HeadsUpAI

Anthropic Details Best Practices for Reliable Claude Computer Use in Production

Anthropic released a production manual for developers building with computer use—the capability for AI to see and control desktop interfaces. The guide introduces specific resolution limits for the Claude 4.6 family and Opus 4.7, alongside a new Teach Mode where users record demonstrations that agents follow and adapt to live UI changes.
Claude 4.6 resolution limit
1.15 megapixels
Opus 4.7 resolution limit
3.75 megapixels
Recommended thinking effort
Medium for 4.6 or High for 4.7
Prompt cache breakpoints
4 total
Safety features
Built-in prompt injection classifiers

This update addresses the reliability gap that has kept Claude's computer use capability in research previews. By standardizing mechanical precision—such as pre-downscaling screenshots to prevent silent API degradation—Anthropic is formalizing the discipline of Agentic Engineering. It solves the near-miss problem where agents misidentify small UI targets.

You can implement these patterns immediately to reduce costs and improve accuracy. Recommendations include setting thinking effort to Medium for most 4.6 models and using prompt caching with a rolling buffer to manage history. Official computer use tools now include built-in, zero-latency prompt injection classifiers to protect agents from malicious web content.

View the full update on claude.com
ClaudeDevs
ClaudeDevs
@ClaudeDevs
X

Computer use turns Claude into an agent that can operate real UIs. New blog post on making it reliable in production: getting click accuracy right, choosing thinking effort levels, keeping long sessions within context, and recording demonstrations Claude can replay: https://t.co/JwSI3YDWcK

152retweets2klikes
View on X

Still wondering? A few quick answers below.

For the Claude 4.6 model family, Anthropic recommends a baseline resolution of 1280x720 pixels to stay within the 1.15 megapixel limit. For Claude Opus 4.7, the limit increases to 3.75 megapixels, allowing for 1080p resolution. Pre-downscaling screenshots to these limits is critical to ensure the model's click coordinates align with the actual image.

Teach Mode allows users to record a demonstration of a specific workflow, capturing screenshots and actions like clicks or typing. These recordings are fed back to Claude as context, including visual markers where the user clicked. Claude then uses this demonstration as a guide to perform the task, adapting its actions if the live interface has changed.

Anthropic recommends setting adaptive thinking effort to Medium for the Claude 4.6 family and High for Opus 4.7. Internal testing shows that Max effort often increases token costs and latency without providing a meaningful lift in accuracy for UI tasks, which are primarily perceptual and mechanical rather than deeply logical or mathematical.

Developers should use prompt caching with four specific breakpoints to reduce costs. A rolling buffer pattern is recommended to prune old screenshots in batches, which keeps the message prefix stable for the cache. For very long sessions, server-side compaction can summarize the history to reclaim context window space while preserving critical user instructions and progress.

Yes, the official computer use tool includes built-in prompt injection classifiers that run automatically on every request. These classifiers scan for adversarial instructions hidden in text or images without adding extra latency or cost. For custom tool implementations, Anthropic recommends using human-in-the-loop confirmations for high-stakes actions and limiting the agent's system permissions.

Share this update