How to write a prompt that gets useful work
To write a prompt that gets useful work, OpenAI, Anthropic, Google, and Microsoft give the same core advice: state the task with a verb, add context about who it is for and why, say what format you want, and refine the answer in follow-up messages. Their frameworks use different names and orders, and none of them requires every element in every prompt.
Writing a prompt that works at work
Most disappointing AI answers start with a request that a new colleague could not act on either. This page is part of day-to-day AI at work and draws on the prompting advice that OpenAI, Anthropic, Google, and Microsoft publish for their assistants, read on November 10, 2026. Anthropic’s prompting best practices give the simplest test: if a prompt would confuse a capable colleague who knows nothing about the job, it will confuse the AI.
The four companies agree on more than their different frameworks suggest. A useful prompt names the task, gives the context, says what the result should look like, and is followed by a second or third message that corrects course. Everything else is refinement.
Four points every company makes
- A clear task, led by a verb: summarize, draft, compare, rewrite, list. Google calls the verb the most important part of a prompt.
- Context: who the result is for, why you need it, and what you already know. Anthropic says explaining the reason behind an instruction helps Claude apply it to cases you did not foresee.
- The format and tone you expect, in measurable terms. OpenAI’s example: instead of “fairly short,” ask for three to five sentences.
- A conversation, not a single shot. All four tell you to refine, ask for options, or regenerate. Microsoft adds that the same prompt can give different answers on different tries.
OpenAI, Anthropic, and Microsoft also advise saying what to do rather than what not to do: “write in plain sentences” works better than “no jargon.”
Where the advice differs
| Microsoft | Anthropic | OpenAI | ||
|---|---|---|---|---|
| Framework | Persona, Task, Context, Format; a few are enough | Goal, Context, Source, Expectations; only the goal is required | Clear and direct, with context and examples | Clear task, context, tone, and style |
| Order | Not specified | Sources last; later parts weigh more | Long documents first, question at the end | Instructions first, text separated by ### or """ |
| Giving a role | One of the four parts | Not mentioned in the pages we read | One sentence is enough, says its developer guide; less needed with recent models, says its general guide | Not mentioned in the pages we read |
| Examples | Not mentioned in the pages we read | Samples as part of the source | Three to five in its developer guide; start with one in its general guide | Show the format; start without, add a few if needed |
Our way through the contradictions on order: keep the request clearly apart from the text it applies to, and with a long document, repeat the request in one line at the end.
On length, Google observed in 2024 that its most successful prompts averaged about 21 words with context, while users typically wrote fewer than nine. Google did not publish a method, but the gap is the point: most prompts are too thin, not too long.
A template to copy
This structure works in ChatGPT, Claude, Gemini, and Copilot. Leave out any line you don’t need.
Task: Draft a reply to the client email below.
Context: The client is a dental practice in Calgary. Our delivery is two days
late because a supplier was late. They are annoyed but polite.
Format: Under 150 words, in three short paragraphs.
Tone: Direct and calm, with one apology.
Source: Use only the email below and our new delivery date, Friday.
---
[paste the email here]
---
Reminder: draft the reply described at the top.
The same structure for a monthly report from a spreadsheet:
Task: Summarize October sales by region from the attached sheet.
Context: Columns A to E are date, region, product, units, and revenue in USD.
Some rows have no region; list them separately instead of guessing.
Format: A table with region, units, revenue, and change from September,
then three sentences on what stands out.
Check: Tell me which columns you used and show the formula for each total.
The last line matters most: asking the assistant to show its work turns an answer you have to trust into one you can check.
Three habits that save rework
- Ask what it needs. Google suggests asking Gemini what questions it has about the task.
- Allow “I don’t know.” Anthropic advises giving explicit permission to say that information is missing, which reduces invented answers.
- Split big jobs. OpenAI and Google both recommend breaking a complex task into several prompts: outline first, then each section.
Save what you repeat
If you type the same background every day, store it once. ChatGPT’s custom instructions hold 1,500 characters on Free and Go and 5,000 on other plans. Claude has account instructions, project instructions, and skills. Google Workspace accepts up to 1,000 instructions shared across Gmail, Docs, Drive, Slides, and Chat, separate from the Gemini app; personal accounts need Google AI Plus, Pro, or Ultra. In Copilot, the setting is under Settings, then Personalization, then Custom instructions, according to Microsoft’s guide to writing prompts.
Emails are the easiest place to start
The template above works best on short, frequent tasks. See drafting emails with AI for the tools built into Gmail and Outlook.
When the prompt is fine but the answer is wrong
A precise prompt reduces mistakes; it does not remove them. Our page on AI mistakes and how to catch them explains what to check.