3-Step Setup for Claude Computer Use Browser Automation

What you'll learn in this article:

• How to identify your best first browser automation candidate
• The 3-part prompt structure that prevents errors
• How to add safety gates without slowing down your workflow
• What to document so the workflow is repeatable by anyone on your team
• The most common failure modes and how to fix them

Step 1: Identify Your Best First Candidate

The best first browser automation is not the most exciting workflow you have. It is the most expensive one, measured in minutes per week. Most executives have one task that costs 30 to 60 minutes every week that they have stopped thinking about because it has always just been part of the job. That is the right starting point.

Run your first candidate through this quick checklist. If it passes all four, it is automation-ready:

• You do the same sequence of steps at least twice per week.
• The web pages involved load consistently without frequent layout changes.
• A mistake can be corrected before it causes permanent damage (nothing sends emails or deletes data automatically).
• The task involves a web application you access through Chrome.

Common candidates that pass: CRM data entry from email, exporting reports from analytics dashboards, staging social media posts in scheduling tools, copying invoice data from email attachments into accounting software.

Step 2: Write the 3-Part Prompt

Every effective computer use prompt has three sections. Skipping any section increases failure rate significantly.

Part A , Context: What is currently open on screen and where the data is coming from. Claude needs to know what it is looking at before it takes any action.

Part B , Task sequence: Exactly what to do, in order. Be specific about app names, field names, and the expected outcome of each step. Vague verbs ("update the CRM") cause more retries than precise ones ("find the contact by first name in HubSpot, click into their record, and paste the text into the Notes field").

Part C , Safety gates: When to stop and wait for your approval. Always include at least one stop-before-submitting instruction and one show-me-what-you-plan-to-do instruction.

3-PART PROMPT TEMPLATE

[Part A , Context]

"I have [App A] open showing [what is visible]. The data I want to transfer is in [source: email / spreadsheet / PDF / URL]."

[Part B , Task sequence]

"Read [source]. Extract [specific fields]. Open [App B]. Navigate to [specific location in App B]. Enter [extracted data] into [specific fields]. Do not click Submit/Send/Save yet."

[Part C , Safety gates]

"Before entering anything, list exactly what you plan to enter and where. After filling all fields, stop and show me the completed form before saving."

Step 3: Document the Working Prompt

The most overlooked step: when the workflow runs cleanly for the first time, save the exact prompt that worked. Do not paraphrase it. Copy it as-is into a shared doc, Notion page, or your company's standard operating procedure. Label it with the task name, the app it targets, and the date it was tested.

This does three things. First, you never have to rediscover the prompt when the workflow needs to run again. Second, anyone on your team can run the workflow. Third, when the app updates and the workflow breaks, you have a documented baseline to diagnose and fix from.

🎯 Action Item

Pick one browser task this week. Run it once with full oversight. If it works, save the prompt and run it again tomorrow. Two successful runs in a row means the workflow is ready to use without close monitoring.

Common Failure Modes and Fixes

Step 4: Turn the Working Prompt Into a Skill

Once a workflow runs cleanly twice in a row, it is ready to become a Skill. A Skill is a saved instruction file that Claude loads on demand , the full 3-part prompt, safety gates, and any task-specific notes, stored in one place. Instead of typing or pasting the prompt at the start of every session, you trigger the Skill by name and Claude reads the instructions automatically.

The difference matters for three reasons. First, token efficiency: Claude loads the Skill from a file rather than receiving it in the conversation, which keeps the session context leaner and reduces cost on repeated runs. Second, consistency: the exact same instructions execute every time, with no risk of accidentally rephrasing a safety gate or omitting a step. Third, team distribution: a Skill file can be shared with anyone on the team who uses Cowork, meaning any team member can trigger the workflow without knowing how it was built.

How to Create a Skill File

A Skill is a markdown file with a specific structure. Create a new file in your Cowork skills folder (inside your workspace's .claude/skills/ directory) and name it after the task it performs. The file needs two things: a frontmatter block that tells Claude what the Skill is for, and the body that contains the full prompt and instructions.

SKILL FILE TEMPLATE

--- name: crm-contact-logger description: Reads the open email thread in Mail, extracts contact details, and logs them into HubSpot. Stops before saving for human review. --- ## Task Instructions [Paste your complete working 3-part prompt here exactly as tested. Include all safety gates and field names.] ## When to Use This Skill Run this skill after every prospect email exchange that needs to be logged in HubSpot. The email thread must be open in Mail before triggering. ## Known Edge Cases - If the contact does not yet exist in HubSpot, create a new record before logging notes. - If the email contains multiple senders, log only the primary prospect, not CC'd parties.

Once the file is saved, you invoke it in Cowork by simply telling Claude to use it: "Run the CRM contact logger Skill." Claude reads the full instruction set, confirms what it plans to do, and executes. The session starts clean and focused rather than spending the first exchange re-establishing context from a long pasted prompt.

📊 Why This Compounds Over Time

A workflow you run 3 times per week saves time three times per week. A Skill you run 3 times per week saves time plus reduces per-session token overhead every time. After 30 days, the difference between a Skill and a pasted prompt is: faster invocation, lower cost, no drift in instructions, and the workflow is distributable to any team member who opens Cowork. One hour to document and save the Skill pays for itself within the first week of use.

The Full 4-Step Workflow

Get the full newsletter every week

This article is part of the PromptHacker Weekly premium deep-dive series. Subscribe to get all 7 articles plus the newsletter every Wednesday. Click here now