Claude Sonnet 5.5 arrives with a prompt to stop agents from building things before you approve

# What Sonnet 5.5 is and where it fits Claude Sonnet 5.5 is Anthropic's updated mid-tier model in the Claude 5.5 family. It's available on the web, the Claude desktop app, and through platform integrations such as Amazon Bedrock and Snowflake Cortex AI. Anthropic positions Sonnet between Opus 5.5 (higher-end) and Haiku 5 (faster, more basic), and makes Sonnet the default choice for free users.

# Why people will notice Sonnet 5.5 Anthropic's internal benchmarks show Sonnet 5.5 matches or comes close to Opus 5.5 in specific areas: agentic coding and desktop/computer-use tasks. Opus still has an edge on "complex, open-ended work requiring sustained judgement," but Sonnet 5.5 aims to deliver many everyday productivity and coding tasks at lower cost and faster speed. Vendors including Amazon and Snowflake announced availability, emphasizing speed, cost-per-task advantages, and integration with platform controls.

# The prompt that prevents wasted usage

In tests cited by PCWorld, Sonnet 5.5 typically returned text-only responses to open-ended prompts, but sometimes generated a chart anyway. Anthropic's recommended prompt reduces that risk by restricting the model's actions to ideation or planning unless the user later authorizes artifact creation.

# Where to apply the prompt The guidance is most useful when applied in places that control default assistant behavior:

• In custom instructions for the user-facing chat so the model consistently avoids automatic builds.
• In developer-side files like CLAUDE.md for Claude Code, which govern coding agents' behavior.
• As an opening line in any new chat when you expect only ideas or plans.

This is a small change in wording but it directly changes the model's action pattern: instead of "do and show," the model responds with "options or plan, await permission."

# Practical effect for users If you use Claude for spreadsheets, code review, or prototyping, this prompt helps avoid accidental usage spikes. It's particularly relevant when an assistant can generate artifacts (charts, slide decks, or runnable code). Implementing the instruction reduces the chance the assistant creates files you didn't ask for and that would count against usage or compute.

# When Sonnet 5.5 makes sense Choose Sonnet 5.5 for focused coding tasks, common knowledge work, document drafting, and interactive agent tasks where cost and speed matter. If you need the highest performance on open-ended tasks that demand sustained judgement, Opus 5.5 remains the recommended option.

# Bottom line