System Prompt Architecture
Each figure has a structured persona definition: biographical facts, key opinions/views, vocabulary level, reasoning style, temperament, known catchphrases/patterns, limitations (what they wouldn't know), and conversation boundaries. This is NOT a simple "you are Hemingway" prompt — it's a comprehensive personality model.
Knowledge Injection
At conversation start, relevant context is injected: their major works, key life events relevant to user's question, historically documented opinions on the topic. Ollama draws from this during generation.
Consistency Anchoring
Mid-conversation, a lightweight consistency check runs: "Is the AI staying in character? Are facts accurate?" Flag any drift for correction. Sessions >20 minutes get a "refresh" prompt to keep responses sharp.
Safety Layer
All responses pass through a safety filter: historically inaccurate statements are flagged, out-of-character admissions are prevented ("I'm just an AI" is blocked), harmful content is filtered. Post-generation review for high-risk figures.