luffyguy

Posted on Apr 13 • Originally published at Medium

how people are actually making money using openclaw.

#ai #technology #automation

Most OpenClaw demos on Twitter are cute. Someone orders food. Someone scrolls Twitter. Flashy, useless.

The real opportunity is boring: automating workflows businesses will pay thousands for. Here’s the complete playbook.

1. The Upwork hack

Go to Upwork. Search “robotic process automation” or “computer use automation.” Filter by $500+ budget.

You’ll find posts like: “$1,000 — need desktop automation for my PDF software business.”

Here’s the move:

Spawn 4–8 OpenClaw sub-agents in parallel
Each one reads a different job post
Each one builds a rough working demo based on the job description
You pick the best demo, attach it to your proposal

Everyone else sends a generic cover letter. You send a working prototype. You win.

This is the fastest path from zero to first paying client. Upwork is a buyer list hiding in plain sight.

2. Main agent = manager. Sub-agents = workers.

Your main OpenClaw shouldn’t do the work. It should delegate.

Think of the main agent as holding a hot coffee. If you ask it to move a desk, it can’t — it’s busy. Same with OpenClaw. If it’s scrolling TikTok for 20 minutes, you can’t talk to it.

Fix: sub-agents. Each one is a specialist with its own skill and its own computer. Main agent orchestrates, sub-agents execute.

This is Anthropic’s Skills pattern applied to computer use.

3. Computer-use agents are the universal API

90% of business software has no clean API. Insurance portals. Zoho. SAP. Old manufacturing tools. Government sites.

APIs can’t touch them. Humans click through them. Computer-use agents click through them too — which means they unlock automation for workflows that were previously impossible.

This is the a16z thesis. Dario Amodei went further on the Dwarkesh podcast: he called computer-use agents the constraint to reaching AGI.

If that’s how Anthropic’s CEO frames it, pay attention.

4. Find the workflow, not the tool

Don’t sell “OpenClaw.” Nobody cares.

Sell “I’ll automate this specific painful workflow that’s costing you 10 hours a week.”

Use the value vs. effort 2x2. Map every automation candidate. Start with high-value, low-effort. That’s your wedge.

Real example from a client: receive email → parse product links → scrape each product page → download reports → parse data → upload to Zoho CRM. One workflow. Clear ROI. Easy to scope.

5. Don’t dump everything on OpenClaw. Build pipelines.

This is where most people fail. They ask OpenClaw to do the whole thing in one prompt. It breaks.

Correct architecture:

Use Claude Code to build proper Python pipelines for the heavy work
OpenClaw’s job is to listen for triggers (email arrives, text received, cron fires)
When triggered, it calls the right sub-agent or script

You’re building specialized AI workers. OpenClaw is the dispatcher.

6. Skateboard before car

Don’t build the full system on day one. Build the smallest working skill. Test it. Break it. Fix it.

Want to automate TikTok trend research? Don’t build a full analytics platform. First, get an agent to open TikTok and scroll. That’s it. Then add screenshot analysis. Then add the database. One step at a time.

7. Always make the agent ask you questions first

Before you tell OpenClaw to build something, tell it: “What do you need from me to build this? Create a plan first.”

This forces scope clarity. It surfaces missing context. It catches bad assumptions before they become bad code.

Non-negotiable step in every build.

8. Pick a vertical

Don’t be everything to everyone.

Good verticals: real estate, manufacturing, distributorships, law firms, small agencies, e-commerce operators.

Avoid early: healthcare, finance. Red tape will kill you before you land a second client.

Unfair advantage doesn’t mean 20 years of experience. It means you understand the customer. Your mom was a real estate agent? That’s enough. Start there. You might have some good knowledge in that area. You know the client and their pain.

9. Two parallelization patterns

Know when to use which:

Split task: one job, 10 sub-agents each handle a chunk. Use when the task decomposes cleanly.
Parallel instances: same task, 10 agents run it independently. Use when you’re searching, scanning, or trying multiple approaches (like the Upwork hack).

Most people only use the first. The second is the unlock for research and prospecting work.

10. Agents are the new SaaS

Old model: build software → sell it → humans press the buttons to create value.

New model: build agents → sell access → agents press the buttons.

The client doesn’t log into a dashboard. They invite you into their workspace and a team of AI employees is already set up, already running, already producing output.

This mental shift is the whole game. If you’re still thinking “what SaaS can I build,” you’re behind.

11. Chat is the final interface

ChatGPT moment: chat box. OpenClaw moment: text message / Telegram.

The interface keeps collapsing into the simplest thing humans already use. You don’t need a UI. You need an agent that can text you back and use tools under the hood.

Build accordingly. Stop designing dashboards.

12. What actually breaks in production

Nobody talks about this. Here’s the list:

Captchas. Cloudflare, hCaptcha, Google reCAPTCHA v3 — agents fail silently. Solution: rotate residential proxies, use captcha-solving APIs (2Captcha, CapSolver), or pick targets without aggressive bot protection.
Session expiry. Cookies die. Agents keep clicking on a logged-out page. Solution: build session health checks and re-auth flows as part of the skill.
Rate limits. Hit a site too fast, get IP banned. Solution: randomize delays, respect robots.txt where possible, use proxy rotation.
Hallucinated clicks. Agent “thinks” it clicked a button but didn’t. Solution: verify state after every action. Don’t trust the model’s self-report — screenshot and re-check.
Cost per run. A long computer-use task can burn $3–8 in tokens. Multiply by 100 runs a day. Solution: cache aggressively, use cheaper models (Kimi, smaller Claude variants) for sub-agent work, reserve Opus for orchestration.
UI drift. Target site redesigns. Your agent breaks overnight. Solution: build assertions into the flow so it fails loudly, not silently.