A system prompt is a set of instructions given to Claude at the start, before the actual conversation.

When you talk to Claude, you send messages, and Claude replies. But in many products, there is also a hidden layer of instructions that is placed at the start of every conversation. This is the system prompt. It does not come from the person chatting. It comes from whoever set up the product. It tells Claude things such as what role to play, what tone to use, what rules to follow, and what background information to keep in mind.

The system prompt and the messages from the user work together. The system prompt is like the stage setting and the actor’s briefing. The user’s messages are the lines of the play that unfold on that stage. The same actor, Claude, can perform very differently depending on the briefing it received, which is one reason Claude can power a friendly study buddy in one app and a formal legal research assistant in another.

flowchart TD
    A["System prompt: role, rules, background"] --> D["Claude reads everything together"]
    B["Earlier messages in the conversation"] --> D
    C["The user's latest message"] --> D
    D --> E["Claude writes the reply"]
        

A few simple words appear again and again in this topic. Here they are in plain English.

The Claude chat apps use a system prompt too, and Anthropic publishes it.

Anthropic’s release notes explain that Claude’s web interface and mobile apps use a system prompt to give Claude up-to-date information at the start of every conversation, such as the current date. The prompt also encourages certain behaviors, for example always providing code snippets in a clear, formatted way. Anthropic says this prompt is updated periodically as it improves Claude’s responses.

Anthropic has also made a point of publishing these app system prompts, in a release notes section of its documentation, along with a log of changes. This was a notable move toward transparency, because it lets anyone see what instructions shape Claude’s behavior in the apps. The release notes make an important clarification: these system prompt updates apply to the Claude apps and do not apply to the Claude API.

They give an assistant a consistent character and clear boundaries.

Without a system prompt, Claude is a general assistant that responds to each message as best it can. That is fine for open-ended chat. But a business or a developer usually wants something more specific: an assistant that always sounds like the brand, stays on topic, uses the right format, and knows the right background. A system prompt provides that consistency, so the assistant behaves the same way for every user, without anyone having to repeat instructions.

Giving Claude a clear role is the best-known way to use a system prompt.

Anthropic’s documentation says that when using Claude, you can improve its performance by using the system parameter to give it a role. It calls this technique role prompting, and describes it as the most powerful way to use system prompts with Claude. The right role, it says, can turn Claude from a general assistant into a virtual domain expert. The documentation lists benefits: improved accuracy in complex areas such as legal analysis or financial modeling, a tailored tone, such as a CFO’s brevity or a copywriter’s flair, and improved focus, because the role keeps Claude within the bounds of your task.

Strong system prompts tend to cover a handful of clear ingredients.

Deciding where an instruction belongs makes prompts cleaner and easier to maintain.

One practical question for builders is whether an instruction belongs in the system prompt or in the user’s message. Anthropic’s earlier guidance put it simply: use the system parameter to set Claude’s role, and put the rest, such as task-specific instructions, in the user turn. A helpful way to think about it is that the system prompt holds what stays the same in every conversation, and the user message holds what changes each time.

flowchart TD
    A["An instruction"] --> B{"Is it the same in every conversation?"}
    B -- "Yes" --> C["Put it in the system prompt"]
    B -- "No, it changes each time" --> D["Put it in the user message"]
        

A system prompt is read again and again, so every sentence should earn its place.

A good system prompt is built through testing, not guessing.

A system prompt shapes Claude’s behavior, but it works within limits.

A system prompt can set a role, a tone, and rules. It cannot make Claude drop its core principles. Claude is built to follow reasonable instructions from the people who set up a product, while still declining requests that would cause serious harm, deceive users in damaging ways, or break its basic guidelines. A system prompt telling Claude to ignore its safety behavior will not work, and it is not a good design goal.

Assume a determined user could find out what your system prompt says.

System prompts are not a safe vault. Users sometimes try to trick an assistant into revealing its instructions, which is called prompt leaking. Anthropic’s documentation includes a page on reducing prompt leak, among its guidance for strengthening guardrails. The safest approach is to assume that anything in the system prompt could eventually be seen, and to avoid putting anything there that would cause harm if it were exposed.

Everyday users have their own ways of giving Claude standing instructions.

You do not have to be a developer to benefit from the same idea. Claude’s apps offer ways for ordinary users to set standing instructions, which work much like a personal system prompt. These are not literally the same thing as the developer’s system prompt, but they play a similar role.

flowchart TD
    A["Anthropic's own instructions in its apps"] --> E["Claude's behavior in a conversation"]
    B["Your profile preferences and styles"] --> E
    C["Project instructions"] --> E
    D["Your messages"] --> E
        

Let us follow an imagined developer named Sana as she builds a study helper for school students.

Sana followed the full recipe: define success, draft, test, adjust one thing at a time, protect secrets, and keep improving. The system prompt became a living piece of the product, not a one-time guess.

Most system prompt problems come from a few avoidable habits.

System prompts are a simple, powerful tool with limits.

A checklist to keep beside you while you write.

Quick answers to the questions beginners ask most often.

A set of standing instructions given to Claude before a conversation begins, usually written by whoever builds or runs the product. It shapes the role, tone, rules, and background that Claude uses.

Yes. Anthropic’s release notes explain that the web and mobile apps use a system prompt to provide information such as the current date and to encourage helpful behaviors, and Anthropic publishes these prompts and logs their changes.

No. Anthropic states that the system prompt updates for its apps do not apply to the API. Developers write their own system prompts.

Giving Claude a role in the system prompt, such as a careful analyst or a patient tutor. Anthropic describes it as the most powerful way to use system prompts, improving focus, tone, and, in complex areas, accuracy.

Stable things such as role, rules, and standing background belong in the system prompt. The specific question, document, or task that changes each time belongs in the user message.

As short as the job allows. Start simple, and add rules only when testing shows they are needed. Long, rambling prompts bury important rules.

No. A system prompt can shape role, tone, and scope, but it cannot override Claude’s core principles and safety behavior.

Not normally, but they may try to trick the assistant into revealing it. Assume it could leak, and keep passwords and confidential details out of it.

Test it on realistic and tricky questions, compare against your definition of a good answer, change one thing at a time, and keep monitoring real conversations after launch.

Yes. Profile preferences, project instructions, and styles in the Claude apps work like personal standing instructions, and the same advice about clarity and specificity applies.

They help when the prompt has several kinds of content. Anthropic’s guidance recommends consistent, descriptive tags to keep instructions, background, and inputs from being mixed up.

In Anthropic’s prompting best practices and role prompting pages, the Console prompting tools, the guidance on strengthening guardrails, and the system prompt release notes.

A system prompt is a set of standing instructions that Claude receives before a conversation begins, shaping its role, tone, rules, and background. Anthropic’s own apps use one, publish it, and log its changes, but those updates do not apply to the API, where developers write their own. Role prompting is the best-known use, and Anthropic says even a single sentence can focus Claude’s behavior. A good system prompt is specific, organized, and short enough to stay clear, holds the things that stay the same in every conversation, and is built by testing and adjusting one change at a time. It cannot override Claude’s safety principles, it can leak, so it is no place for secrets, and it should be backed by other safeguards where mistakes would be serious. Everyday users can use the same idea through preferences, project instructions, and styles.