Every Messages API request can carry a system parameter: instructions that sit above the conversation and apply to everything Claude does in that request. In a demo, a one-line system prompt is fine. In production — where hundreds of users hit the same prompt with inputs you didn't anticipate — the system prompt becomes a policy document, and it deserves the same care. Everything here applies identically whether you call Claude first-party or through Amazon Bedrock, Google Vertex AI, Microsoft Foundry, or Claude Platform on AWS: the models, and therefore the prompting behavior, are the same.
What belongs in the system prompt
Anthropic's guidance is to treat Claude like a brilliant but brand-new employee. The golden rule from the official prompting best practices: if a colleague with minimal context would be confused by your prompt, Claude will be too. A production system prompt typically covers four things:
Role. Who is Claude in this application? Even a single sentence — "You are a claims-triage assistant for an insurance back office" — measurably focuses behavior and tone. (More on this in role prompting.)
Task and scope. What should it do, and — just as important — what is out of scope? Be explicit about "above and beyond" behavior rather than implying it; Claude follows stated expectations more reliably than assumed ones.
Context with reasons. Explain why an instruction matters. The official docs give a memorable example: telling Claude output will be "read aloud by a text-to-speech engine, so never use ellipses" works better than a bare "no ellipses" — Claude generalizes from the explanation to cases you didn't enumerate.
Format rules. Output structure, length expectations, and what to do when the input is malformed or the answer is unknown.
Structure it so Claude can parse it
Long system prompts benefit from the same discipline as long documents: wrap distinct sections in named XML tags — <role>, <rules>, <output_format>, <examples> — using consistent, descriptive tag names, and nest tags where there's hierarchy. This is Anthropic's documented recommendation, and it makes the prompt easier for your own team to diff and review too. See structuring prompts with XML tags for the full pattern.
Failure modes that only show up in production
The prompt that silently breaks caching. Prompt caching is a strict prefix match on exact bytes. A timestamp, request ID, or user name interpolated into the system prompt invalidates the cache on every request, and you pay full input price without any error telling you so. Keep dynamic values out of the system prompt, or place them after the cached prefix.
The prompt that grew by accretion. Contradictory rules added over months ("always be concise" next to "always explain your reasoning in detail") produce inconsistent output. Review the prompt as a whole when you add a rule, and delete rules that no longer apply.
Relying on retired mechanisms. Prefilled assistant responses — starting Claude's answer for it to force a format — are no longer supported starting with the Claude 4.6 generation; such requests return a 400 error. Move format control into system-prompt instructions or, better, structured outputs.
Assuming the eval failure is a prompt problem. The official prompt-engineering overview is blunt about prerequisites: clear success criteria, empirical tests against them, and only then prompt iteration. Sometimes the fix is a different model, not a longer prompt.
One platform note
On Claude Opus 4.8 only, the API additionally accepts mid-conversation {"role": "system"} messages, letting an operator inject instructions after cached history without invalidating the cached prefix. Other models return a 400 if you try this — so don't build cross-model tooling around it. The portable pattern remains a single top-level system parameter.
Where to go next
Pair this with reusable prompt templates to keep system prompts versioned and testable, and see the feature matrix for what each platform supports.