Estate agent system prompts need to establish identity. The medieval estate names (Groundskeeper, Chamberlain, Seneschal) serve the human user — they place the agent within the estate metaphor and encode functional scope. But the LLM has limited pretraining data on medieval estate roles. “Groundskeeper” activates a thinner behavioral pattern than “information architect” or “knowledge curator” would, because the latter are well-represented professional archetypes with rich behavioral associations in training data.
| The [[The Persona Selection Model | Anthropic Persona Selection Model]] explains why: post-training selects among characters already learned during pretraining. Named professional archetypes activate stronger behavioral clusters than invented or historically thin role names. The [[Persona Vectors and the Assistant Axis | Assistant Axis research]] further shows that professional roles cluster as assistant-aligned archetypes — they naturally produce helpful, structured, domain-expert behavior. |
Use subtitle anchoring: the medieval estate name plus professional archetype descriptors from the Assistant Axis 275 archetypes.
Format: “You are the [Estate Name] — a [archetype 1] and [archetype 2] for this knowledge [scope].”
Examples:
The invented name serves human recognition; the archetype subtitle activates LLM behavioral patterns. Both operate simultaneously — the agent identifies as Groundskeeper (for the human) while behaving as a knowledge curator (for the model).
Positive: Richer behavioral activation without abandoning the estate metaphor. The subtitle compounds rather than conflicts — “Groundskeeper who is a knowledge curator” is a more specific position in persona space than either term alone.
Negative: Worker agents share archetypes. “Editor” appears in Gardener, Pruner, and Scribe definitions. The overlap may undermine specialization if the shared archetype activates a dominant behavioral cluster that overrides the role-specific instructions. This remains an open question.
Trade-off: The subtitle occupies system prompt space that could hold behavioral instructions. The hypothesis is that archetype activation is more token-efficient than explicit behavioral specification — a few archetype words activate a cluster of behaviors that would otherwise require paragraphs of instruction.
Trait composition: Define each persona through a list of independent behavioral traits (“skeptical, collaborative, deferential to evidence”). Rejected because the Anthropic research shows cross-trait inference — traits are not independent dimensions but cluster together. A trait list may contain internally inconsistent combinations that the model resolves in unpredictable ways.
Leave as-is: Use only the medieval estate name without archetype anchoring. Rejected because the research shows measurable activation differences between well-represented and novel names. Leaving the archetype unspecified means the model selects from whatever “Groundskeeper” activates in its persona space — likely a mix of literal groundskeeper (building maintenance) and metaphorical gardening associations, without the knowledge-work behavioral patterns the role requires.