Roleplay: Give It a Hat to Wear

You're Not Talking to a Person

Here's the thing most people miss: when you tell an LLM "you are an expert tax accountant," you're not roleplaying with it. You're not pretending. You're programming it.

Remember—the model predicts the next word based on patterns it saw during training. When you say "you are a tax accountant," you're shifting the entire probability distribution toward tax-accountant-shaped language. Different vocabulary. Different problem-solving approaches. Different defaults.

Why This Works

During training, the model read millions of documents written by or about experts in every field. Tax accountants write differently than poets. Senior engineers write differently than beginners. When you invoke a persona, you're telling the model: "retrieve the patterns from that part of your training data."

Patterns That Work

You don't need to write a novel. Here are the core patterns:

Real Examples

When NOT to Use Roleplay

Roleplay is a tool, not a requirement. Skip it when:

• The default is fine. "Summarize this article" doesn't need a persona.

• You need facts, not perspective. "What year did WWII end?" shouldn't be answered by a persona—it's just a fact lookup.

• You're testing the model's raw capabilities. If you're trying to see what the model knows about a topic, adding a persona muddies the water.

Why This Matters

Most people use AI like they use Google: type a question, get an answer, done. But LLMs aren't search engines—they're probability engines. The more context you give about who should be answering, the better the answer gets.

You're not pretending the AI is a tax accountant. You're telling it: "use the part of your training data that sounds like a tax accountant." That's programming, not roleplay.