Role & Persona Prompting That Actually Works
Multi-dimensional personas that shape voice, taste, and decision-making — not a frankenstein average of 10,000 examples.
“"Write like a CEO" gives you a frankenstein average of every CEO the model has ever read. "Write like Jensen Huang in 2024 after a quarterly earnings beat" gives you Jensen.”— SHE · YOUR AI GUIDE
Rumelhart's 1980 schema-activation theory predicted this 40 years before LLMs existed. The brain — and, structurally, every transformer trained on human text — stores knowledge as schemas: bundles of facts, voices, frames, and decision rules linked to specific identities. When you name a generic role ("CEO"), you activate the average across all CEO schemas in the model's training distribution. The output reads like the AVERAGE because mathematically it IS the average.
When you name a specific person, era, and context ("Jensen Huang in 2024 after a quarterly earnings beat"), you activate a tightly-clustered set of schemas with much lower variance. The voice sharpens, the decision frame locks in, the taste calibrates. The Rumelhart effect compounds: every additional dimension you specify (decision frame, anti-preferences, era, audience) narrows the activated cluster further.
The production move is the 4-axis persona spec. Most people stop at expertise ("You are a marketing expert"). Elite prompters specify all four: expertise (what they know), voice (how they sound), taste (what they reject), decision frame (the lens through which they evaluate). All four together compress 30% of your context into one role line — and the model behaves like the persona instead of pretending to.