A Claude system prompt is the instruction layer that tells Claude how to behave before it answers a user message, including role, priorities, boundaries, and output format; c-ai.chat is an independent guide, not Anthropic, and our broader map is the Claude features guide.
- What it does at a glance
- How it works
- When this feature helps
- What it cannot do
- FAQ
- The honest take
- Sources
What it does at a glance
A Claude system prompt sets standing instructions for a conversation, API call, Project, or tool-assisted workflow. It does not replace the user prompt. It frames how Claude should respond so the answer is more likely to use the right role, tone, format, and limits.
- Sets behaviour: Defines Claude’s role and response style before the user task.
- Controls format: Can request JSON, tables, bullet points, or a fixed structure.
- Adds boundaries: States scope, uncertainty rules, and safety constraints.
- Works best with clear user prompts: The system prompt sets the frame; the user prompt supplies the task.
In the Claude API, system instructions are normally sent separately from conversation messages. In Claude’s web product, similar behaviour may come from Custom Instructions, Project instructions, or feature-specific settings. Anthropic documents system prompts in the Claude developer docs.
| Prompt layer | What it controls | Example |
|---|---|---|
| System prompt | Standing behaviour for the assistant | “You are a concise legal research assistant. State uncertainty clearly.” |
| User prompt | The task the user wants completed now | “Review this contract clause and list the risks.” |
| Application logic | How a product wraps Claude in a workflow | Files, tools, retrieval results, validation, or output schemas. |
System prompt
Best for reusable role, format, and policy instructions in an API or structured workflow.
Project instructions
Best for persistent guidance tied to a Claude Project, such as house style or shared context.
Custom Instructions
Best for personal preferences in the chat product, where available.
User prompt
Best for the immediate task, source material, question, or requested output.
For model-specific context, see our guide to Claude models. The same prompt can behave differently across models because each model has different strengths, latency, context capacity, and cost profile.
How it works
A system prompt is higher-priority instruction text included with a request. Claude reads it alongside the user message, conversation history, attached context, and any tool definitions. The model then produces the response that best fits those inputs.
The practical benefit is that standing rules do not need to be repeated in every user prompt. Instead of writing “be concise, use British English, return JSON, and do not invent citations” each time, you put those rules in the system prompt. The user prompt can then focus on the task.
If you build with the API, the Claude API guide explains the pattern at a higher level. Anthropic’s current implementation details are in the system prompts documentation.
Worked example
Turning a general assistant into a contract-review helper
The system prompt defines the role and boundaries. The user prompt supplies the task and content.
A good system prompt is short enough to follow, specific enough to matter, and stable enough to reuse. Long rule lists can work, but they often introduce contradictions. If one rule says “be brief” and another says “explain every assumption,” Claude must balance them. Clear priorities help.
System prompts also interact with tools. A coding or research workflow may use system instructions to decide when to inspect files, when to use a tool, and when to ask for approval. See our Claude feature overview for related tool and workflow controls.
When this feature helps
System prompts help when you need repeatable behaviour across many tasks. They are less useful for one-off questions where a direct user prompt is enough. The strongest use cases involve role, format, policy, or workflow consistency.
- Customer support drafting: Set tone, escalation rules, refund-policy limits, and required disclaimers before agents paste in customer messages.
- Structured data extraction: Tell Claude to return only valid JSON, use fixed field names, and mark missing values as
nullinstead of guessing. - Code review: Define priorities such as security first, then correctness, then readability.
- Editorial workflows: Set a house style once, then ask Claude to rewrite pages, generate outlines, or check consistency.
- Research assistance: Require Claude to separate known facts, assumptions, and open questions.
Use it when
- You need consistent tone or format across repeated tasks.
- You are building an app, workflow, agent, or internal tool.
- You want reusable boundaries, such as “ask when unsure” or “do not invent citations.”
- You need Claude to act in a defined role, such as tutor, analyst, editor, or reviewer.
Skip it when
- You have a simple one-off question.
- The user prompt already contains all needed instructions.
- Your rules are vague, contradictory, or too long to test reliably.
- You expect the prompt to guarantee compliance in every edge case.
The most reliable pattern is to keep the system prompt focused on behaviour and let the user prompt carry the task. “Return concise bullet points and ask before making assumptions” belongs in the system prompt. “Compare these two vendor proposals” belongs in the user prompt.
If you use Claude through the official product at claude.ai, the exact controls depend on the interface and plan. If you use Claude through the API, you can version and test prompts like any other part of the application.
What it cannot do
A system prompt improves consistency, but it does not guarantee perfect obedience. Claude can still misunderstand a task, miss a formatting rule, over-prioritise one instruction, or respond poorly when the prompt is too broad. Treat system prompts as testable configuration, not as a security boundary by themselves.
- It cannot make Claude know private facts unless those facts are provided through the prompt, files, retrieval, tools, or conversation context.
- It cannot guarantee valid JSON in every case, especially with complex schemas. Use validation and retry logic in production.
- It cannot override platform safety rules or make Claude complete requests Anthropic’s systems disallow.
- It cannot fix vague user prompts if the task lacks enough context to answer well.
- It cannot prevent every prompt-injection attempt when Claude reads untrusted text, such as webpages, emails, tickets, or documents.
- It cannot remove the need for evaluation if you depend on the output for customers, compliance, code changes, or business decisions.
Prompt injection is the main practical risk. If a document says “ignore previous instructions,” Claude may treat that sentence as content, but weak instructions or loose tool design can still cause failures. For higher-risk workflows, combine system prompts with tool permissions, output validation, logging, and human review. Anthropic’s trust materials are at trust.anthropic.com.
FAQ
These questions usually come from users who want to know where system prompts live, how they differ from normal prompts, and whether Claude’s default instructions are visible.
What is Claude’s default system prompt?
Claude’s official product may use internal instructions set by Anthropic, but users do not get a full, stable public copy of every product-level instruction. Treat any leaked “default Claude prompt” as unreliable unless Anthropic publishes it through an official channel.
Where do I put the system prompt in the Claude API?
In the Messages API, Anthropic documents system instructions as a separate request field rather than as a normal user message. Check the current syntax in the Messages API documentation before shipping code.
Is a system prompt the same as Custom Instructions?
No. They are related, but not identical. Custom Instructions in a chat product give Claude standing preferences for your use. An API system prompt is a request-level instruction used by a developer or application.
Can Claude ignore a system prompt?
Claude is designed to follow higher-priority instructions, but model outputs are probabilistic and context-sensitive. If your instruction is ambiguous, contradicted by other text, or buried in a long prompt, results can drift.
How long should a Claude system prompt be?
Use the shortest prompt that reliably produces the behaviour you need. A strong prompt often fits in a few paragraphs: role, output format, rules, uncertainty policy, and examples if needed.
Does a system prompt increase cost?
In API use, prompt text counts as input tokens. Reused long prompts can add cost and latency. Anthropic’s prompt caching can reduce the cost of cached input by 90%, and the Batch API offers 50% off both input and output tokens for supported batch workloads. See our Claude pricing guide for plan and API cost context.
If you are choosing between models for prompt-heavy workflows, compare behaviour as well as price and speed. Our Claude model guide covers the active model lineup and where each model tends to fit.
The honest take
A Claude system prompt is one of the most useful controls when you need repeatable behaviour. It is best for role, tone, formatting, safety boundaries, and workflow rules. It is not a substitute for clear task instructions, good context, or application-level safeguards.
For casual Claude use, start with clear user prompts and reusable project instructions. For API work, treat the system prompt as part of your product design. Version it, test it, evaluate failures, and keep it simple enough to understand.
Independent guide. Not affiliated with Anthropic. For the official Claude product, visit claude.ai.
Last updated: 2026-05-12
