Persona Depth: How Specific Should Your Role Be?

TL;DR: More persona detail doesn't buy more accuracy. A 2026 study found a minimal, five-token persona cost less MMLU accuracy than a full, ~150-token one (68.0% vs. 66.3%, against a 71.6% baseline, no persona). But extra depth does earn real gains on tone, style, and safety refusals. The right depth depends on the task, not a template.

"You are a world-class [role] with 20 years of experience" is probably the single most copy-pasted line in prompt engineering, and almost nobody asks how much of it is doing anything. Is the job title enough? Does the fabricated tenure help? Does piling on more biography make the output better, or just longer? Persona prompting gets covered as a yes-or-no question a lot: does it work, does it not. This post asks the question underneath that one, the one the popular templates never test: once you've decided to use a persona, how much of one do you actually need?

What Do We Actually Mean by "Persona Depth"?

Think of it as a dial, not a switch. At one end there's no role at all. Moving up:

1. A bare label. "You are a lawyer." One clause, no other detail.
2. A one-line role. "You are a senior corporate lawyer advising a startup founder." Adds seniority and audience in a single sentence.
3. A short persona. A paragraph: role, audience, what it optimizes for, maybe one constraint.
4. A full persona. Multiple paragraphs: biography, credentials, voice attributes, explicit refusals, what it does when it doesn't know something.

A 2026 USC preprint on persona effectiveness formalizes almost exactly this scale for its experiments, generating each of its 12 test personas at three granularity levels it calls full, short, and minimum, running roughly 150 tokens down to around 5. That's the closest thing to a controlled version of the question every prompt writer is actually asking when they wonder whether to trim their persona or pad it.

In practice, level 2 or 3 on that scale is where most real prompts already sit, whether the writer thought about it or not. "You are a helpful assistant" (a bare label) and a three-paragraph biography with a name, an alma mater, and a list of past clients (a full persona) are both common in the wild, and they are not the same instruction. The question this post is answering is which end of that range earns its keep, and for what kind of task, rather than treating "add a persona" as a single move with one effect.

Does a Deeper Persona Make the Model More Accurate?

No, and the direction of the effect is the opposite of what the popular templates assume. Three separate pieces of research land on the same conclusion from different angles.

An EMNLP 2024 Findings paper by Zheng et al. ran the most exhaustive version of this test to date. The paper's own framing: "We curate a list of 162 roles covering 6 types of interpersonal relationships and 8 domains of expertise." The evaluation itself ran across 2,410 factual questions and four families of LLMs. Its conclusion is blunt: "adding personas in system prompts does not improve model performance across a range of questions compared to the control setting where no persona is added." Where a persona did help on a specific question, the paper found the effect wasn't attributable to any obvious property of that persona: "the effect of each persona can be largely random."

Wharton's Generative AI Labs ran a similar test on MMLU-Pro and GPQA in a December 2025 technical report and reached the same place from a different benchmark: expert personas didn't reliably beat a no-persona baseline, low-knowledge personas ("Toddler," "Layperson") reliably hurt it, and matching the persona's domain to the question's domain didn't rescue it either. The report surfaced a failure mode neither earlier study had flagged: given an out-of-domain expert persona, Gemini 2.5 Flash refused to answer roughly 42% of trials, typically saying it lacked the relevant expertise.

Neither of those studies varied persona depth on its own, though. The 2026 USC paper did, and this is the number that actually answers this post's question. Against a 71.6% no-persona baseline on MMLU, "all expert persona variants damage accuracy, with the minimum persona suffering the least (overall: 68.0% vs. 71.6% baseline)." Adding more detail didn't recover any of that loss. It cost more: "the minimum persona achieves 68.0% vs. 66.3% for the long persona (both below the 71.6% baseline)", with the full-length version losing the most.

The same paper reports the persona underperforming the no-persona baseline consistently across all four MMLU subject categories it tested, not just on average across a mixed set. One caveat worth stating plainly rather than glossing over: the paper also found this direction flips for reasoning-distilled models on its broader task set, where expert personas improved overall scores instead of hurting them. So "depth costs accuracy" is the pattern for standard instruction-tuned models on knowledge lookup specifically, not a universal law for every model architecture. Combined with the 162-role and MMLU-Pro/GPQA results above, though, three independent research groups, using three different benchmarks, land on the same general place for the models most people are actually using: a persona doesn't add facts to a model that weren't already there, and loading in more instruction-following behavior leaves less room for the model to just retrieve the answer.

So Where Does Extra Depth Actually Pay Off?

Not on facts, but not nowhere either. The same USC study measured persona effects on two other axes: generative quality (MT-Bench) and safety alignment (adversarial-prompt benchmarks), and depth behaved completely differently there.

On MT-Bench, the paper found "long expert personas yield the strongest category improvements (e.g., Extraction +0.65, STEM +0.60)", the same direction as the accuracy result above, just reversed in effect. On the safety side, a dedicated "Safety Monitor" persona built specifically to hold a behavioral line showed the same pattern: "the long Safety Monitor achieves +17.7% on JailbreakBench vs. +8.9% for the minimum prompting variant." More detail, more gain, on both.

Why would depth cut in opposite directions depending on the task? The most plausible read, and the one the paper itself offers, is that persona text competes with the model's pretrained factual recall for its attention: more persona detail means more instruction to follow, which crowds out fact retrieval on a lookup task but reinforces a behavior on a task that's actually about behavior (write in this voice, extract in this shape, refuse this category of request). Depth isn't good or bad in the abstract. It's a lever that pulls toward "follow this instruction more faithfully," and whether that helps depends entirely on whether the instruction was the whole job or just decoration on top of a fact-retrieval job.

What Do OpenAI, Anthropic, and Google Actually Tell You to Do?

Their own current documentation, read directly rather than through a blog post about their documentation, is consistent with the research and with each other: keep it short by default.

None of the three vendors' own reference examples resemble the elaborate, multi-paragraph "world-class expert with 20 years of experience, published in top journals, known for..." template that circulates as received wisdom. Every official example is one to three sentences. That's not proof the longer version never helps (the USC data above says it sometimes does, on the right kind of task), but it is evidence that the vendors who built these models don't treat elaborate depth as the default, and neither should you.

Is "20 Years of Experience" Load-Bearing, or Cargo Cult?

Mostly cargo cult, and it's worth saying plainly rather than hedging around it. No vendor documentation and no study cited in this post tests a specific tenure claim against another. Nobody has published a result showing 20 years beats 10, or that "world-class" beats "senior." The phrase is popular because it reads like credibility signaling, the same reason a resume pads with adjectives, not because any experiment found the number moves the output.

What the research does show as load-bearing: the audience the response is written for, the format it should take, what the persona explicitly won't do, and what it should say when it doesn't know something. Those are the fields the USC and Wharton work actually varied and measured effects from. A fabricated biography is not one of them. If you're going to spend words on persona depth, spend them on those four, not on an invented career history.

Here's the same role at three depths, so you can see exactly where the words go:

MINIMUM
You are a financial analyst.

SHORT
You are a financial analyst who reviews quarterly earnings for a
retail investor audience. Flag anything a non-specialist would
misread. State clearly when a number requires a filing you don't
have access to, instead of estimating it.

FULL
You are a financial analyst reviewing quarterly earnings for a
retail investor audience who reads confidently but isn't a
professional. Optimize for: catching the one number that changes
the story, not restating the whole release. Never speculate on
stock price direction — flag that as outside scope. Vocabulary:
plain English with the technical term given once, then a plain
restatement. When a figure requires a filing you don't have, say
exactly what's missing rather than estimating it.

These are illustrations, not a test we ran. But notice that the difference between SHORT and FULL isn't biography, it's more constraints: a sharper "optimize for," an explicit refusal, a vocabulary rule. That's the kind of depth the research above says actually changes behavior, not a résumé.

How Should You Actually Decide Persona Depth for a Given Prompt?

Match the depth to the job, in roughly this order:

• Factual lookup, math, or code correctness: skip the persona, or cap it at one clause. Spend the words you saved on constraints and output format instead, where they're proven to help.
• Writing, tone, or extraction tasks: a short-to-full persona is worth it. This is exactly where the USC data shows the biggest gains, and where audience and vocabulary rules do real work.
• A safety or compliance boundary you need enforced strictly: invest in a full, explicit persona built around the boundary itself (what it refuses, and why), the same shape as the "Safety Monitor" condition that outperformed a minimal one.
• Anything else, or when you're not sure: start minimal. Add one dimension, audience, a refusal, a format rule, at a time, and check whether the output actually changed before adding the next one. Most prompts stop needing more after one or two additions.

Can You Just Test This on Your Own Prompt, Instead of Trusting Any of the Above?

Yes, and for anything that actually matters, you should. The studies above ran thousands of trials across benchmark questions built for exactly this kind of comparison. Your prompt is one task, usually without a labeled right answer sitting next to it, so the same experiment on your own work will be noisier, but the method still transfers: run the same request through a minimal persona and a full one, change nothing else, and judge only the thing you actually care about, not which answer sounds more confident. Confidence is cheap for a model to produce regardless of whether the persona is one clause or ten; a "world-class expert" voice reads as more authoritative whether or not the underlying content is more correct, which is exactly the trap the research above is warning about. If a longer persona isn't measurably changing the one thing you're optimizing for, the shorter one was the right call, and it's also the one that costs fewer tokens on every single call afterward.