Every system prompt mixes two ingredients. A persona tells Claude who to be: "You are a meticulous financial analyst at a mid-market bank." An instruction tells Claude what to do: "Round all figures to two decimals; never speculate about unreleased earnings." Both are legitimate — Anthropic's own guidance endorses each — but they trade off differently on predictability, scope, and maintenance cost.
What a persona buys you
Per Anthropic's prompting best practices, giving Claude a role via the system prompt — even a single sentence — focuses behavior and tone. That is the persona's superpower: compression. "You are a careful pediatric nurse educator" implies hundreds of small behaviors — vocabulary level, caution around dosage, warmth of tone — that would take pages to enumerate as rules. Personas generalize to situations you didn't anticipate, because Claude fills gaps by asking "what would this person do?"
The cost is edge-of-scope unpredictability. A persona is an inference engine, and inferences at the boundary are yours to discover in production. Does the "friendly support agent" offer refunds? You never said; the persona decided. When a behavior must not vary — legal disclaimers, data-handling rules, escalation thresholds — inference is the wrong mechanism.
What explicit instructions buy you
Instructions are the opposite trade: verbose, narrow, and predictable. Anthropic's golden rule of clear prompting — if a colleague with minimal context would be confused by your prompt, Claude will be too — pushes toward explicitness: state exactly what to do, in what format, under what conditions. Rules are also auditable: a reviewer can read "never provide tax advice; direct users to a licensed professional" and verify it exists, which matters when compliance signs off on your prompt.
Two documented refinements make instructions stronger. First, explain the reason: Anthropic's guidance shows that adding context ("this output is read aloud by a text-to-speech engine, so never use ellipses") lets Claude generalize the rule instead of pattern-matching its letter. Second, don't rely on implication — the docs advise explicitly asking for "above and beyond" behavior rather than hoping the model infers it.
Choosing: a working rule
| Dimension | Persona ("you are a…") | Instruction (explicit rules) |
|---|---|---|
| Coverage | Broad — generalizes to unlisted cases | Narrow — covers exactly what's written |
| Predictability | Softer at the edges | High, rule by rule |
| Prompt length | One or two sentences | Grows with every requirement |
| Auditability | Hard to verify behavior from text | Each rule is checkable |
| Best for | Tone, judgment, domain framing | Format, compliance, boundaries, escalation |
In practice: layer them, in that order
Production system prompts almost always use both, structured so they don't blur. Anthropic recommends organizing prompt sections with consistent, descriptive XML tags — a natural fit here: a one-to-two-sentence role, then a tagged block of hard rules, then formatting requirements, then 3–5 examples in <example> tags showing persona and rules operating together. Examples are the reconciliation mechanism: they demonstrate how the warm persona and the strict refund rule coexist in one reply.
Keep personas honest, too. "You are the world's leading expert with 40 years of experience" adds no capability; it's decoration, and it can nudge outputs toward overconfident phrasing. A useful persona describes function and disposition ("a skeptical procurement reviewer whose job is finding gaps"), not credentials. And whichever mix you choose, test it: Anthropic's prompt-engineering overview is blunt that success criteria and empirical evals come before prompt tuning. When an eval fails, the diagnosis usually tells you which layer to fix — tonal drift points at the persona, boundary violations point at a missing or unclear rule.
Where to go next
Go deeper on each half with role prompting and system prompt design. When rules pile up, instruction hierarchy covers precedence, and negative instructions covers the "never do X" rules specifically.