Personas don't transfer: the role prompt that helps on one model can hurt on the next

Overview

A study published on 19 July gave two frontier models the same saved persona — a research librarian — and asked each for code across twelve tasks. On Claude Opus, the persona produced in-character disclaimers in 55 of 60 responses, twelve outright refusals to write code at all, and a drop in mean correctness from 0.92 to 0.67. On GPT-5.5: no refusals, and correctness essentially unchanged. Same words, opposite behaviour.

That saved persona line in your prompt library — “act as a senior analyst”, “you are a meticulous editor” — is not a quality setting that travels between models. It is a behavioural policy that each specific model interprets its own way, and it is the line most likely to break silently when the model underneath you changes.

The content

The comfortable assumption behind every prompt library, custom instruction, and shared team agent is that a persona is a stable dial: written once, useful everywhere. The librarian study’s authors read their results as personas acting as “model-dependent behavioral-policy biases rather than universal quality interventions.” The same identity that sharpens one model can push another into character so hard it stops doing the job.

This is not one quirky result. The peer-reviewed EMNLP 2024 study ran 162 personas over 2,410 factual questions across four model families, and beneath its headline (personas didn’t improve accuracy — a finding this site has covered) sits the portability evidence: on Llama3-70B, more personas had negative effects, while the Qwen2.5 models were insensitive to all 162; even the correlation between a persona prompt’s fluency and its accuracy points in opposite directions on different families. Trying to select the best persona automatically often performed no better than random. A May 2026 study of expert-role prompting adds that the effects are “highly conditional rather than universal” — role prompts reshaped the character of answers (deeper expertise framing, less clarity) rather than raising capability, and whether that trade helped depended on the domain and question type.

Why this matters now, and not just in the lab: the model behind your workplace AI changes without telling you. The silent engine swap is its own module here, and the software-engineering literature has documented for years that hosted models update silently enough that existing prompts need regression testing. When the engine changes, the context lines in your saved prompt (facts about you, your data, your format) mostly keep working — they’re information. The persona line is different in kind: it’s an instruction about who to be, tuned — whether you meant it or not — to the temperament of the model you wrote it on. A team prompt gallery standardised on last quarter’s model can be carrying personas that now suppress exactly the capability they were written to sharpen, and nothing on the screen will say so.

Two honest limits. The librarian numbers come from one industry research team’s study — two models, twelve coding tasks — and while the study is pre-registered, the headline persona numbers are its exploratory post-hoc finding, not the pre-registered endpoint. Treat them as a clean existence proof that the same persona can flip from harmless to harmful across models, not as a universal constant. And none of this says personas are useless: a role reliably steers tone and register, which may be exactly what you want for a client-facing draft. The claim is narrower and more practical — whatever your persona is doing for you, it is doing it on this model. On the next one, it may do something else.

Try it

Take one saved prompt you actually rely on — a custom instruction, a project prompt, a shared team agent — and split it:

1. Mark each line as one of:
   CONTEXT — facts about you, your data, your standards
   FORMAT  — how the answer should be structured
   PERSONA — who the model is told to be
2. Run a real task twice: once as-is, once with the PERSONA lines deleted.
3. Compare the two outputs against your own judgement of what good looks like.

If the persona-free version is as good or better, the persona was costing portability for nothing — cut it, or replace it with the concrete behaviour you actually wanted (“cite every claim”, “flag uncertainty”, “challenge my assumptions”); behaviour instructions travel between models far better than identities do. If the persona version is better, keep it — and re-run this same comparison next time your tool’s engine changes, because that’s the moment the persona line is most likely to have quietly changed sides.

Where it breaks: this test won’t catch a persona whose effect only shows on task types you didn’t test. The point isn’t a one-off audit — it’s knowing which lines in your saved prompts are model-specific, so you know what to re-test when the model changes.

Additional reading

• The Librarian Who Refused to Code — arXiv (19 July 2026) — pre-registered, though the librarian result is the paper’s exploratory post-hoc analysis; 480 completions across two frontier models and twelve coding tasks; the research-librarian persona triggered in-character disclaimers in 55/60 responses and twelve refusals on Claude Opus (correctness 0.92→0.67) with neither effect on GPT-5.5; personas as “model-dependent behavioral-policy biases rather than universal quality interventions.”
• When ‘A Helpful Assistant’ Is Not Really Helpful — Findings of EMNLP 2024 (November 2024) — 162 personas, 2,410 questions, four model families (FLAN-T5, Llama-3, Mistral, Qwen2.5): no aggregate improvement; more personas hurt on Llama3-70B while Qwen2.5 models were insensitive to all 162; the perplexity–accuracy correlation flips sign between families; automated best-persona selection often no better than random.
• When Does Persona Prompting Actually Help? — arXiv (28 May 2026) — 1,140 questions, 38 expert roles: role prompting traded expertise depth against clarity rather than improving capability; effects “highly conditional rather than universal,” helping advisory questions in some domains and hurting explanatory ones in others. Single-generator study — read it for the conditionality pattern, not the magnitudes.
• (Why) Is My Prompt Getting Worse? — arXiv (November 2023) — hosted LLM APIs update silently, degrading previously working prompts; the case for regression-testing your prompts across model updates — the maintenance habit this module applies to the persona layer specifically.

Editor’s note

Twelve months ago, persona crafting was part of conventional prompt engineering. I recall the expert-led classes I attended explaining that framing the persona enabled a model to more accurately interpret your expectations. Now, that idea is essentially thrown out entirely. Give the model your expectations; don’t ask it to infer them based on what you hope is a shared imagination of a character. It seems obvious now, but that’s only because models and our relationships with them have evolved. These modules are aimed at moving that evolution.