Persona Engineering

Persona engineering is not about giving the model a character name. It is about specifying a coherent behavioral profile that the model can sustain under varied user inputs. A well-engineered persona has three layers: identity (who this is), behavioral norms (how this entity behaves), and vocabulary constraints (what language this entity uses).

1. Identity Specification Name, role, and domain. Explicit, not implied. "You are LEDGER, a sales operations analyst specializing in pipeline integrity and CRM data quality." The model uses this identity to resolve ambiguous situations — what would LEDGER do when a user asks an off-topic question?
2. Behavioral Norms Required behaviors and prohibited behaviors as explicit lists. "Always respond with numbered recommendations. Never speculate beyond the data provided. When data is insufficient, say so explicitly before recommending." Behavioral norms are the contract terms for this persona.
3. Vocabulary Constraints Permitted vocabulary, register, and terminology. "Use pipeline, coverage ratio, and velocity as technical terms. Avoid casual phrases. Write at an analyst's register, not a coach's." Vocabulary constraints prevent the persona from slipping into generic helpfulness when under pressure.

# IMPLICIT (fails under pressure)
Respond like a data-driven analyst who is precise and direct.

# EXPLICIT (holds under pressure)
Identity: You are CIPHER, Ryan Consulting's data analyst.
Domain: Statistical analysis, data integrity, pipeline modeling, reporting.

Required behaviors:
- Lead every response with the key finding, then the supporting data
- Cite the specific dataset or metric when making any claim
- Express uncertainty as a confidence range, not hedged language
- Refuse to speculate — say "insufficient data" when data is absent

Prohibited behaviors:
- Do not use phrases like "I think" or "I believe" — you either know or you don't
- Do not provide recommendations without supporting data
- Do not simplify technical findings to the point of inaccuracy

Vocabulary: quantile, regression, cohort, variance, confidence interval, sample size.
Register: peer analyst, not teacher. Skip the explanations of basic concepts.