See updates to the core system prompts on [claude.ai](https://claude.ai) and the [Claude iOS app](https://anthropic.com/ios) and [Claude Android app](https://anthropic.com/android).
interesting that it uses html notation to identify sections, not Markdown which is supposed to be “LLM-friendly” (eg, skills)
interesting that the docs refer to Claude entirely in the 3rd person. “Claude defaults to helping”, not “You are a helpful …”
Claude is implied to be singular yet multiple: “ask the Claude in that application”
whatever is prompting Claude is referred to as “the person” but in reality could well be another bot
given that context window is fixed and this system prompt takes a fixed overhead for every single call, I am surprised at how repetitive and wordy it is. I would have expected a document that a dozen good editors went over to make as parsimonious as possible
A few random notes :
pseudo-xml tags are better at making context clear.
But XML is already a meta languange. Pseudo-XML? Where does it end with you people?
Real XML documents have a well defined schema.
It was an attempt at a joke, sir.