What's this about?
Anthropic's release notes now include the system prompt for each Claude model, and the archaeology across versions turned out to be the interesting part. Early system prompts ran just over 300 words; current ones exceed 3,000. Simon Willison rebuilt the whole collection as a git commit history so the changes between models can be read as diffs.
What exactly is a system prompt?
It's a block of instructions placed before your conversation that the model reads first and that you don't see — the standing directions about who it is, what it should refuse, how long its answers should be, what today's date is. It's the difference between the raw model and the assistant you actually talk to. Publishing them matters because everything the model does that seems like personality or policy might be either trained-in behavior or just a line of text someone wrote — and until now, outsiders couldn't tell which.
What's the community saying?
- The routing disclosure was the standout find. Opus 5's prompt contains instructions explaining that the user may have selected Claude Fable 5 and been served Opus instead — a product behavior that was visible only in the prompt.
- The growth from 300 to 3,000+ words. Commenters read the expansion as the accumulated sediment of every edge case and incident, and noted the earliest prompts contained no child-safety guardrails at all, where current ones have several bullet points.
- "Why don't they bake it into the model?" A recurring question with real economics behind it: users pay for these tokens on every API call.
- Contradictions in the prompt. An experienced practitioner suggested asking Claude to enumerate the inconsistencies in its own system prompt, noting that contradictory instructions degrade quality even outside the area of the contradiction.
- The obligatory joke. "Claude keeps responses focused, brief, and concise to avoid overwhelming the person" — quoted, followed by "Claude and I must have a different idea of what brief means."
- A request aimed at open harnesses. Several asked why open-source harnesses like Pi or Opencode don't expose a customizable system prompt, since that's the natural surface for building a custom agent.