LLMs, GenAI, RAG, agents, evaluation · insight-0535

Supply context instead of personas, because role prompts do not add knowledge

Situation
A prompt starts by telling the model it is a world-class revenue-management expert, and the team expects better pricing analysis as a result.
Action
Replace or supplement the persona with the actual inputs (booking curves, fare rules, constraints) and the decision criteria; test any persona against a no-persona control on the eval set.
Reason
A large study found adding personas to system prompts did not improve performance on factual questions relative to no persona, and the effect of a given persona was largely random.
Boundary
Analysis, Q&A and advisory prompts.

Symptom or risk

The output sounds authoritative but uses no route data, so tone is mistaken for expertise.

Personas can usefully set tone, audience or reading level, which you should evaluate separately.

Real context costs tokens and data plumbing.

Keeping a persona because one demo looked better.

Zheng et al. found across 162 roles and four model families that adding personas to system prompts did not improve performance on factual questions versus a no-persona control.
When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models