Simon Willison dropped a system prompt diff

Simon Willison dropped a system prompt diff analysis that the HN thread barely noticed. 53 points. That is a shame because this is the most useful piece of release documentation Anthropic has shipped in months, and it came from a researcher scraping the actual model instead of reading the marketing page.

The Real Release Notes Live In The System Prompt

The difference between Opus 4.6 and 4.7 system prompts reveals 15-plus changes that affect every single conversation you have with the model. Most of them are not in the official documentation.

Claude now lists new integrated tools that 4.6 never mentioned. Claude in Chrome (browser agent), Claude in Excel, Claude in Powerpoint, and something called Claude Coworker. These are not listed on Anthropic's main product page. They are in the system prompt. The model knows about them and will reference them in certain contexts.

More practically: Claude now has a tool_search mechanism that checks available tools before telling you it cannot do something. If you ask Claude to look up a file, a calendar event, or location data, it checks whether those tools are actually loaded in the current session before refusing. This is a real behavior change. If you have been working around Claude's "I cannot access" responses for specific tools, that workaround behavior might be outdated in 4.7.

The Child Safety Change Has Real Workflow Implications

This one should get more attention than it is getting. The child safety section expanded significantly. If Claude refuses one request for child safety reasons in a conversation, every subsequent request must now be approached with extreme caution. One refusal colors the rest of the session.

For agency operators building multi-turn automation that handles user-generated content, this matters. If your automation accepts user inputs and passes them to Claude, and one of those inputs triggers a child safety refusal, the rest of that conversation session runs under a modified safety constraint. You might not notice it in testing. You might notice it in production when Claude starts declining requests that should be straightforward.

Build your automation to isolate turns. Do not let a single user input lock down an entire session. The behavior change is documented in the system prompt even if it is not in the marketing.

Claude Got Quieter And More Confident

Two changes worth noting together. The system prompt added an instruction to be less verbose. At the same time, it removed restrictions on emotes and removed the instruction to avoid words like genuinely, honestly, and straightforward.

These changes partly cancel each other out. Less verbose means shorter responses. The removed restrictions mean Claude might use stronger language when it wants to. In practice, what I have seen is that 4.7 feels more direct than 4.6 but occasionally too clipped on complex tasks where you actually wanted the longer explanation.

If your workflows depend on Claude producing thorough, long-form output, test whether the verbosity instruction is affecting output quality on your specific task types. The instruction is in the system prompt so it affects the base model behavior before any of your custom instructions apply.

The new disordered eating section and screenshot attack protection are narrow safety additions that most agency workflows will never encounter. They are worth knowing exist so you can recognize the behavior if it shows up unexpectedly.

Anthropic Is The Only Lab Doing This

Every other major AI lab keeps their system prompts internal. Anthropic publishes theirs. Simon Willison has been archiving them and running diffs, which means you can actually track how model behavior changes between versions instead of guessing from marketing copy.

This is genuinely useful and underused. The system prompt is the real specification for how the model behaves. When something changes in production and you cannot explain why, the answer is often in the system prompt diff, not in the feature announcements.

Bookmark the Simon Willison research repository. Subscribe to his posts on model updates. Treat system prompt diffs as the real release notes for every Anthropic model update.

One Action To Take This Week

Pull your production conversations from the past two weeks and look for anything that broke after you moved to Opus 4.7. Specifically look for cases where Claude stopped asking clarification questions, started refusing requests that should have been straightforward, or gave you shorter answers than you expected.

If you find patterns, check the system prompt diff for the corresponding instruction. There is a good chance the behavior change is documented there even if it was never announced.

Anthropic ships the real changelog in the system prompt. Start reading it.

Sources: - Simon Willison system prompt analysis: https://simonwillison.net/2026/Apr/18/opus-system-prompt/ - HN discussion: https://news.ycombinator.com/item?id=47823270 - GitHub diff archive: https://github.com/simonw/research/commit/888f21161500cd60b7c92367f9410e311ffcff09