The Persona Frame - Replacing Forty-Seven Rules With One Voice

The problem

A retail bank's complaint-response assistant had accumulated forty-seven rules over eighteen months. Never say "sorry for the inconvenience." Always mention the case number. If the complaint mentions a wire transfer, cite Reg E. Every rule traced back to a real incident, someone got burned, someone wrote a rule, the rule went into the prompt. That's how rule lists actually grow in production. Nobody sits down and designs forty-seven rules on day one. They accrete, one incident-driven patch at a time, until the prompt reads like a legal team's revenge on plain English.

I tested that assistant against twenty complaints. Ten fell into categories the rules explicitly covered. Ten were novel, categories the rule list never anticipated, because the incidents that would have produced those rules hadn't happened yet. One of the novel ones was a mobile-check-deposit bug complaint, a category that didn't exist when most of the forty-seven rules were written. On the covered categories, the assistant performed the way you'd expect from forty-seven carefully tuned instructions. On the novel ones, it fell apart in one of two ways. Sometimes it defaulted to generic, flat boilerplate that technically responded to the complaint without actually engaging with it. Sometimes it invented a rule-sounding constraint that wasn't real, a fabricated policy that read exactly like the other forty-seven and was wrong in a way nobody could catch just by skimming the output.

The gap between how the assistant handled covered categories and how it handled everything else is the actual failure signature. A rule list is a lookup table dressed up as judgment. It handles exactly what it was built to handle and offers nothing coherent once the input steps outside that table, because nothing in forty-seven discrete instructions establishes an underlying register the model can fall back on. There's no voice underneath the rules. There's just the rules, and the moment a complaint doesn't match one, the model is reasoning from nothing.

The pattern

The fix I tried was almost insultingly simple on paper. Replace the forty-seven-rule enumeration with a short persona: you are Maren, a senior complaints resolution specialist who has handled thousands of cases. You're calm, specific, and never defensive. You always name the concrete next step and the case number. That's it. That's most of the prompt now.

A handful of rules stayed explicit rather than folding into Maren's voice. The Reg E citation requirement is the clearest example. Tone and judgment can generalize to a complaint category nobody wrote a rule for, because a coherent persona carries an implicit sense of how to handle an unfamiliar situation the same way an experienced person would. A legal citation requirement can't generalize that way. Either the complaint mentions a wire transfer and Reg E applies, or it doesn't, and that's not a judgment call a voice can make on its own. So the compliance-critical rules stayed as explicit constraints layered on top of the persona, and everything else, the register, the tone, the instinct for what a real complaints specialist would say next, moved into the persona itself.

What I actually measured was instruction count more than any specific accuracy percentage: roughly ten total instructions running the persona version against the baseline's forty-seven. On the novel categories, the ones the original rule list had no answer for, the persona version closed most of the gap. Maren doesn't need a rule for the mobile-check-deposit complaint because Maren has an implicit sense of how a senior specialist responds to a customer with a confusing, half-explained problem: acknowledge what's actually wrong, name a concrete next step, give the case number, don't get defensive. None of that required anticipating the specific complaint category in advance. It required a voice consistent enough that the voice itself does the generalizing.

The instruction-count comparison is the load-bearing part of this for me. Forty-seven rules is forty-seven places for an edit to introduce a regression, forty-seven things a new hire has to internalize before touching the prompt, forty-seven potential conflicts between rule twelve and rule thirty-one that nobody notices until a complaint triggers both at once. Ten instructions, nine of them a persona description and one an explicit compliance rule, is a system a person can actually hold in their head.

Design considerations

Here's the part I don't want to soften. Independent research published since I started leaning on this pattern complicates it without invalidating it, and the complication matters enough that I check it every time I reach for a persona frame now. A study called "Playing Pretend: Expert Personas Don't Improve Factual Accuracy" found that persona framing can reduce factual accuracy on knowledge-heavy tasks even while it improves how well the output matches what a human expects the tone to sound like. A separate paper, "The Prompt Makes the Person(a)," found the effect is highly sensitive to exactly how the persona gets worded, which means two people trying the same technique with slightly different phrasing can get meaningfully different results and neither one is doing it wrong. Further work on persona-conditioned reliability and on persona effects in strategic-reasoning tasks adds to a picture that's now fairly consistent across several independent groups: persona effects are task-dependent, sometimes negative, and never free.

None of that breaks what I did with Maren, because the design already kept the fact-sensitive, compliance-critical instruction, the Reg E citation, explicit and outside the persona's voice. The persona was never carrying the factual weight. It was carrying tone, register, and the judgment calls that a rule list can't anticipate in advance. But I'd be lying if I said the caveat doesn't narrow where I'll reach for this. Persona framing is validated, in my own testing and in the literature, for tone and register generalization. It is not something I'd trust for knowledge-heavy factual accuracy, and the current research leans negative on exactly that use case.

That distinction is the calibration question I sit with before building one of these. What part of the task is actually judgment and voice, the kind of thing an experienced specialist would handle by feel, and what part is a fact that's either true or false regardless of who's speaking. A complaint's tone response is judgment. Whether Reg E applies to a wire transfer complaint is a fact. Fold the judgment into a persona and it generalizes. Fold the fact into a persona and the compliance requirement now depends on how consistently the model stays in character, which is a bad thing to leave riding on vibes. The rule list problem was that everything lived in one undifferentiated bucket, forty-seven flat instructions with no distinction between voice and fact. The fix is to sort all forty-seven before deciding what the persona has to carry.