claude-code-computer-use

Claude Code Computer Use

Run Claude Code as a bounded UI-control agent that observes screenshots, takes one action, re-checks state, and repeats until success or stop conditions.

Required loop

1. Restate the user goal in one sentence.
2. Define a max step budget (default 20).
3. For each step, do exactly one UI action.
4. After each action, capture a fresh screenshot.
5. Verify expected change before continuing.
6. Stop on success, ambiguity, or budget exhaustion.

Tool contract to enforce

Require these tool primitives (or closest equivalents):

• screenshot()
• click(x,y) / double_click(x,y)
• type(text) / key(combo)
• scroll(delta)
• Optional: ocr(region) / accessibility_tree()

Never take multiple UI actions in a single reasoning step.

Safety policy

• Allowlist app/site scope before starting.
• Ask before destructive actions (delete, send, purchase, publish, submit).
• If target is unclear for 2 consecutive steps, pause and ask human.
• Log every step with: intention, action, result.

Claude Code invocation pattern

Use non-interactive Claude Code mode for deterministic logs:

claude --print --permission-mode bypassPermissions "<task prompt>"

Task prompt template:

You are controlling a UI via tools.
Goal: <goal>
Allowed scope: <apps/sites>
Max steps: <N>
Rules:
- One action per step.
- After each action, screenshot and verify.
- If uncertain for 2 steps, stop and ask.
- Ask before destructive actions.
Return a step log and final status: success | blocked | needs-human.

Completion criteria

Declare completion only when:

• Goal state is visible in screenshot evidence, or
• You are blocked and provide the exact blocker + required human input.

Always return concise evidence:

• Final status
• Last 3 step logs
• Screenshot/file references (if available)