Under the Hood · Part 3 · 5 min read
Tell it who to be
How a persona quietly changes every answer you get.
By Aikansh ·
Ask the same person a question as "a cautious accountant," then again as "a bold founder," and you will get two different answers. Same brain, different hat. AI works the same way — except the effect is larger, and most people never reach for the hat.
A model learned from millions of people, so inside it are many possible stances on almost anything. Ask it cold and you get the bland average of all of them — careful, hedged, forgettable. The moment you tell it who to be — a role, a set of principles, a named expert — you select which voice answers. And that does not just change the tone. It changes the substance: what it considers, what it weighs, what it bothers to warn you about.
WHY A PERSONA CHANGES THE ANSWER
- It read all of us. The model holds many stances at once. Unprompted, it hands you the average of them.
- A role narrows it. Name a role or a principle and it collapses to one coherent stance — sharper, steadier, more useful than the average.
- Famous names are shortcuts. Borrowing a well-documented figure or framework hands it a rich, ready-made way of thinking to imitate.
- The wheel turns both ways. A persona steers powerfully — including confidently in the wrong direction. Steering is not the same as being right.
Here is the move, made concrete. Instead of "review my plan," try: "You are a skeptical CFO who has watched a hundred startups fail. Review my plan." The first gets you a tidy summary. The second gets you the three things that will actually sink you. Same model, same plan, a different hat.
One thing to hold onto, so this doesn't collapse into "just prompt better": a persona is not magic, it's a constraint. You aren't adding knowledge — you're narrowing the model to one stance out of the many it already holds. Point it at the right expert and it sharpens; point it wrong and it's confidently off. Steering, not truth.
Once you see it
The people getting something close to magic out of these tools are usually not asking cleverer questions. They cast the model in the right role first, and only then ask. It is the cheapest, most underused lever there is — and it is sitting in the first sentence of your prompt.
You're not just asking a question.
You're casting it in a role.
TRY IT — 2 MINUTES
Open Claude. Ask a real question — say, "review this paragraph" — and read the answer. Then ask the exact same thing twice more, changing only the first line:
- "You are a blunt copy editor who hates filler. Review this paragraph."
- "You are a kind writing teacher. Review this paragraph."
Same engine, same question — three different answers. You changed who's driving without touching a word of the actual task.
Next: a driver who knows who they are still can't move without wheels — the tools that let it touch the road.
FOR THE CURIOUS — WHAT'S ACTUALLY HAPPENING
This is role conditioning. A model predicts the most likely continuation of the text it is given; a persona changes what is likely. Set the opening context to a skeptical CFO and the high-probability continuations shift toward risk, cash flow, and failure modes. The persona does not add knowledge — it re-weights what the model already has.
Named figures and established frameworks work because the training data is dense around them: the model has seen enough of how they think to imitate it coherently. Standing tenets — the system prompt, a set of principles — act as a persistent constraint on every answer, which is the same reason good teams write down leadership principles instead of re-deciding their values each morning. The caveat is real: a confident persona can amplify bias and tell you what you want to hear. A steering wheel is only as good as the driver pointing it.
A PERSONA PATTERN LIBRARY
Don't prompt the average — cast the expert. A few I reach for, drawn from the systems I've built:
- “Act as a supply-chain operator reviewing this fulfillment plan. Where does it break under peak load?”
- “Act as a marketplace GM reviewing this seller-onboarding strategy. What kills liquidity in the first 90 days?”
- “Act as an engineering director reviewing this AI architecture. Where will it page someone at 3am?”
- “Act as a skeptical CFO reviewing this small-business acquisition. What in the numbers would make you walk?”
The hat does half the work. The other half is checking whether the expert is actually right.