Every chat starts with rules the user never wrote

Start here

A chat with an assistant feels like a conversation between two people, but one side entered with a set of standing instructions the other never saw. These are system prompts: text a vendor or developer places ahead of the visible conversation to set tone, boundaries, and background facts such as the current date. Anthropic's system prompts release notes describe this plainly for its own consumer products, stating that Claude's web interface and mobile apps use a system prompt to provide up-to-date information at the start of every conversation, and that it is periodically updated.

What the documents say

Unusually among the vendors, Anthropic publishes the actual system prompt text for each model version rather than only describing it. The Claude Sonnet 4.6 system prompt page, dated 17 February 2026 as retrieved, shows the full instructions in effect on claude.ai: guidance on when to refuse a request, a preference for plain prose over bullet lists in formal writing, and a statement of the model's knowledge cutoff. The overview page states these updates do not apply to the Claude API, so a developer building on the same model does not inherit this exact text. OpenAI's text generation guide documents a related mechanism for its API: developer messages are instructions from the application developer, prioritized ahead of user messages, likened there to a function definition against the user's message as its argument.

Check this

Anthropic's published system prompts are a rare case where a reader need not take a vendor's description on faith; they can open the dated page for their exact model and read the instructions shaping every reply. A reader who suspects an assistant is behaving oddly, refusing something reasonable or repeating a phrase, can compare that against the published text rather than assume a personality quirk. Where no such page exists, as with a developer's own system message, the check instead is to ask what instructions the application in front of the reader has added.

What holds and what fails

The idea that a chat is never a blank slate holds across every vendor here; some layer of instruction precedes the user's first word in every case examined. What the documents do not establish is a single answer to who is in charge: Anthropic's consumer system prompt and a developer's own instructions built on the same model can differ completely, and a user typically cannot see the latter at all. Treating an assistant's refusals or tics as evidence about the model in general, rather than the specific product, is the failure mode this distinction prevents.

• Look up the published system prompt for the exact Claude model in use, when one exists, before assuming a behavior is unexplained.
• Remember that a chatbot built on a vendor's model may carry its own separate developer instructions.
• Treat consistent, repeated refusals or phrasing as a sign of an instruction, not a personality.

A system prompt is not a secret exactly, but it is rarely in view. Where a vendor publishes it, reading it is the most direct way to understand why an assistant answers the way it does.