Automate Freelance Work With Claude AI—In 10 Minutes

Last October, I spent 90 minutes manually pulling comparable property prices from three different Portuguese real estate portals, copying them into a spreadsheet, and formatting a market analysis report for a client in Funchal. Ninety minutes for work that should take fifteen. That’s the kind of inefficiency that quietly kills a solo operation. Then I started testing Claude’s computer use feature — and that specific task now runs while I drink my morning coffee.

Claude’s computer use capability (part of the Claude API, powered by Anthropic’s tool-use architecture) lets the AI actually control a computer: click buttons, fill forms, read screens, copy data, and move between applications. It’s not a chatbot giving you instructions. It does the work. For freelancers running lean operations, that difference is enormous.

This tutorial walks you through exactly how I set it up for my real estate consulting business in Madeira — and how you can adapt the same approach for your freelance workflow.

What You’ll Build by the End of This Tutorial

By the time you finish this guide, you’ll have a working Claude computer use setup that can:

• Open web browsers and extract structured data from multiple sites
• Fill in forms and populate documents automatically
• Move files between folders based on naming rules
• Compose and draft emails using information pulled from your desktop
• Run repeatable research workflows without you touching a keyboard

The tutorial is built around real estate workflows, but every step applies directly to any freelance operation — copywriters, consultants, virtual assistants, designers.

Prerequisites Before You Start

You need a few things in place before step one:

• Claude API access — Computer use runs through the API, not the standard Claude.ai chat interface. Sign up at anthropic.com and grab your API key. As of 2026, the Claude 3.5 Sonnet model supports computer use and costs approximately $3 per million input tokens and $15 per million output tokens.
• Python 3.10+ installed — The reference implementation Anthropic provides runs on Python. If you’re not a developer, don’t panic — I’ll give you the exact commands to copy-paste.
• A Linux or macOS machine, or a Linux virtual machine on Windows — Computer use works best in a sandboxed Linux environment. Anthropic’s own quickstart uses Docker. I run it in a lightweight Ubuntu VM on my MacBook.
• Docker Desktop installed — This is the easiest way to run the sandboxed environment. Free to download at docker.com.
• 20 minutes of uninterrupted time for the initial setup.

If you’ve never touched an API before, this will feel unfamiliar for about the first 30 minutes. Push through it. The payoff is real.

Step 1: Clone Anthropic’s Computer Use Demo and Set Up Docker

Anthropic provides an official quickstart repository. Open your terminal and run:

git clone https://github.com/anthropics/anthropic-quickstarts.git
cd anthropic-quickstarts/computer-use-demo

Now build and run the Docker container. Replace your_api_key_here with your actual Anthropic API key:

export ANTHROPIC_API_KEY=your_api_key_here
docker build -t computer-use-demo .
docker run 
    -e ANTHROPIC_API_KEY=$ANTHROPIC_API_KEY 
    -v $HOME/.anthropic:/home/user/.anthropic 
    -p 5900:5900 
    -p 8501:8501 
    -p 6080:6080 
    -it computer-use-demo

Once it’s running, open your browser and go to http://localhost:8501. You’ll see a Streamlit interface on the left and a live view of the virtual desktop on the right. That desktop is what Claude controls.

The first time I saw Claude move a mouse cursor across a screen by itself, I genuinely stopped what I was doing. It’s jarring in the best way.

Step 2: Write Your First Freelance Task Prompt

The interface has a chat box where you give Claude its instructions. The key to getting good results is writing prompts that are specific, sequential, and scoped. Vague instructions produce wandering behavior.

Here’s a prompt template I use for property research. Adapt the site names and fields for your own industry:

Open the Firefox browser. Go to idealista.pt. 
Search for apartments for sale in Funchal, Madeira. 
Filter results to show only listings priced between €200,000 and €400,000 
with at least 2 bedrooms.

For the first 10 results, collect the following information:
- Listing title
- Price (€)
- Size (m²)
- Number of bedrooms
- Neighborhood
- URL of the listing

Open LibreOffice Calc. Create a new spreadsheet with those columns.
Paste the collected data row by row.
Save the file to the Desktop as "funchal_comps_[today's date].ods".

When finished, tell me how many listings you found and flag any 
listings where the price per m² exceeds €3,500.

Hit Enter. Then watch. Claude will open the browser, type the search, click through filters, read listing cards, open a spreadsheet, and populate it. The whole run takes 8 to 12 minutes depending on how fast the target website loads. My manual version of this task used to take 40 to 50 minutes.

Step 3: Build a Repeatable Prompt Library for Your Workflow

One-off prompts are useful. A library of tested, reusable prompts is where the real time savings stack up. I keep mine in a plain text file called claude_tasks.txt on my desktop. Here are three templates I use weekly:

Template A: Lead Follow-Up Email Draft

Open the file "leads_this_week.txt" on the Desktop.
Read the first lead entry. It will contain: name, property interest, 
last contact date, and any notes.

Open the Firefox browser and go to gmail.com. Log in if needed.
Compose a new email to the email address listed in that lead entry.

Write a follow-up email in Portuguese (European) with the following structure:
- Friendly opening using their first name
- Reference the property type they enquired about
- Mention 1-2 new listings that match their stated criteria 
  (use the listings from "funchal_comps_[today's date].ods" on the Desktop)
- Suggest a call this week with 2 specific time slots: 
  Tuesday 10am or Thursday 3pm Lisbon time
- Close professionally as Robson Penassi, Real Estate Consultant, Madeira

Save the email as a draft. Do NOT send it.
Tell me the subject line you used so I can review it.

Template B: Property Description for Listings

Open the file "property_details_[address].txt" on the Desktop.
It contains: property type, size, rooms, floor, view, renovation status, 
asking price, and 3 selling points noted by the owner.

Write a property listing description in English with the following specs:
- 180 to 220 words
- Opening sentence that leads with the strongest feature
- Paragraph 2: practical details (size, rooms, layout)
- Paragraph 3: location context specific to Madeira (mention the 
  island's climate, access to Funchal, or proximity to the coast 
  as relevant)
- Final line: call to action for private viewings
- Tone: warm but professional, no superlatives like "stunning" or "amazing"

Open LibreOffice Writer. Paste the description.
Save as "listing_description_[address].odt" on the Desktop.

Template C: Weekly Social Media Content Batch

Open the file "weekly_content_brief.txt" on the Desktop.
It contains: 3 topic ideas for LinkedIn posts about Madeira real estate.

For each topic, write one LinkedIn post with these specs:
- 100 to 150 words
- First line must work as a standalone hook (no "Hey everyone!" openings)
- Include one specific fact or number about Madeira property market
- End with one direct question to encourage comments
- No hashtag spam — maximum 3 hashtags per post

Open LibreOffice Writer. Create a document titled "LinkedIn Posts - Week [X]".
Paste all three posts with a clear separator between them.
Save to Desktop.

Step 4: Set Up File Naming Conventions Claude Can Follow

Computer use works best when your file system is consistent. Claude reads filenames and folder names to orient itself. If your Desktop looks like most people’s — random downloads, inconsistent naming, folders within folders — the agent will get confused or waste tokens trying to figure out where things are.

I use a flat folder structure on the VM desktop:

• /Desktop/inputs/ — raw data files Claude reads from
• /Desktop/outputs/ — files Claude creates or saves to
• /Desktop/tasks/ — my prompt library text files

Every input file follows the pattern: [task_type]_[identifier]_[date].txt. For example: lead_santos_joao_20260318.txt. When I give Claude a prompt that says “open the most recent lead file,” it can sort by date and find the right one without guessing.

Spend 20 minutes organizing your VM’s file structure before you run anything complex. It will save you from failed runs caused by Claude opening the wrong file.

Step 5: Monitor, Verify, and Correct — Don’t Walk Away

This step is where most tutorials skip over the important part. Computer use is not set-and-forget. Not yet. Claude makes mistakes: it clicks the wrong button, misreads a form field, or gets stuck on a CAPTCHA. During any task longer than 5 minutes, you should be watching the screen view in the Streamlit interface.

When something goes wrong mid-task, type a correction in the chat box. Claude responds to mid-task instructions. I’ve done things like:

Stop. You clicked the wrong filter. The price range should be 
€200,000 to €400,000, not €200,000 to €300,000. 
Go back to the search results page and re-apply the correct filter.

It corrects course and continues. The key is catching the mistake early before it compounds across 10 rows of a spreadsheet.

My Real-World Experience Using Claude Computer Use in Madeira

I run a one-person real estate consultancy. No assistants, no junior brokers. Every hour I spend on admin is an hour I’m not spending with clients or generating new leads. In January 2026, I committed to testing Claude computer use for a full 30 days on real tasks, tracking time carefully.

The most impactful change was in market analysis reporting. Every time I take on a new buyer client, I build a comparable properties report — pulling prices, sizes, locations, and days on market from Idealista, Imovirtual, and occasionally direct agency sites. Before computer use, this took me between 75 and 90 minutes per report, depending on how active the market was that week. I was producing roughly 6 of these reports a month, so we’re talking 7.5 to 9 hours of work.

After setting up the data extraction prompt in Step 2 of this tutorial and refining it over about 8 runs, I got that report-building time down to under 15 minutes per report. The AI does the browsing and spreadsheet population. I spend my time reviewing the output, flagging any obvious anomalies, and writing the 2-paragraph narrative summary at the top. Six reports a month now costs me roughly 90 minutes total instead of 9 hours. That’s 7.5 hours back per month — in a one-person business, that’s meaningful.

The property description workflow also changed significantly. In February 2026 I had 9 new listings to write copy for in a single two-week stretch — unusual for Madeira’s market but it happened. Using Template B from this tutorial, I batched all 9 descriptions in one afternoon. Each run took Claude about 6 minutes. My review and editing added maybe 5 minutes per description. Total time: roughly 100 minutes for all 9. My old process, writing them manually with ChatGPT assistance, would have taken me close to 4 hours.

One specific thing I noticed: Claude does better with European Portuguese site interfaces than I expected. Idealista.pt’s filters use Portuguese labels and the AI reads them correctly without any special prompting. That surprised me — I’d assumed I’d need to write prompts that mapped Portuguese UI labels to English concepts. Mostly I didn’t need to.

I’ve also used computer use for the less glamorous admin: renaming batches of property photo files from camera roll format (IMG_4521.jpg) to the structured naming my website requires (funchal_apartamento_estrela_01.jpg). A task I used to do manually for 20 minutes after every photo shoot now runs in 3 minutes with a simple prompt. Small, but it adds up fast when you’re doing it every week.

Where Claude Computer Use Falls Short

I want to be honest here because some of the coverage of this feature is breathlessly positive and it does not reflect what actually happens in day-to-day use.

CAPTCHAs stop it cold. Several Portuguese real estate portals trigger CAPTCHAs after 3 or 4 page loads from the same IP. When that happens, the task fails completely and Claude doesn’t always tell you clearly why — it sometimes just stops moving. I’ve had to manually solve the CAPTCHA in the VM browser and then re-prompt Claude to continue. This is manageable but annoying, and it means you can’t fully walk away from longer tasks.

Speed is inconsistent. Some runs that should take 8 minutes take 20. Token consumption is also hard to predict. I’ve had data extraction tasks cost me $0.40 in API calls and nearly identical tasks cost $1.20 because Claude took more steps to complete the same work. Budget in some buffer if you’re doing this at volume.

Dynamic JavaScript-heavy sites are harder. Sites that load data asynchronously after scrolling or after user interaction sometimes confuse the agent. It clicks in what appears to be the right place but the data hasn’t fully rendered yet. I’ve learned to add explicit wait instructions in my prompts: “After clicking the search button, wait 5 seconds before reading any results.”

It is not autonomous. I want to be clear: this is not a tool you set up on Monday and forget about. It requires active supervision on any task over 10 minutes. That’s fine — the supervision time is still much less than doing the work manually — but go in with realistic expectations.

How Claude Computer Use Compares to Other Automation Options

Tool	Setup Difficulty	Handles Unstructured Sites?	Needs Coding?	Monthly Cost (Solo)
Claude Computer Use	Medium	Yes	Minimal (copy-paste setup)	Pay-per-use (~$10–40 depending on volume)
Make.com + Browserless	High	Partial	Some required	$16–29 (Make) + Browserless costs
Zapier + Chrome Extension	Low	No	No	$20–49
UiPath (RPA)	Very High	Yes	Yes	$420+ (enterprise pricing)
Playwright/Puppeteer (DIY)	Very High	Yes	Yes (developer-level)	Free but requires dev skills

For a solo freelancer who isn’t a developer, Claude computer use sits in a sweet spot: more flexible than Zapier, vastly cheaper than enterprise RPA tools, and accessible without writing actual code.

Troubleshooting the Most Common Problems

Claude stops moving mid-task

Usually means it hit a CAPTCHA, a login wall, or a page that loaded differently than expected. Check the screen view. If you