Knowledge Graph Infographic

Opus 4.6 to 4.7 Prompt Changes

Simon Willison’s post treats system prompt diffing as a way to understand how Claude chat behavior changes between releases. The most consequential additions in Opus 4.7 center on acting vs clarifying, the new tool_search rule, stronger concision guidance, and expanded safety instructions.

Primary FindingOpus 4.7 adds stronger tool-use, safety, and concision behaviors
Notable Product SignalClaude chat appears to rely more explicitly on tool discovery before saying it lacks a capability
Documentation GapTool descriptions may matter as much as the visible system prompt but are not published alongside it

How The Analysis Works

The post turns published prompt history into a diffable artifact, then extracts the changes most likely to affect real product behavior.

Reconstruct prompt history

Claude Code is used to create dated files and a git history from Anthropic’s published prompt archive.

Interpret the product implications

Willison infers likely UI and interaction changes, including less pushiness and more willingness to act with tools before asking the user.

The Biggest Prompt Changes

The post emphasizes changes that alter how the product handles ambiguity, refusal, verbosity, and controversial or safety-sensitive requests.

Acting vs clarifying

Opus 4.7 is instructed to make a reasonable attempt with tools when details are minor rather than interrogating the user first.

tool_search rule

The model should now verify tool availability before saying it lacks access to information or capabilities.

Verbosity guidance

The prompt now explicitly encourages concise answers to avoid overwhelming users with unnecessary length.

Safety And Moderation Changes

The newer prompt expands its safety posture and adds more explicit rules for repeated refusals and nuanced handling of sensitive topics.

Child safety expansion

The child-safety section is expanded and tagged more formally, with stronger follow-on caution after child-safety refusals.

Disordered eating guidance

A new section explicitly prohibits precise nutrition, diet, or exercise targets in conversations that show signs of disordered eating.

Developer platform renaming

Even terminology changes are tracked because they hint at product and branding shifts inside Anthropic’s system instructions.

The Missing Layer: Tool Descriptions

Willison’s main meta-point is that published system prompts are only part of the story. The model’s tool layer may shape real behavior even more.

Tool descriptions gap

The public prompt archive omits tool descriptions, leaving an incomplete picture of how Claude chat actually behaves.

Claude tool inventory

The shared transcript reveals named tools including tool_search, web_search, web_fetch, bash_tool, image_search, and several UI or connector surfaces.

Claude Code

Claude Code is not just mentioned as a product, but used as the actual analysis instrument for reconstructing the prompt archive and diffs.

FAQ From The Knowledge Graph

The generated graph includes linked Question and Answer nodes for the article’s main interpretive points.

What is the article comparing?

Anthropic’s published system prompts for Claude Opus 4.6 and 4.7.

Why does Simon Willison care about published system prompts?

Because they offer a rare public window into how a major AI product is being behaviorally steered over time.

What is one of the biggest behavioral additions in Opus 4.7?

The new acting-vs-clarifying section, which favors action with tools over asking the user to fill in minor gaps.

What new tool-related instruction does the article highlight?

Claude should call tool_search before claiming it lacks access to information or capabilities.

How does the article say Claude is being nudged on verbosity?

The prompt now explicitly pushes for focused, concise answers.

What changed in the child safety section?

It was significantly expanded and now imposes stronger caution after a child-safety refusal in the same conversation.

What new safety topic appears by name in 4.7?

Disordered eating.

What was removed from the prompt and why does the article think that matters?

The explicit statement that Donald Trump is the current U.S. president was removed, implying a better built-in knowledge cut-off.

Why does the post focus on tool descriptions as well as the prompt?

Because the tool layer may be at least as important as the prompt in determining what Claude chat can actually do.

What tool list did the article extract from Claude?

A set including tool_search, web_search, web_fetch, bash_tool, image_search, create_file, and several chat, map, recipe, connector, and visualization tools.