XNews

Prompting best practice: Ask AI to simulate domain experts

Ask LLMs to simulate experts, not 'you'. - Andrej Karpathy advised treating LLMs as simulators, not entities. - Avoid asking LLMs “what do you think?” because “you” invokes a default persona embedding. - Instead ask which real-world experts or stakeholders are most appropriate. - Prompt template suggested: ask the model to find the best-suited people and simulate their answers. - The approach is presented as yielding richer, multi-perspective responses than default-persona prompts.

Key insights

Treat LLMs as simulators, not entities

Andrej Karpathy's guidance emphasizes that large language models function as powerful simulators rather than autonomous entities. The posts state that asking 'what do you think' causes the model to adopt a generic persona via learned personality embeddings rather than producing perspectives grounded in specific domain expertise. The recommended practice is to reframe prompts to request simulations of particular expert groups or stakeholders so the model can generate multiple, role-specific perspectives.

Avoid the default 'you' persona to get deeper answers

The posts warn that directly addressing an LLM as 'you' often triggers a default, averaged persona from the model's training data. That default response pattern can be serviceable but tends towards generic answers. Changing prompt framing to ask 'which experts would best answer this and what would they say' is presented as a method to elicit more targeted, varied, and substantive responses.

Practical prompt template: request best-suited respondents

A practical prompt phrase circulated: 'Help me find the people in the world best suited to answer this question; how would they answer it?' This template is suggested as a straightforward way to get the model to enumerate relevant expert roles and to generate answers from those perspectives, replacing vague role assignment or default-persona prompts.

Source posts