How to Use OpenClaw
I remember the first time I got OpenClaw running and just stared at the Telegram chat, cursor blinking, thinking “okay, now what do I actually say to this thing?” It felt weirdly like the first time I opened a blank terminal after installing Linux. Infinite possibility and zero guidance.
So this is the post I wish I had read then.
Your first message
Start simple. Do not try to build a complex automation on day one. Send something like:
“Go to news.ycombinator.com and tell me the top 5 stories right now.”
This tests the full pipeline: the agent receives your message, opens a browser, navigates to a URL, reads the page, extracts information, and sends it back. If that works, everything is wired up correctly. If it does not, you know to check your setup before going further.
Writing good prompts
OpenClaw is not ChatGPT. You are not asking it to generate text. You are asking it to do things on a computer. That shift changes how you should write prompts.
Be specific about the output you want. “Research CRM tools” is vague. “Visit the pricing pages of HubSpot, Salesforce, Pipedrive, and Zoho, and create a comparison table with plan names, monthly prices, and contact limits” gives the agent a clear task with a defined end state.
But you also do not need to micromanage every step. The agent figures out navigation, handles popups, deals with cookie banners. Tell it what you want, not how to click through each page. I have found that the sweet spot is describing the desired output in detail while leaving the method up to the agent, because it often finds faster paths than you would have prescribed anyway.
Real tasks people run
These are pulled from Reddit threads and the OpenClaw community Discord, lightly edited:
Research and reports. “Research the top 5 project management tools for freelancers. Compare pricing, key features, and user reviews. Format as a report I can share with my team.”
Web monitoring. “Check this apartment listing page every 6 hours. If any new 2-bedroom listings under $2,000 appear, send me the links.”
Data processing. Send a CSV file and say: “Clean this up. Remove duplicate rows, standardize the date format to YYYY-MM-DD, and flag any rows where the email column is empty.”
Code tasks. “Write a Python script that reads a folder of invoice PDFs, extracts the total amount and vendor name from each, and outputs a summary CSV.”
Scheduling and reminders. “Every Monday at 8am, check the exchange rate for USD to EUR and send me the current rate.”
Building on what works
Once you have a few tasks running, the interesting part starts. People tend to build increasingly sophisticated workflows by chaining simple tasks together. A freelancer on the OpenClaw subreddit described a setup where the agent monitors three client inboxes, drafts responses based on templates, and queues them for approval every morning, all triggered by a single scheduled task.
The trick is to start with one step, verify it works, then add the next. Trying to describe a 10-step automation in a single message usually leads to confusion. Break it up, test each piece, and combine them once you trust the individual parts.
When it gets stuck
OpenClaw is not perfect. It sometimes misreads a page, takes a wrong turn during research, or fails to extract data from a particularly messy website. When that happens, just tell it. “That’s not right, the pricing is on /pricing not /plans” or “Skip that site, try this one instead.” The agent adjusts.
For tasks that fail consistently, check whether the target website is heavily JavaScript-dependent or uses aggressive bot detection. Most sites work fine, but some fight back. The OpenClaw GitHub issues are useful for finding workarounds when you hit a wall.
Going further
The architecture guide explains what is happening under the hood. The Telegram guide covers the messaging setup if you have not configured that yet. And if you are still in the planning stage and haven’t installed anything, the install guide walks through every method.
Or if you skipped the self-hosting route entirely and went with OpenClaw Cloud, all of the above works exactly the same. You just did not have to set up a server first.