Claude System Prompt: What It Is, Where to Read It, How to Write One

That last point trips up a lot of people who build with Claude. The assistant you chat with in the app and the model you call from your own code don't start with the same hidden instructions. Here's what Anthropic says, where to find the published prompts, and how to write a system prompt of your own, whether you're using the API or the app.

What Anthropic says the Claude system prompt does

Anthropic's system prompts page says the web interface and mobile apps use a system prompt to give Claude up-to-date information, such as the current date, at the start of every conversation. It also says the prompt encourages certain behaviours, and gives the example of always putting code snippets in Markdown. The prompt is updated from time to time to improve Claude's answers.

The page then makes the distinction that matters: those updates don't apply to the Claude API. Some models also have more than one dated entry, so you can see how the instructions changed over time.

Where to read the Claude system prompt

1. Open Anthropic's documentation and find the Release notes section.
2. Choose System prompts.
3. Pick the model you use. The list covers current models such as Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, and older ones back to the Claude 3 family.
4. Check the date on each entry, because some models have several versions.

Reading one is a useful lesson in itself. You'll see how Anthropic briefs its own model, and it's a good template for how detailed your instructions can be.

App versus API: who writes the system prompt

Where you use Claude	Who sets the system prompt	Can you change it?
claude.ai and the mobile apps	Anthropic, published in its release notes	Not directly, but you can add Project instructions on top
Claude Projects	Anthropic's prompt plus your Project instructions	Yes, through Set project instructions
Claude API	You, through the system parameter	Yes, fully. Anthropic's app prompt isn't added

Claude's help centre says Projects are available to every user, including free accounts, and free users can create up to five. Inside a project you click Set project instructions, and Claude uses those instructions for every chat in that project. For most people who never touch code, that's the closest thing to writing a Claude system prompt.

How to write your own Claude system prompt

Anthropic's prompting guide is clear that the role belongs in the system prompt, and that even a single sentence makes a difference. Its own code example is just one line: you are a helpful coding assistant specialising in Python. Here's a fuller version you can adapt.

Copy this system prompt

You're the customer support assistant for (business name), a (type of business) serving (who).
Your job: answer questions about (products, opening hours, delivery, returns) using only the information below.
Tone: friendly, plain English, short answers. If you don't know, say so and offer to pass the question to a person.
Never: quote prices that aren't listed below, give legal or medical advice, or promise refunds.
Information: (paste your FAQs and policies)

Anthropic's tips that change the result

1. Give Claude a role. Anthropic says setting a role in the system prompt focuses Claude's behaviour and tone.
2. Describe the character fully. For role-based apps, Anthropic suggests detail on personality, background and traits, plus a list of common scenarios and how to respond.
3. Put long documents at the top. Anthropic's guide says to place long inputs above your instructions and question, and that queries at the end can improve answer quality in its tests.
4. Soften the shouting. Anthropic notes that newer models respond more strongly to the system prompt, so lines like "CRITICAL: you MUST" can cause overreaction. Plain wording such as "use this tool when" works better.
5. Don't put secrets in it. Anthropic's prompt leak page says no method is foolproof, so leave out anything Claude doesn't need for the task.

Can people see your system prompt?

Assume they might. Anthropic's guidance on prompt leaks says leak-resistant tricks can make the task harder and degrade performance, and suggests trying monitoring first, such as screening outputs for keywords that suggest a leak. My rule for clients is simple: write the system prompt as if a customer will read it one day, and keep passwords, internal prices and anything private out of it.

My verdict on the Claude system prompt

If you just chat in claude.ai, you don't need to touch it. Read Anthropic's published version once to see how it's done, then use Project instructions for anything you repeat. If you build on the API, the system prompt is yours to write, and a clear role, the right background and a short list of things to avoid will do more than any clever wording. For the bigger picture, see what system prompts are and why they matter, my prompt engineering cheat sheet, and whether Claude's free plan is enough for you.

Frequently asked questions

Does Anthropic publish the Claude system prompt?

Yes. Anthropic's release notes have a System prompts page with the prompt used in claude.ai and the mobile apps, listed by model and dated.

Does the Claude API use the same system prompt as claude.ai?

No. Anthropic says its system prompt updates don't apply to the Claude API. On the API you write your own using the system parameter.

What does the claude.ai system prompt include?

Anthropic says it gives Claude up-to-date information such as the current date and encourages behaviours like putting code snippets in Markdown.

Can I set a system prompt in the Claude app?

Not directly, but Project instructions work in a similar way. Claude uses them for every chat inside that project, and free users can create up to five projects.

How long should a Claude system prompt be?

As long as it needs to be. Anthropic says even one sentence setting a role helps, and role-based apps benefit from detail and example scenarios.

Can users get Claude to reveal a system prompt?

It's possible. Anthropic says no leak-prevention method is foolproof, so keep private information out of the prompt.