35 min read
Three ways to see it
A model that has read everything cannot, at the same time, sound like everyone. Without an instruction about who to be, it defaults to a kind of generic, slightly American, slightly customer-service voice that does not match a Pakistani office. The fastest way to fix this is to tell the model who it is. That single instruction, often called role prompting or persona prompting, shifts the output more than any single trick except chain-of-thought. It is also one of the most misunderstood. People in Pakistani training rooms hear "role prompting" and start writing "you are an expert in everything". This produces worse output, not better. This lesson teaches role prompting at the level of actual control.
Way one: the expert role. You assign the model a specific professional identity. "You are a senior FBR tax officer with twenty years of experience reviewing returns from textile exporters in Faisalabad." Notice the precision. Not "you are a tax expert". Not even "you are an FBR officer". A specific seniority, a specific institution, a specific domain. The more specific the role, the narrower the distribution of style and content the model will produce, and the more it will sound like that role. A vague role gives a vague output. A precise role gives an output that survives the test of being read by an actual FBR officer.
Way two: the audience role. Here you do not tell the model who it is, you tell it who it is talking to. "Explain this clause of the new SECP company law to a small business owner in Sialkot who never studied past matric." The same explanation, with the same legal content, will come out in plain Urdu with no jargon. Switch the audience to "a partner at a Karachi law firm" and the same prompt produces dense legal English with case references. The model is doing the same intellectual work; the audience tag is choosing which register, which vocabulary, which level of background assumption it uses. For internal training documents in Pakistani organisations, audience role often outperforms expert role.
Quick check
Quick check: what makes modern AI different from a rule-based program?
The why-tree
Why role prompting works. The model has seen vast amounts of writing tagged with the kind of person who produced it. When you assign a role, you are pulling the model towards the cluster of writing that matches that role. The model still uses all its knowledge, but it filters its expression through the style of the named role. Role does not change what the model knows. It changes how the model speaks.
Sources
Sources. OpenAI's prompt engineering guide, section "Adopt a persona". Anthropic's documentation on giving Claude a role. Lilian Weng's prompt engineering blog post, section on role and instruction prompting. The 2023 paper "Role-Play with Large Language Models" by Murray Shanahan et al., which gives the cognitive science framing for why role prompts work.