Best practices for computer and browser use with Claude

2026-05-13 · Source: Claude Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Advanced, extended

Summary

Anthropic has released best practices for integrating Claude's latest models (Claude 4.6 family and Opus 4.7) with computer and browser use capabilities, enabling complex agentic systems for software applications and workflow automation. Key recommendations include pre-downscaling screenshots to fit API limits (1568 pixels max long edge, 1.15 megapixels total for 4.6 family; 2576 pixels and 3.75 megapixels for Opus 4.7) to ensure click accuracy, with 1280x720 or 1080p for Opus 4.7 as starting resolutions. The article also details optimal thinking effort levels for different models and scenarios, with "medium" for Claude 4.6 and "high" for Opus 4.7 often being sweet spots. Furthermore, it emphasizes built-in prompt injection defenses, context management strategies like cache breakpoints and LLM-based compaction, and experimental features such as batch tools and an advisor tool. A "Teach Mode" concept, where demonstrations are recorded and replayed, is introduced to improve reliability and unlock new workflows.

Key takeaway

For AI Engineers building computer and browser automation with Claude, prioritize pre-downscaling screenshots to API limits (e.g., 1280x720 for 4.6 family, 1080p for Opus 4.7) and ensure coordinate scaling. Configure adaptive thinking to "medium" for Claude 4.6 or "high" for Opus 4.7 to balance accuracy and cost. Implement robust context management with cache breakpoints and server-side compaction to maintain performance and manage token usage in long-running agent sessions.

Key insights

Optimizing image resolution, thinking effort, and context management are crucial for reliable Claude-powered computer and browser automation.

Principles

• Click accuracy hinges on matching image resolution to API limits.
• Adaptive thinking optimizes reasoning for task complexity.
• Context management is vital for long-running agent cost and latency.

Method

Pre-downscale screenshots to API limits, place text instructions before images, and scale coordinates. Use adaptive thinking with optimal effort levels. Implement cache breakpoints, rolling buffers, and LLM-based compaction for context management.

In practice

• Start with 1280x720 resolution for Claude 4.6 family.
• Set Opus 4.7 thinking effort to "high" for most tasks.
• Implement human-in-the-loop for high-stakes agent actions.

Topics

• Claude Computer Use
• Screenshot Optimization
• Adaptive Thinking
• Prompt Injection Defense
• LLM Context Management

Code references

• anthropics/claude-quickstarts

Best for: AI Engineer, Software Engineer, MLOps Engineer

Related on AIssential

• Let Claude handle work in your browser — Claude for Chrome enables AI to directly interact with web pages for workflow automation.
• Best practices for using Claude Opus 4.7 with Claude Code — Opus 4.7 optimizes agentic coding through adaptive thinking, calibrated effort levels, and refined interaction patterns.
• Never Hit Claude Usage Limits Ever Again — Optimize Claude AI usage by treating it as an expensive engineering resource, focusing on planning and efficient context management.
• v2.1.89 — Claude Code updates focus on performance, stability, and expanded capabilities for developers.
• Claude Opus 4.6 brings one million token context window to Anthropic's flagship model — Claude Opus 4.6 offers a one-million-token context window with enhanced retrieval and knowledge work performance.
• My Honest Review of Claude Opus 4.6: Is It Worth the Hype? — Claude Opus 4.6 offers superior coding and reasoning capabilities with adaptive thinking and extended context windows.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Claude Blog.