What it is
One or two sentences about the person who will read the answer. Name who they are, what they already know, and what they will do once they have read it. On the ten checks this is Purpose, the second check, and it changes nearly every choice the model makes: what goes first, which terms need explaining, how much detail survives, what can be left out.
The part people skip is the last one. “For my manager” is a start. “For my manager, who has to approve the budget by Friday” tells the model the answer is really a decision, so the cost and the deadline belong at the top and the history can shrink to a line.
When it helps
It helps whenever the reader is not you. A summary for a colleague who missed the meeting, a reply to a customer, an explanation of a policy for staff who will never read the policy itself: each has a reader whose needs the model cannot see.
It helps most when the reader knows less, or something different, than the material assumes. Technical writing for people outside the field, legal wording for a customer, an engineer’s notes for a sales team. Saying what the reader already knows tells the model which words to explain and which to leave alone.
And it helps when one piece of material goes to several people. The board, the team and the customer read the same news for different reasons, and naming each reader in turn gets you versions that differ where they should. The summary for different audiences template is built that way.
When it does not
For your own questions it adds little. If you are asking for a spreadsheet formula or the meaning of a term, you are the reader, and the model can see the question you asked.
A job title on its own can backfire. “For executives” can get you short sentences and big words, a picture of what executives are supposed to like. “For the finance director, who will decide whether to fund the pilot and cares most about payback” gets you the cost and the payback period. Describe what the reader needs and knows, not just their rank.
A document with several readers pulls in several directions. If the engineers and the board both get the same page, one of them gets the wrong one. Pick the reader the document is mainly for, or ask for separate versions.
You do not need a name. A role and a need do the work, and leaving real names out of the prompt is a good habit anyway; the guide to anonymizing text explains why.
Before and after
Three requests where the only change is the reader. Each score comes from the Prompt Grader’s free ten-check estimate, written for any chat model.
Summarize this incident report. """ [the incident report, pasted in full] """
Summarize this incident report for our chief operating officer, who has to decide by Thursday whether to delay next week's launch. They were not on the incident call, and they know the product well but not how the database is set up. """ [the incident report, pasted in full] """
The first version gets a summary in the report’s own order: timeline, cause, fix. The second gets one ordered for the decision, with what the incident means for the launch near the top and the database detail either explained in plain terms or left out.
Explain our new expense policy. """ [the expense policy, pasted in full] """
Explain our new expense policy to warehouse shift leads, who will read it out to their teams at the start of a shift. Most of their staff rarely file expenses; the two things they ask about are meal allowances on overtime and mileage between sites. """ [the expense policy, pasted in full] """
Without a reader, the model walks through the whole policy at the same depth. With one, it can lead with the two questions staff actually ask and write sentences that still make sense when someone reads them aloud in a loading bay.
Write release notes for version 6.2 from the changelog below. """ [the changelog, pasted in full] """
Write release notes for version 6.2 from the changelog below. They are for our customers' office managers, who set the app up for their staff and want to know what will look different on Monday and whether they need to do anything. """ [the changelog, pasted in full] """
Release notes for no one in particular can list every change in the order it was made. For office managers, the model can sort by what their staff will notice and pull the one change that needs action to the top.
Each pair moves one point, because only one check changed. The answers can move further than the scores: once the model knows who is reading, it also knows what it can drop.
The line to add
This is for {{who_will_read_it}}, who already knows {{what_they_know}} and will use it to {{what_they_will_do}}.
If you can only fill one blank, fill the last. What the reader will do with the answer, whether that is approve, reply, follow the steps or pass it on, tells the model more than who they are.
In ChatGPT, Claude, Gemini and Copilot
No assistant needs special wording for this; a plain sentence near the top works in all four. The vendors describe it in their own terms. Microsoft’s guide to Copilot prompts calls it context and frames it as two questions, why you need it and who is involved; one of its sample prompts describes the audience as professionals who work in a hybrid environment.
Anthropic’s guide for Claude recommends giving the reason behind an instruction. Its example tells Claude the response will be read aloud by a text-to-speech engine, which explains why to avoid ellipses, and notes that Claude can generalize from the explanation. Who will read the answer, and how, is the most common reason there is. In Google’s prompting guide for Gemini in Workspace, the four areas are persona, task, context and format; persona is the role you give Gemini, so the reader fits under context.
Naming the reader is different from giving the model a role. “You are a financial analyst” changes the voice; “this is for a finance director deciding on a pilot” changes what goes in. The role pattern covers when the first is worth adding too.
Key takeaways
- Say who will read the answer, what they already know and what they will do with it.
- The last part matters most: a reader who has to decide needs a different document from one who has to follow steps.
- A job title alone invites a stereotype. Describe the need instead.
- One document, one main reader. Ask for separate versions when there are several.
- A role and a need are enough. Leave real names out.
Questions
How do I tell ChatGPT who my audience is?
Say it in a sentence near the start of the prompt: who will read the answer, what they already know and what they will do with it. “For new customers who have never used an invoicing app and need to send their first invoice today” does more than “for beginners”. If you write for the same readers every day, put them in your custom instructions so you do not have to repeat them.
Should I give AI a role or describe the audience?
For most work writing, start with the audience. A role changes the voice the model writes in; the audience changes what goes in and what is left out. You can use both, and a one-sentence role is plenty. The role pattern covers when it is worth adding.
Can I use one prompt for several audiences?
Yes. Ask for one version per reader, each labeled, and say what each reader needs from it. A board, a team and a customer read the same news for different reasons, so the versions should differ in what comes first as well as in wording. The summary for different audiences template is set up this way.