Do role prompts still help with GPT-5?

A Beginner's Guide to ChatGPT teaches a simple order for a good request: give ChatGPT a role, add context, state the task, ask for a format, then iterate and verify. With GPT-5 now the default model in ChatGPT, it is fair to ask whether the first step still earns its place. GPT-5 is better at working out what you mean, so does "You are an experienced travel agent" still change anything?

The short answer is that roles still help, but mostly as a compact way to pass on expectations. A role tells the model who the reader is, what standard to meet and what tone to use. It does not make the model know more facts than it already knows.

What the evidence says

Two sources are useful here. The first is a research study, published on arXiv and accepted at a 2024 conference, that tested 162 different roles, such as "You are a lawyer" or "You are a helpful friend," on 2,410 factual questions across four popular families of language models. Adding a role to the system prompt (the background instructions a chatbot receives before your message) did not raise accuracy overall compared with no role at all. Some roles helped on some questions, but the authors could not predict which ones in advance; picking the right role automatically worked about as well as picking one at random.

The second is OpenAI's own GPT-5 prompting guide, published with the model. Its sample prompts still open by telling the model what it is and what it is for, for example a scheduling assistant for a clinic, and then list detailed rules. The guide's main warning is about something else: GPT-5 follows instructions very closely, so instructions that are vague or contradict each other cost it more effort and can lower quality. In other words, a role is a frame, and what goes inside the frame matters more.

A worked test: with and without a role

Here is a test you can run yourself. Open two new chats so neither one remembers the other.

Chat A, no role:

Explain what a heat pump is and whether it makes sense for a house in a cold climate. Keep it under 200 words.

Chat B, a role that carries information:

You are an energy advisor who explains home heating to homeowners with no technical background. You give balanced advice and say plainly when something depends on local prices or the house itself.

Explain what a heat pump is and whether it makes sense for a house in a cold climate. Keep it under 200 words.

When you compare the two, look at four things: the vocabulary, whether the answer admits what it depends on, how the advice is balanced, and whether any facts differ. The outcome below is an illustration of the kind of difference to look for, not a measured result, and every run varies. Both answers will usually get the basic facts right. Chat B may use plainer words, adds a line such as "it depends on your electricity rates and insulation," and ends with a suggestion to get a local quote. Chat A is accurate but reads more like an encyclopedia entry.

That pattern fits the research. The role did not supply new facts; it supplied an audience (homeowners without technical background) and a standard (balanced, honest about what depends on local conditions). You could get much of the same effect without the words "You are," by writing "Explain this for a homeowner with no technical background, and say what depends on local prices."

What to put in a role for GPT-5

A useful role answers three questions in one or two sentences:

• Who is speaking? A job or point of view that implies a level of expertise, such as "a pharmacist," "a hiring manager," "a primary-school teacher."
• Who is listening? The reader's level and situation.
• What standard applies? Balanced, cautious, concise, warm, strict about sources, and so on.

Keep it consistent with the rest of the prompt. If the role says "You are a concise expert" and the format asks for a detailed, beginner-friendly guide, GPT-5 will try to satisfy both, and OpenAI's guide notes that such conflicts can hurt the result. When something in the answer is off, OpenAI also suggests asking GPT-5 itself which words in your prompt to add or remove to get the behavior you want.

When is a role just extra words?

A role is extra words when it adds no information the rest of the prompt doesn't already give. "You are a helpful assistant" is the classic case: the model is already trying to be helpful, so the sentence changes nothing. The same goes for flattering titles such as "You are the world's greatest expert": they do not unlock better knowledge, and the study above found no reliable gain from roles on factual questions.

A role is also extra when your context already says everything. If you write "I'm preparing a two-minute talk for parents at a school meeting; explain screen-time research in plain words and point out what is still uncertain," adding "You are an education researcher" adds little. The audience, purpose and standard are all there.

Finally, a role never replaces checking. If the answer includes figures, names, dates or medical, legal or financial advice, verify them against a reliable source, whatever expert the model was asked to play. Keep the role when it tells the model something it would otherwise have to guess, and drop it when it only sounds impressive.