Role Prompting & Anthropomorphic Steering

Telling an LLM "You are a senior software architect with 20 years of experience" is often dismissed as pseudo-scientific superstition by critics. However, mechanistic interpretability research demonstrates that role conditioning narrows the vocabulary activation distribution to match high-quality technical corpora.

To maximize the effect of persona prompting: 1. Don't just assign a title; assign an **epistemic philosophy**. (e.g. "You favor boring, battle-tested distributed patterns over trendy microservice frameworks"). 2. Explicitly outline the persona's skepticism and review methodology.