A team wants Claude to write technical documentation in a specific voice. They've tried writing detailed style instructions ("be concise, use active voice, avoid jargon, write at a 9th-grade reading level, prefer short paragraphs, use code examples") with mixed results — the model follows some rules and forgets others, especially as documents get longer.
What's the right addition?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Role prompting integrates style guidelines into a coherent identity. When the model is "writing as Stripe documentation," it draws on patterns it associates with that style — concise, active voice, well-structured, code-forward. Six isolated rules produce isolated rule-following ("oh right, I need to use active voice" — but only when it remembers). A role lets the rules cohere. This is one of the most powerful and underused prompt engineering techniques: when style matters, name a persona that embodies the style.
Adding rules has diminishing (and eventually negative) returns. The model has limited attention; 30 rules guarantees that some are ignored on any given output. Roles compress style into a single coherent target.
Temperature controls output randomness, not which rules the model follows. Lowering temperature might make wrong-style output more consistent — exactly the opposite of what you want.
Smaller models follow nuanced style instructions worse, not better. They produce more variance in style and miss more rules. The constraint capability scales with model size.