How to Make OpenClaw Actually Work
A non-technical, step-by-step reference guide. 10 steps, each with checkable action items and inline detailed instructions with exact terminal commands.
Pro Tip: Once OpenClaw is running, send it this page.
Copy this page's URL, open a chat with your OpenClaw bot, and say: "Go to [URL] and implement all the setup steps in the guide that haven't been done yet." OpenClaw can read web pages and edit its own config files — many users complete 80% of this setup automatically.
What Is OpenClaw?
OpenClaw is a personal AI agent that lives on your computer. Unlike ChatGPT (which just answers questions), OpenClaw can do things for you — read and write files, browse the internet, send messages, manage your calendar, and run tasks on a schedule, all while remembering what it has learned about you over time.
Think of it as the difference between a consultant you call occasionally and a full-time employee who is always at their desk. The heartbeat feature wakes it up every 30 minutes to check if there is anything it should be doing — even when you are not asking.
| Feature | ChatGPT | Claude Code | OpenClaw |
|---|---|---|---|
| Lives on your computer | ✗ | ✓ | ✓ |
| Remembers things over time | Limited | Limited | ✓ |
| Works in Telegram / WhatsApp | ✗ | ✗ | ✓ |
| Runs tasks on a schedule | ✗ | ✗ | ✓ |
| Proactive (wakes up on its own) | ✗ | ✗ | ✓ |
| Extensible with skills/plugins | Limited | Limited | ✓ |
The 10-Step Setup Guide
Click "Show Action Items" to expand. Click "Detailed Instructions" under any item for step-by-step guidance with exact commands. Progress is saved in your browser.
Create a Troubleshooting Baseline
Before anything else
When things break, you need a smart helper to figure out why. This step creates your dedicated support system so you can solve 99% of setup issues on your own. The most critical part is getting the OpenClaw documentation installed inside OpenClaw itself — AND into your own Claude project — so both the agent and you can answer questions instantly. The secret weapon here is Context7: a free service that has already converted the entire OpenClaw documentation into a single, compressed, AI-ready Markdown file.
"Context7 is the secret weapon. One URL gives you the entire OpenClaw documentation as a single AI-ready file. Put it in your workspace folder AND your Claude project. The agent can now troubleshoot itself, and so can you."
Personalize Your Assistant
Make it sound like you
OpenClaw needs to know who you are and how it should behave. Without this, it gives generic, unhelpful answers. Think of it as writing an employee handbook for your new digital hire. There are four critical files. AGENTS.md is the most important — it defines how the agent behaves in every situation. SOUL.md defines its personality. IDENTITY.md gives it a job title. USER.md tells it about you.
"Goal: make responses specific, opinionated, and useful from day one. These four files are the foundation — AGENTS.md is the most important of all."
Fix and Optimize Memory
Make it actually remember
OpenClaw forgets things unless you explicitly tell it how to remember. Most memory problems happen because memory was never set up to save in the first place. Here is why: as your conversation gets longer, the context window (the agent's short-term memory) fills up. When it gets full, OpenClaw does a 'compaction' — it summarizes everything to free up space, and in doing so, loses details. The fix is to enable two settings that force the agent to write everything to a file BEFORE the compaction happens.
"The #1 memory mistake: giving important rules only in conversation. Put everything permanent in MEMORY.md or SOUL.md."
Choose the Right Model (The Brain)
Save money, stay reliable
OpenClaw needs an AI model to power it. Choosing the wrong method can cost you far more than necessary. The OAuth method lets you reuse a subscription you already pay for instead of paying expensive per-message API costs. Always set up backup models so work never stops.
"Use OAuth to avoid API costs. Always configure a fallback model. Switch models instantly in Telegram by typing 'models'."
Organize Telegram Groups and Topics
Prevent context confusion
If you talk to OpenClaw in one long, mixed conversation, it gets confused and loses track of what it's supposed to be doing. Separate Telegram groups or topics keep each task focused. Each group gets its own system prompt — its own job description.
"Separate Telegram groups with group-specific system prompts prevent context bleed and keep each task focused."
Understand the Three Browser Modes
Browse the web safely
OpenClaw can browse the internet for you, but it needs to know how to do it safely. There are three browser modes, each for a different purpose. Using the wrong one can accidentally expose your personal accounts to automation risks. Moritz demonstrated this live in the video: he has a skill that opens the managed browser, goes to his grocery delivery app (Rewe, like Instacart), and orders his weekly groceries automatically — completely hands-free.
"Managed profile for automation. Chrome relay only when you need real logged-in browser state. Brave API for search."
Install Skills
Give it new superpowers
Skills are like apps on your phone — they give OpenClaw new abilities. There are built-in skills, a marketplace of community-made skills at clawhub.com, and you can create your own. The 'summarize' skill alone is worth the entire setup process. Notable built-in skills: Summarize, 1Password, Apple Notes, Notion, OpenAI Whisper (voice transcription), Nano PDF, and Nano Banana Pro.
"If you repeat a task 2–3 times, skill it. The summarize skill alone justifies the entire setup."
Optimize the Heartbeat File
Keep it reliably alive
The 'heartbeat' is what wakes OpenClaw up every 30 minutes to check if it needs to do anything. Without proper setup, scheduled tasks silently fail and you never know. A well-configured heartbeat is the difference between a reliable assistant and a flaky one.
"Heartbeat for routine monitoring. Cron for exact timing. Combine both for a fully automated daily workflow."
Security Basics and Prompt Injection
Lock the doors
There are two distinct security risks with OpenClaw. Risk 1: Someone gets backend access to your machine (hacks in remotely). Risk 2: Prompt injection — a malicious email or webpage contains hidden instructions that trick the agent. Moritz's recommendation: run OpenClaw on your local Mac (not a cloud VPS) to minimize Risk 1. Use a strong model and add a security rule to AGENTS.md to minimize Risk 2.
"Two types of risk: backend access (use local Mac, not VPS) and prompt injection (use strong model + AGENTS.md security rule). Both are manageable."
Least Access Principle and Agent-Owned Accounts
Don't give it your keys
The safest approach is to give OpenClaw its own separate digital identity instead of access to your personal accounts. Think of it exactly like onboarding a new employee: you wouldn't hand a new hire the keys to your personal Gmail, calendar, and bank account on day one. You'd set them up with their own work accounts and give them access only to what they need. Same principle applies here.
"Clean separation, safer permissions, easier auditability. Give the agent its own identity — not yours."
Use Cases & Bonus Recommendations
Real-world examples, the most important mindset shift, and how to let OpenClaw set itself up.
Pro Tip: Let OpenClaw Implement These Steps For You
Once OpenClaw is running (even partially), send it this page and say: 'Read this guide and implement all the steps that haven't been done yet.' OpenClaw can read web pages, edit its own config files, and set up its own workspace. Many users complete 80% of this setup by simply pointing the agent at this guide.
How to do it:
The 'No AI Slop' Content System (7 Steps)
Moritz built a full 7-step automated content machine that creates authentic short-form videos for YouTube, Instagram, and TikTok — with minimal time investment. The key is that the content is still HIM on camera, not an AI avatar. OpenClaw handles everything except the 10-minute filming session.
How to do it:
Conversational CRM
Moritz built a conversational CRM using a Google Sheet as the database. He can ask it 'Who do I need to follow up with today?' and it checks his leads sheet, Gmail, and Calendar to give a complete answer. It can also write WhatsApp messages and Gmail drafts on his behalf. The big unlock: connecting Google Sheets + Gmail + Calendar in one workflow.
How to do it:
Treat It Like an Employee, Not a Tool
The overarching theme of the video is a mindset shift. Don't treat OpenClaw like Google Search — a tool you query and discard. Treat it like a new employee. Give it an onboarding manual (USER.md + SOUL.md), a dedicated workspace (agent-owned accounts), specific daily tasks (HEARTBEAT.md), and regular feedback. The more context you give it, the more valuable it becomes over time.
How to do it: