. Q3 included a one-off IT cost."
Structure	What should the output look like?	"Five bullets, at most 120 words, numbers in EUR."
Reasoning	Should it show its thinking?	"Show the calculation for each deviation."
Stop	What must it not do?	"Do not guess causes that are not in the data. Mark them as [ASK]."

You do not need all six every time. Role, objective and context do most of the work. Stop is the part people skip, and it is the one that guards against the most expensive mistake: a confident, invented number in a document that leaves the building.

How do Microsoft's and Google's prompt guides compare?

The vendors publish their own short formulas. They agree with each other more than they differ.

ROCSRS	Microsoft Copilot	Google Gemini
Role	Not named	Persona
Objective	Goal	Task
Context	Context, Source	Context
Structure	Expectations	Format
Reasoning	Not named	Not named
Stop	Not named	Not named

All three ask for a goal, context and a format. ROCSRS adds two lines that matter in work documents: whether to show reasoning, and where to stop. Microsoft's "source" is worth borrowing: in Copilot, naming the exact file, email thread or meeting changes the answer a lot.

How much effort should a prompt get?

The more important the result, the more time the prompt deserves. We teach three levels.

Level	Time	When	How
1. Just ask	About 10 seconds	Lookups, ideas, rewording your own text	One or two lines, then refine in the chat
2. Structured prompt	2-5 minutes	Anything that goes to a client or a manager	ROCSRS: at least role, objective, context, structure and stop
3. Let the AI write the prompt	5-15 minutes	Tasks you repeat every week	The AI interviews you, writes the prompt, you test and save it

For level 3, start with this:

I want a reusable prompt for [task]. Ask me questions one at a time until you know the role, the objective, the context, the structure of the output, whether reasoning should be shown and what you must never do. Then write the prompt with [placeholders] for the parts that change each time.

When a level 3 prompt is used every week by several people, turn it into a team AI assistant.

Set a personal master prompt once

Most assistants let you save standing instructions that apply to every chat. In Claude this is under Settings > Your Claude > Instructions; in ChatGPT under Settings > Personalization > Custom instructions. A few lines save you from repeating yourself:

Answer in [language] unless I ask otherwise. I work as [role] at a [type] company. Keep answers short and concrete; use lists only when I ask. Before you answer, tell me what is missing or contradictory in my request. Do not invent prices, deadlines, quantities or names. Do not promise anything to clients that I have not confirmed.

10 prompt templates by role

Copy, replace the [brackets], and add your own files. Each template follows ROCSRS in plain sentences.

Finance: management report commentary

You are a controller. Using the attached budget versus actual table for [month], explain the three largest deviations so that [CEO] can decide whether action is needed. Five bullets, at most 120 words, amounts in EUR. Show the calculation for each deviation. Do not guess causes that are not in the data; mark them [ASK].

Sales: reply to an enquiry

You are an account manager at [company]. Draft a reply to the client enquiry below that answers every question and proposes one next step. Our tone is friendly and short. Use the attached price list only. At most 150 words. Where a price, delivery date or quantity is not in the input or the price list, write [CHECK]. Do not offer discounts.

Customer service: complaint reply

You are a customer service specialist at [company]. Draft a reply to the complaint below (names removed) that acknowledges the problem, explains what happens next and gives a date for an update. Calm, human tone, no corporate phrases. At most 120 words. Do not admit liability or promise compensation; write [CHECK WITH MANAGER] where that decision is needed.

HR: job advertisement

You are an HR specialist at a [size] [industry] company. Write a job ad for [role] from the brief below that a strong candidate would read to the end. Structure: about us (2 sentences), the job, what we expect, what we offer, how to apply. At most 300 words. Before writing, list any information the brief is missing. Do not invent salary or benefits.

Operations: letter to a supplier

You are a purchasing manager. Write a letter to [supplier] about the delivery deviation described below: what happened, what it cost us if known, and what we need from them by [date]. Firm but polite. At most 200 words. Use only the facts in the input; mark anything uncertain as [CHECK].

Management: meeting recap

You are the chair of the meeting. From the notes below, write a recap for people who were not there. Three sections: decisions, open questions, next actions (each with owner and date). At most 250 words. Include only what the notes support; list unclear points as open questions instead of guessing.

Marketing: LinkedIn post from project notes

You are a B2B marketer at [company]. Turn the project notes below into a LinkedIn post that shows how we work, for [audience]. One concrete detail in the first line, no hype, no emojis, at most 120 words. Do not name the client or use any figure that is not in the notes.

Procurement or public sector: reading tender documents

You are a bid manager. Read the attached tender documents and list every requirement we must meet, in a table: requirement, document and section, deadline, mandatory or optional. Then list anything ambiguous we should ask the contracting authority. Quote section numbers exactly; if you cannot find something, say so instead of assuming.

Public sector: reply to a citizen

You are a specialist at [institution]. Draft a reply in plain language to the enquiry below (name and address removed). Answer the question first, then explain the next steps. At most 200 words, no officialese. Where a legal reference is needed, write [CHECK LEGAL REFERENCE] instead of citing from memory.

Everyone: a spreadsheet formula

I use [Excel or Google Sheets] in [language]. Column A has [what], column B has [what]. I need a formula in column C that [result]. Give the formula, one sentence on how it works, and one test I can do to check it. Use the function names for my language version.

What are the most common prompting mistakes?

• One line for an important document. Level 1 effort for a level 2 result.
• Pasting data with no instruction. The AI guesses what you want, usually a summary.
• No stop line. The result looks complete because the AI filled the gaps with plausible inventions.
• Accepting the first answer. Assistants tend to agree with you and to please. Ask what is missing or contradictory, and ask it to argue the other side. In one of our demos, an assistant reversed its conclusion on survey data the moment it got the context it had been missing.
• Wrong account for the data. A perfect prompt on a personal account is still a data problem. Our AI policy template sets out what may go where.

What to do this week

Next 15 minutes. Take the last AI answer that disappointed you. Rewrite the prompt with role, objective, context, structure and stop, and run it again.

Next 7 days. Set your personal master prompt. Copy the two templates closest to your work, adapt them and save the versions that work.

Next 30 days. Share the prompts that work with your team in one document, each with an owner. Turn the one used most into a team assistant.

In our team workshops everyone practises this method on their own documents, in the tool the company already licenses. Start with a discovery call.

Sources

Microsoft Support, Learn about Copilot prompts. Google, Gemini for Google Workspace prompting guide 101. Product facts checked on 30.10.2026.