# Computer Use Best Practices

> Instructions

- Skill: `uygnoey/computer-use-best-practices` (Agent Skill, multi-file: 11 files)
- Install (CLI): `npx skillmds@latest add uygnoey/computer-use-best-practices`
- Raw SKILL.md: https://api.skillmd.com/api/skills/uygnoey/computer-use-best-practices/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: uygnoey (https://skillmd.com/u/uygnoey)
- Updated: 2026-09-21
- Page: https://skillmd.com/skills/uygnoey/computer-use-best-practices

---


## Instructions
Use this skill when you are implementing an agent that operates a UI via screenshots (computer use / browser use) and you need reliability, cost, and safety guardrails.

### 1) Screenshot sizing (downscaling)
- Downscale screenshots before sending them to the API.
- Start with 1280×720 as a simple default; for Opus 4.7 start at 1080p.
- Avoid extremely low resolutions (below ~960×540) and avoid sending native 4K.
- If you want maximum fidelity within limits, compute a “max API fit” resolution that preserves aspect ratio.

Implementation helpers:
- `scripts/screenshot_pipeline.py` (prepare_screenshot + compute_max_api_fit)

### 2) Coordinate scaling
If you resized the screenshot before sending it, scale the API-returned coordinates back to native screen space before executing the click.

Implementation helpers:
- `scripts/screenshot_pipeline.py` (scale_coordinates)

### 3) Message ordering
When sending a text instruction with an image, put the text block before the image block in the message content.

### 4) Small targets
- Use zoom for dense UIs (tool config).
- Prefer keyboard navigation (tab, shortcuts) for tiny targets.
- If you control the UI, increase target size via scaling/zoom.

Reference:
- `references/small-targets.md`

### 5) Thinking effort defaults
Use model-appropriate defaults for computer use:
- Opus 4.7: default `high`
- Claude 4.6 family: default `medium`

Reference:
- `references/thinking-effort.md`

### 6) Prompt injection safety
Treat all web content as untrusted:
- Use built-in prompt injection classifiers via `computer_20251124`.
- Require human confirmation before irreversible actions.
- Scope permissions and log actions.

Reference:
- `references/prompt-injection-safety.md`

### 7) Long-running context management
- Use cache breakpoints: one on a stable prefix, plus up to three on recent tool results.
- Prune old screenshots in batches; consider server-side compaction for very long sessions.

Implementation helpers:
- `scripts/cache_helpers.py` (set_trailing_cache_control)
- `scripts/prune_helpers.py` (prune_old_screenshots)
- `templates/compact_prompt.txt` (example compaction prompt)

### 8) What did not help (avoid wasting time)
See `references/optimization-myths.md`.

## Examples
### Example: prepare screenshot + send instruction before image
See `examples/python_api_example.md`.

### Example: enable zoom for dense UIs
See `examples/enable_zoom_snippet.md`.

## Source
https://claude.com/blog/best-practices-for-computer-and-browser-use-with-claude