Claude Computer Use for Admin Tasks: The Right Way

I spent 47 minutes last Tuesday doing nothing but copying property details from email threads into my CRM, formatting them, and pasting the cleaned data into a spreadsheet for my quarterly market report. Forty-seven minutes. On a task that requires zero thinking. That’s the kind of work that slowly kills a solo operation — not the hard stuff, but the endless, brainless repetition that eats your afternoons.

Claude’s computer use feature is Anthropic’s answer to that problem. It lets Claude actually operate your computer — move the mouse, click buttons, type into fields, read what’s on screen — rather than just generating text you then have to act on yourself. I’ve been testing it seriously since early 2026, and this tutorial walks you through exactly how I set it up and use it for real estate admin work. No hype, no vague promises. Just the actual steps, the real results, and the parts where it fell flat.

What You’ll Build by the End of This Tutorial

By the time you finish reading and following these steps, you’ll have Claude computer use configured and running at least one automated admin workflow. Specifically, you’ll know how to set it up through the API, write instructions Claude can actually follow, and apply it to repetitive tasks like data entry, file organization, and copy-paste work between apps. I’ll use my real estate context throughout, but the logic applies to any solo business with dumb, recurring admin work.

Prerequisites Before You Start

Before step one, make sure you have these in place:

• An Anthropic API account with access to Claude 3.5 Sonnet or higher (computer use is not available on the standard Claude.ai chat interface — you need API access)
• Basic comfort running Python scripts or using a tool like Cursor or VS Code — you don’t need to be a developer, but copy-pasting a script shouldn’t terrify you
• A Mac or Linux machine — Anthropic’s computer use demo works best here; Windows users can run it inside a Docker container (I’ll mention this in the troubleshooting section)
• Docker installed if you want to use Anthropic’s pre-built demo environment (recommended for beginners)
• A clear, specific admin task in mind. Vague tasks produce bad results. “Help me with emails” will fail. “Open Gmail, find all emails with the subject line ‘Viewing Request’, copy the sender name and property address into column A and B of this spreadsheet” will work.

Budget note: API usage costs vary by model and volume. For moderate admin tasks — say, 30 minutes of computer use sessions per week — expect to spend roughly $15–$40/month depending on how complex your instructions are. This is not a free tool.

Step 1 — Set Up the Anthropic Computer Use Demo Environment

Anthropic provides an open-source demo that gives Claude a sandboxed desktop to operate in. This is the safest way to start. It runs inside Docker so Claude can’t accidentally delete your files or mess with your real system.

Open your terminal and run:

git clone https://github.com/anthropics/anthropic-quickstarts.git
cd anthropic-quickstarts/computer-use-demo
export ANTHROPIC_API_KEY=your_api_key_here
./setup.sh

Once the Docker container is running, you’ll get a browser URL — usually http://localhost:8080 — that shows you a virtual desktop. This is what Claude sees and operates. Open that URL. You should see a basic Linux desktop. That’s your sandbox.

If setup.sh fails on Mac with a permissions error, run chmod +x setup.sh first, then try again.

Step 2 — Write Your First Task Instruction (This Is Where Most People Go Wrong)

Claude computer use lives or dies on how precisely you describe the task. Think of it like giving instructions to a new assistant who has never seen your computer before and has no context about your business. Every assumption you leave unstated becomes a mistake.

Here’s the difference between a bad instruction and a working one:

Bad:

Go into my email and pull out the property inquiry information and put it in a spreadsheet.

Working:

Open the Firefox browser. Go to gmail.com. I am already logged in. 
Search for emails with the subject line containing "Viewing Request" received in the last 7 days.
For each email you find:
1. Open the email
2. Copy the sender's full name
3. Copy the property reference number (it appears after the text "Property Ref:")
4. Copy the date and time they requested for the viewing
5. Open the LibreOffice Calc spreadsheet already open in the taskbar
6. Paste those three values into the next empty row: Column A = sender name, Column B = property ref, Column C = requested date/time
7. Return to Gmail and move to the next email
Repeat until all matching emails in the last 7 days are logged. Save the spreadsheet when done.

Specific. Sequential. No assumptions. That’s the format that works.

Step 3 — Run the Task and Monitor the First Session

In the demo interface, paste your instruction into the prompt field and hit Send. Watch what Claude does in the virtual desktop window. Do not walk away on your first run. Claude will narrate each action in the chat panel — “I’m clicking the Firefox icon,” “I’ve typed the search query” — and you can catch errors in real time.

Common things to watch for on first runs:

• Claude clicks the wrong element because two buttons look visually similar — you’ll see it pause or backtrack
• A login screen or CAPTCHA appears that Claude can’t pass — you’ll need to handle auth manually before handing off the task
• The spreadsheet column layout doesn’t match what you described — Claude will either stop and ask, or guess wrong

On my first run with a Gmail-to-spreadsheet task, Claude successfully processed 9 of 11 emails correctly. It got confused on two where the property reference was formatted differently. That’s a 82% success rate — useful, but not hands-off yet. After I added a line to my instruction explaining the two possible formats, the next run hit 100% on 14 emails.

Step 4 — Identify and Build Your Repetitive Admin Workflows

Once you’ve run one successful task, the question is: what else can you hand off? Here are the task types that work reliably with Claude computer use, and the ones that don’t.

Task Type	Works Well?	Notes from Testing
Copy-paste data between apps	✅ Yes	Best use case. Reliable when layouts are consistent.
Email triage and logging	✅ Yes	Works well with structured emails. Struggles with messy threads.
File renaming and organization	✅ Yes	Give it exact naming conventions. Works fast.
Form filling (web forms)	⚠️ Partial	Breaks on dynamic fields, dropdowns with JavaScript.
Social media posting	⚠️ Partial	Platform UI changes break it frequently. Not stable.
Tasks requiring judgment calls	❌ No	Claude will stop and ask, or make wrong assumptions.
Multi-app report assembly	✅ Yes	Pulling data from 2-3 sources into one doc works well.

Step 5 — Build a Prompt Library for Your Recurring Tasks

Don’t rewrite instructions every time. Once a task runs cleanly, save the instruction as a text file. I keep mine in a folder called /claude-tasks/ with filenames like weekly-inquiry-log.txt and monthly-listing-export.txt. When I need to run them, I open the file, paste it into the prompt field, and go make coffee.

Here’s a real prompt template I use for logging new property inquiries from email into my tracking sheet:

TASK: Weekly Inquiry Logging — Madeira Property Inquiries
DATE RANGE: Last 7 days from today

Open Firefox. Navigate to gmail.com. I am logged in as [your email].

Search for: subject:("property inquiry" OR "viewing request" OR "interested in") 
Filter: Last 7 days, Unread only

For each result:
1. Open the email
2. Extract: Sender full name | Sender email | Property name/address mentioned | Date received
3. If no property name is mentioned, write "Not specified"
4. Open LibreOffice Calc file at: Desktop > Inquiries_2026_Q1.ods
5. Find the first empty row in Sheet 1
6. Enter data: A=Date received, B=Sender name, C=Sender email, D=Property mentioned
7. Mark the email as Read in Gmail
8. Return to inbox, open next result

When all emails are processed:
- Save the spreadsheet (Ctrl+S)
- Take a screenshot of the completed rows and save it to Desktop as "inquiry_log_[today's date].png"

Stop and report how many rows were added.

That last line — “stop and report how many rows were added” — is important. It gives you a built-in audit. You can glance at the number and cross-check against your inbox without reading every row.

My Real-World Experience Using Claude Computer Use in My Madeira Real Estate Business

I run a one-person real estate consultancy in Madeira. No assistants, no team. Just me, a laptop, and a client base that expects fast responses. The admin work that comes with managing 15–25 active listings at any given time is relentless. Every inquiry needs to be logged. Every property detail needs to be formatted correctly for different portals. Every week I pull together a mini market summary for my newsletter. None of this work is hard — it’s just time-consuming in a way that’s hard to justify when I could be doing calls or viewings.

I started testing Claude computer use in February 2026. My first serious workflow was the inquiry logging task I showed above. Before Claude computer use, I spent about 35 minutes every Monday morning going through the previous week’s emails and manually updating my tracking spreadsheet. I also frequently missed one or two inquiries when my inbox was busy — not good when you’re a solo operator and every lead matters.

After three weeks of iterating on the prompt, the task now runs in 6–8 minutes with me doing nothing except opening the demo environment and pasting the instruction. That’s roughly 27 minutes saved per week, or about 2 hours per month on just that one task. Doesn’t sound massive until you stack it with the two other workflows I’ve built: a monthly file-organization task that sorts signed contracts into the right client folders (used to take 20 minutes, now takes 4), and a quarterly data-pull that assembles price-per-square-meter figures from three different browser tabs into a single table for my market report (used to take 55 minutes, now takes around 12).

Combined, I’m recovering around 3.5 hours per month from tasks that I was doing manually with zero cognitive value. That’s not revolutionary on paper, but in practice it means I stopped losing Friday afternoons to admin catch-up. I spend those hours on calls with prospective buyers instead.

One specific win I didn’t expect: accuracy. My manual logging had a small but real error rate — wrong date, wrong property name, sender email mistyped. Claude computer use, once the prompt is dialed in, makes none of those errors. The data in my spreadsheet is cleaner now than it’s ever been, which also made my quarterly reporting faster because I spend less time fixing inconsistencies.

The API costs for my usage have been running at about $22–$28 per month. For 3.5 hours of recovered time, that math works clearly in my favor.

The Real Limitations I Hit During Testing

I want to be straight about where this falls apart, because the demos Anthropic shows look smoother than reality.

It is slow. Claude computer use does not operate at human speed. It takes a screenshot, processes what it sees, decides on an action, executes it, takes another screenshot, and repeats. A task that would take me 4 minutes of focused clicking might take Claude 9–12 minutes. For tasks where I can genuinely walk away and do something else, that’s fine. But it means you can’t use it for anything time-critical.

Dynamic web interfaces break it. I tried to use it to update listings on one of the major Portuguese property portals. The portal uses a lot of JavaScript-rendered dropdowns and auto-save fields. Claude got through about 40% of the form before losing track of its place in the interface. I spent more time troubleshooting than the task would have taken me manually. I dropped that use case after 4 days of attempts.

It needs supervision on new tasks. The first 2–3 runs of any new workflow require you to watch closely. Claude will confidently do the wrong thing sometimes — clicking a “Delete” button when it meant “Archive,” for example — and if you’re not watching, the mistake compounds. Build in review time for new automations.

Authentication walls stop it cold. Any task requiring two-factor authentication or a login that doesn’t persist will stall the session. You have to handle logins yourself before handing off the task, which limits full end-to-end automation.

Troubleshooting Tips When Things Break

Claude keeps clicking the wrong element: Add more visual context to your prompt. Instead of “click the Save button,” write “click the blue Save button in the top-right corner of the form.” Screenshots of the UI pasted into the prompt also help Claude orient itself.

Docker container won’t start on Windows: Enable WSL2 in your Windows settings first. Run wsl --install in PowerShell as administrator, restart, then try Docker again. Anthropic’s demo is built for Linux environments and WSL2 is the stable path on Windows.

Task stops halfway with “I’m not sure how to proceed”: Your instructions have a gap. Claude hit a state it wasn’t prepared for. Go back to your prompt, add a conditional: “If you see a popup asking X, click Dismiss and continue.”

API calls are costing more than expected: Long tasks with lots of screenshots burn tokens fast. Break complex tasks into shorter, discrete instructions rather than one long run. Five 10-step instructions will usually cost less than one 50-step marathon session.

Claude is operating too slowly for your patience: This isn’t really fixable — it’s an architectural limit of the screenshot-action loop. Set the task running, switch windows, and come back. Watching it operate in real time makes it feel slower than it is.

Rating and Practical Summary

My rating: 3.8/5 — Genuinely useful for high-frequency, low-complexity data tasks, but the slow execution speed and fragility on modern web interfaces mean it’s not a complete admin replacement for a solo operator yet.

Here’s where Claude computer use earns its place in my workflow and where it doesn’t:

• Use it for: recurring copy-paste tasks between stable apps, email-to-spreadsheet logging, file organization, multi-source data assembly for reports
• Skip it for: dynamic web portals, anything requiring active logins, tasks that need judgment or contextual decisions, anything time-sensitive
• Budget for: $20–$40/month in API costs for moderate solo business usage
• Invest upfront in: writing precise, step-by-step prompts — the first hour you spend doing this saves you weeks of frustration

If you’re running a solo business and you have even three tasks that you do every week on autopilot — same steps, same apps, same output — Claude computer use is worth the setup time. The payoff isn’t dramatic in week one. But by month two, when those tasks are just running while you focus on actual work, you’ll stop noticing the time you used to lose to them. That’s exactly what happened for me in Madeira.

Start with one task. Get it working cleanly. Then build the next one. Don’t try to automate everything at once.

If you want to see the other Claude workflows I use in my real estate business — including how I combine Claude Artifacts with computer use for client report generation — browse the Claude AI section of the site. Everything there comes from actual testing, not product demos.