Claude publishes its opening instructions: Why the same AI can act differently
Anthropic（AN-throp-ik）
The AI company that makes Claude.
system prompt
Instructions a service gives an AI before a user’s message.
Claude API
A way for developers to use Claude in their own software.
What happened
Anthropic, the company behind Claude, points people to a public record of the system prompts used by Claude’s web service and mobile apps. A system prompt is an instruction given to an AI before a person starts typing. Anthropic says these prompts can provide current information, such as the date, and encourage response habits, such as using Markdown for code examples.
The page was shared on Hacker News, where it received 587 points and 241 comments. Those figures show attention from that community. They do not prove that the source material is correct. The claims in this article come from Anthropic’s documentation.
The background: instructions arrive before a question
When a person asks an AI to explain a topic, the visible question is not always the first instruction in the conversation. The service can also send instructions that set context or formatting expectations. That earlier instruction is the system prompt.
Anthropic says it periodically updates the prompts for Claude’s web interface and mobile apps to improve responses. This matters because a model and a system prompt are not the same thing. The model is the part that generates language. The prompt is a set of directions used in a particular product setting. Changing those directions can change how an answer is presented, even when the model name stays the same.
Why it matters
Making these instructions easier to inspect gives users more context for judging an AI answer. A reply may reflect both the person’s request and directions supplied by the service.
It also draws a useful line for developers. Anthropic states that these system-prompt updates do not apply to the Claude API. In other words, an experience in the Claude website or phone app should not be assumed to match an application that calls Claude through the API.
That distinction can help when people compare results. A difference between two Claude experiences may come from the product context, not only from the underlying model.
What is confirmed
Anthropic’s documentation confirms that the web interface and mobile apps use a system prompt at the start of each conversation. It gives the date and code-formatting behavior as examples of what the prompt can do. The company says the prompt is updated from time to time.
The page also says that, starting with the Claude 4.6 generation, each model ID refers to one fixed snapshot. Therefore, those model IDs have one listed entry on the page.
What remains unknown
The document does not say how much any one update changes everyday answers. It also does not establish that every Claude surface will show the same behavior in every situation. The practical effect of a change would need separate testing under specific conditions.
What to watch next
The key question is whether future updates remain easy to find and compare. Users can also check which Claude product they are using before drawing conclusions from a response. An AI’s behavior is shaped by more than a model label: the instructions placed before the conversation can matter too.
Claude has instructions before you type
📰 Full story: Claude publishes its opening instructions: Why the same AI can act differently
Claude reads service instructions before it reads your question.
system prompt
A starting instruction for an AI chat.
Claude API
A tool that lets an app use Claude.
💡 The gist
- Claude gets instructions before each chat.
- The web and phone apps use them.
- Claude API users do not automatically get those updates.
Anthropic makes Claude, an AI assistant. Before Claude reads your question, it can read a system prompt. A system prompt is a private starting instruction from the service.
It can tell Claude today’s date. It can also guide how Claude shows code. For example, it can ask for Markdown formatting.
This is different from the AI model itself. The model produces the words. The system prompt guides the chat setting. So a changed prompt can change an answer’s style. The model name may still stay the same.
Anthropic says it updates these prompts sometimes. The company says this helps improve responses. The prompts described on this page are for Claude’s website and mobile apps.
They do not automatically apply to Claude API. An API lets developers use Claude in their own apps. So a developer’s app may not behave exactly like the Claude website. That difference does not automatically mean something is broken. The two products may start with different instructions.
The documentation page was popular on Hacker News. It got 587 points and 241 comments there. That tells us people in that community noticed it. It does not prove every claim on the page. For the product facts, Anthropic’s documentation is the source to check.
We still do not know how much each prompt update changes answers. The page does not measure that effect. It is still useful to remember this: an AI reply can be shaped by your question and by the instructions it received first.
💬 Do long AI instructions help?
Claude starts with a set of rules. People are debating why those rules have grown and whether that is good.
- A user comparison says the rules grew from about 300 words to more than 3,000. The added material appears to cover safety and newer information.
- One view is that rules grow after problems are found, like safety rules for buildings. Another warns that too many rules, especially conflicting ones, can make answers worse.
- People disagree on the effect on performance. Some users report that long rules use valuable context; others say “performance” might mean answer quality, not speed, and the technical cost is more complicated.
- Keeping rules as text makes them easier to change for different products and tasks. Reusing the same opening text through caching may reduce repeated work.
- Safety should not depend on the rules alone. Commenters describe using training and separate checks before and after the main model as well.
initial digest at 241 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
Claude reads a first note
📰 Full story: Claude publishes its opening instructions: Why the same AI can act differently
Claude reads a note before it talks.
Anthropic（AN-throp-ik）
The company that makes Claude.
system prompt
Claude’s first note.
Anthropic makes Claude.
Claude reads a first note. It reads it before your question.
The note is called a system prompt.
It can say today’s date. It can say how to show code.
Claude on the web uses this note. Claude on phones uses it too.
The note can change sometimes.
Claude API does not get these changes. An API helps other apps use Claude.
Hacker News talked about this page. Many people looked at it there.
That does not make every claim true. We check Anthropic’s page for the facts.
💬 The AI rule book
An AI reads a rule book before it talks.
- People say the rule book has become much longer. It includes more safety notes.
- A longer book may help. But some people worry it can confuse the AI.
- The rule book is not the only safety helper. Other checkers can help too.
initial digest at 241 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.
💬 Debating Claude’s system prompts
The thread discusses the growing size of published prompts, possible effects on quality and cost, where safeguards belong, and why prompts remain editable. Much of it is user observation or inference, not confirmed causation.
initial digest at 241 comments (revision 1). We fetched 100 comments and sampled 100 across the thread. These are HN users’ reports, not independently verified facts.