ChatGPT Role Prompt: How to Make AI Sound Professional

Why ChatGPT Sounds Like A Student

Ask ChatGPT a question with no role attached and you get the average of everything it has ever read on that topic. Averages are safe. Averages are also bland, over-explained, and weirdly enthusiastic — which is exactly what a senior reader notices first.

Here is what most guides get wrong. They treat the role as a costume — say “act as a marketing expert” and the words get fancier. It is closer to a filter. Naming a job narrows which vocabulary, which structure, and which concerns come forward first. A finance director and a communications lead summarise the same quarter differently, and not because one knows more.

The three tells of AI writing

Senior readers do not consciously spot AI. They spot a pattern. Three things give a draft away almost every time:

• An opener that restates the question before answering it
• Vocabulary nobody says out loud — pivotal, testament, robust, spearheaded
• Relentless positivity, even when the news is genuinely bad

Same request. Same model. The only difference was three lines at the top.

The Three-Line Role Card Explained

The Three-Line Role Card is a ChatGPT role prompt written as exactly three lines: the badge (job title, seniority, years), the room (who reads this and what they need from it), and the fence (the words, tone, and moves that are banned).

The fence is the line almost everyone leaves out, and it is the one that does the heaviest lifting. Telling the model what to avoid is a distinct skill worth learning properly — we cover it in depth in our guide to negative constraints in AI prompts.

How to Write a ChatGPT Role Prompt

You need about ten minutes for your first card. Every card after that takes under a minute, because you will be editing the last one rather than starting fresh.

Skip the adjectives. “Growth guru” and “world-class strategist” pull the model straight into marketing-speak. A job title, a seniority level, and a number of years is enough, and the number matters more than people expect — it sets how much explaining the model does.

Who reads this, what do they already know, and what mood are they in when it arrives? A board that has seen the numbers needs a different opening from a team hearing them for the first time. This is the line that kills over-explaining.

List the words you never want to see, the tone you refuse, and the structural habits that annoy you. Your fence should be personal. Mine bans exclamation marks and any sentence beginning with “In today’s”.

You are a [job title] with [number] years of experience in [industry or specialism].

This is written for [who reads it] who already know [what they know] and currently feel [mood or concern].

Do not use corporate cliches such as delve, synergy, testament, pivotal, spearheaded, tapestry or journey. No exclamation marks. No passive voice. No opening line that restates my request. Say clearly when something needs a human decision rather than guessing.

Paste the card, then your actual request underneath it. Do not merge them — keeping the role separate from the task is what lets you reuse the same card across twenty different jobs.

Five Role Cards for Real Work

These are complete cards, not fragments. Each one is a full ChatGPT role prompt in the badge, room, fence shape. Copy one, swap the brackets, and add your request underneath.

The first is the hardest email most managers ever send. Notice how much of the card is spent on the room rather than the badge — when the news is bad, the reader’s state matters more than the writer’s credentials.

You are a senior HR business partner with 15 years of experience in organisational change.

This goes to 50 people in one department who will read it on their phones and immediately wonder whether their own job is safe. They have heard rumours for three weeks.

Be direct in the first two sentences. No warm-up paragraph. Do not describe the change as exciting or as a journey. Do not use passive voice to hide who decided what. Under 300 words. Flag anything that legal should review before this goes out.

The second is the memo that gets read in a lift on the way to the meeting. The badge here is doing real work: a CFO writing for non-financial readers automatically drops the jargon a finance analyst would keep.

You are a CFO presenting to a board where three of the seven members have no financial background.

They will read this for four minutes before the meeting and want to know what changed, what it costs, and what you want them to decide.

Lead with the decision you are asking for. No more than three numbers in the opening paragraph. Do not explain accounting terms unless I ask. Do not invent figures — if a number is missing from what I give you, write MISSING and move on.

Three shorter cards cover most of what is left. Each one follows the same badge, room, fence shape, just compressed.

You are a consultative account executive who sells to operations leaders and has never used a hard close in your life.

This goes to a VP who was interested but distracted, and who gets fifteen vendor emails a week.

Under 120 words. No pleasantries about hoping they are well. No bullet list of features. One clear next step, phrased so that saying no is easy.

You are a managing editor who has rejected far more drafts than you have published, and who explains exactly why.

This is for me, the writer. I want the honest version, not the encouraging version.

Do not compliment anything before criticising it. Quote the specific sentence you are objecting to. Mark every claim I have not evidenced. Do not rewrite the whole piece — fix the three worst paragraphs only.

You are a principal programme manager who escalates to executives roughly once a quarter and is trusted because you never overstate.

This goes to a sponsor who has not followed the detail and will decide in five minutes whether to intervene.

State the delay, the cause and the impact on the critical path in that order. No blame directed at a named team or vendor. Two options with a recommendation. No more than one page.

Where to Keep Your Role Card

A ChatGPT role prompt you retype every morning is one you will abandon by Thursday. ChatGPT gives you three places to store one, and they behave differently.

Custom Instructions is where a role stops being a prompt and becomes a setting. On web and desktop you will find it under Settings, then Personalization, then Custom Instructions. On iOS and Android the same panel is called Customize ChatGPT. According to OpenAI’s Custom Instructions documentation, it is available on every plan, with Free and Go accounts capped at 1,500 characters and Plus and above at 5,000.

That cap answers a question people ask constantly: how long should a role be? Short enough to fit twice over. If your card is pushing 1,500 characters, it has stopped being a role and started being a brief.

For work that has its own audience — one client, one report series, one board — put the role in a Project instead. Our walkthrough of how to organise your AI work with ChatGPT Projects covers the setup, and the difference between account-level and project-level rules is worth understanding properly if you are juggling several. We unpack that hierarchy in system prompts versus custom instructions. If you want a permanent default role for everything you do, our full Custom Instructions guide is the place to start.

When Your ChatGPT Role Prompt Fails

Three failures account for nearly everything people complain about. All three are fixable from the card itself.

The first one is worth seeing side by side, because the difference in the input looks trivial and the difference in the output really is not.

Drift is the other one people find genuinely annoying. By the fifth or sixth exchange the replies start to sound generic again. The cheap fix is a re-anchor: “staying in the HR business partner role, redraft paragraph two.” The permanent fix is moving the card into settings. If output variation frustrates you generally, our piece on why ChatGPT gives inconsistent results goes deeper into the structural causes.

What a Role Cannot Do

This is the part the prompt-list articles skip, and it is the part that protects your reputation.

The confidence problem gets worse with a role, not better, which surprises people. An authoritative badge produces authoritative-sounding fabrications. We explain the underlying behaviour in our guide to ChatGPT hallucination, and OpenAI’s own guidance in its prompt engineering best practices for ChatGPT makes the same point about being specific rather than assuming the model fills gaps correctly.

There is a data question too. Role prompts tend to pull real workplace material into the chat — draft announcements, board notes, client emails — and that is where the actual risk sits.

One thing worth knowing: on personal ChatGPT accounts, your conversations may be used to improve the models unless you opt out in settings. That is true on the free plan and on the paid personal plans alike — the guaranteed no-training position comes with business and enterprise accounts, not with paying more as an individual. We cover the practical implications in is ChatGPT safe to use at work.

Where to Take This Next

Most people move through three stages, and the jump that matters is the second one — from a costume to a card.

Write one card this week for the task you do most often. Save it somewhere you will find it again. Then write the second card only when the first one visibly fails to fit — that is how you end up with three good cards instead of thirty mediocre ones.

Role is the first of four elements. Once your ChatGPT role prompt is reliable, the next gain comes from the Goal line, and after that from Context and Format. The AI Execution Loop is where the four fit together into a system you stop having to think about.