From docs/THOUGHTS.md — a real system prompt captured from the demo's LLM inspector:
"You are a personality-neutral radio DJ writing live station patter. …"
"Backstory: A quirky DJ, full of talent…"
"Style: bubbly, energetic, expressive"
The neutral boilerplate and the persona's Style block pull in opposite directions — at best the model ignores one, at worst it splits the difference into blandness. Audit the prompt template layering so the base prompt is persona-agnostic scaffolding (format rules, no-emoji, no stage directions) and personality comes only from the persona block.
Second part: consider exposing the effective system prompt (or its editable parts) in the Admin UI — the LLM inspector already shows it per call; a settings-level view would let operators see what their persona edits actually assemble into.
From docs/THOUGHTS.md — a real system prompt captured from the demo's LLM inspector:
The neutral boilerplate and the persona's Style block pull in opposite directions — at best the model ignores one, at worst it splits the difference into blandness. Audit the prompt template layering so the base prompt is persona-agnostic scaffolding (format rules, no-emoji, no stage directions) and personality comes only from the persona block.
Second part: consider exposing the effective system prompt (or its editable parts) in the Admin UI — the LLM inspector already shows it per call; a settings-level view would let operators see what their persona edits actually assemble into.