Type “summarize this” above a page of meeting notes and ChatGPT will give you something. Say who the summary is for, what it has to decide, what shape you want back and what it must not guess, and you get something you can forward. That gap is most of what prompt writing for work comes down to.

Google saw the same pattern with its own assistant. Its Gemini for Workspace prompting guide (October 2024 edition) says the most fruitful prompts from its users averaged around 21 words with relevant context. The prompts people actually tried were usually under nine words.

A four-part structure: goal, context, format, limits

OpenAI’s prompt engineering best practices for ChatGPT start with one rule: be clear and specific, and give enough context. Splitting that into four parts makes it easy to remember on a busy day.

1. Goal: start with a verb

Say what you want done: draft, summarize, rewrite, compare, turn into a table. Google’s guide calls the verb or command the most important part of a prompt. “Write a reply to this customer” beats “customer email”.

2. Context: who it is for and why

Add the reader, the situation and what the output is for. A recap for your manager and a recap for a client need different things. Google’s version adds a persona (“You are a program manager in…”), a quick way to set the angle.

3. Format: describe the shape you want back

Ask for headings, bullets, a table or a word count. The ChatGPT help page says OpenAI’s API guides apply to ChatGPT too, and its API prompting best practices swap fuzzy requests for exact ones. “Use a 3 to 5 sentence paragraph” beats “fairly short”. The same article says models respond better when shown an example of the format.

4. Limits: what to use, and what to do when something is missing

This is where most work prompts are weakest. Name the material to use, the tone to keep and what to do when information is missing. OpenAI’s API guide advises saying what to do instead of only what not to do.

Two small habits help too. Put the instruction first and fence off pasted material with triple quotes or ###, as the API article suggests. And plan for a second round: the ChatGPT help page treats refining the prompt after the first answer as normal.

Sample prompts for four common team tasks

Each sample follows the four parts. They are starting points written for this guide, so swap in your own details. Square brackets mark what you fill in.

Meeting recap

Sample: Turn the notes below into a recap for the people on this project who missed the call.

Context: we are deciding whether to move our December sale forward by a week.

Format: three headings (Decisions, Action items with owner and due date, Open questions), under 150 words.

Limits: use only what is in the notes. If an owner or date is missing, write [OWNER?] or [DATE?] instead of guessing.

Notes: “””[paste notes]”””

If your team records calls, the built-in note takers in Meet, Zoom, Teams and Notion can produce the raw notes. We compared them in our guide to AI meeting notes for small teams.

Customer reply

Sample: Draft a reply to the customer message below.

Context: their order arrived late and they want a refund. Our policy allows a refund or a 15% voucher.

Format: under 120 words, friendly and plain, no bullet points.

Limits: do not promise a delivery date. If there is no order number, ask for it.

Message: “””[paste message]”””

The policy and the 15% voucher are made up for this sample. Before you paste a real customer’s name or order history into any AI tool, read our checks for putting customer data into AI tools at work.

Spreadsheet cleanup

Sample: Clean up the rows below.

Context: a supplier list exported from two systems, so formats differ.

Format: the same columns, dates as YYYY-MM-DD, phone numbers with country code.

Limits: do not fill in empty cells; write [MISSING]. List likely duplicates under the table instead of deleting them.

Rows: “””[paste 20 to 50 rows]”””

For anything longer than a screenful, upload the file instead. OpenAI’s page on whether ChatGPT tells the truth says its data analysis tool is what handles accurate calculations.

Project brief

Sample: Write a one-page brief for a freelance designer.

Context: we need a new label for our coffee bags, and the designer has never worked with us.

Format: sections for Background, What we need, Audience, Must include, Deadline and Budget.

Limits: use only the facts below. Where a section has no facts, write [ASK ME] so I know what to fill in.

Facts: “””[paste notes]”””

For more templates, OpenAI publishes a prompt pack for work in OpenAI Academy, last updated 17 September 2026. One reads: “Summarize these meeting notes into a structured recap. The notes are rough and informal. Organize them into categories: key decisions, next steps, and responsibilities.”

What a better prompt cannot fix

A good structure makes answers easier to use. It does not make them true. OpenAI’s help page lists the usual failures: wrong dates and facts, fabricated quotes and citations, and overconfident answers to ambiguous questions. It adds that the model may sound confident even when it is wrong.

Two problems keep coming up at work. The first is missing data: if the sales figure is not in what you pasted, no wording will conjure it. The second is made-up figures: asked for a number it lacks, a model may fill the gap with a plausible one.

Ask for placeholders instead of guesses

That is why every sample above says what to write when something is missing. Anthropic’s guide to reducing hallucinations in Claude recommends the same move: give the model explicit permission to say it does not know. One of its examples tells the model to remove any claim it cannot back with a quote and mark the gap with empty brackets.

The advice is written for Claude, but the idea works as a habit with any chatbot. A visible [MISSING] is easy to spot and fill. A confident wrong number in a client email is not.

Even then, treat the result as a first draft. OpenAI tells users to verify quotes, data and references. Anthropic’s page ends the same way: these techniques reduce hallucinations but do not eliminate them.

Where this approach stops working

Prompts like these suit drafting and tidying: text in, better text out. They fall short when the job needs facts the model cannot see, such as today’s stock levels, unless you upload them or turn on search. Software that completes tasks in other apps on its own is a different kind of tool, covered in our explainer on how an AI agent differs from a chatbot.

Longer prompts also take longer to write. For a quick question, a one-line prompt is fine. The structure earns its keep on tasks you repeat every week, where a saved prompt pays back the minutes it took.

What to do this week

Pick one task you hand to ChatGPT every week, such as the Monday recap or a common customer reply. Rewrite your usual prompt with a goal, context, format and limits, plus a [MISSING] rule. Run it on real material and fix what went wrong in the prompt itself. Then save the final version in a shared doc so your team can reuse it.

Sources

Anna Jakovleva

Anna Jakovleva is the senior editor for AI tools at AI Observer, leading coverage of agents and automation, design and product, and the site's tutorials. She worked in operations and product teams before joining, so she judges tools by setup time, cost and where they break. Her writing is hands-on and practical.

Spotted a mistake? Tell us on the corrections page.

Conversation

Every comment is reviewed before it appears, so there can be a short delay.