- Author
- Simon Willison
- Published
- 18th April 2026
- Kind
- Personal blog post, with a diff of published prompts and one tool-listing test
- Original
- simonwillison.net/2026/Apr/18/opus-system-prompt/
- Subject
- Harnesses and agent loops
What it is
Simon Willison writes a weblog on AI tools. The post compares two Claude.ai system prompts. It then lists the chat tools Claude reports.
What it says
- The post says Anthropic is the only major lab publishing chat system prompts.
- The 4.7 prompt renames the developer platform and lists browser, spreadsheet and slides agents.
- The child safety section is expanded, with extreme caution required after a child safety refusal.
- A new section says to attempt vague requests now and use tools before asking.
- New text asks Claude to respect a user who ends a chat.
- Rules on asterisk actions and certain words were removed. A disordered eating section was added.
- Tool descriptions are not published, so the author asked Claude to list its tools.
What evidence it gives
The post quotes passages from the git diff and offers the author's own highlights. It links the Claude Code prompt used to build the history. It links a shared transcript of Claude's tool list. It does not show the full diff inline.
Interests and limits
This is the author's selection of changes, so omitted changes are unknown. The author interprets removals, such as guessing the new model no longer misbehaves. Tool descriptions come from asking the model, not from Anthropic. The author believes the tool list is unchanged since 4.6.