Draft messages, fill web forms and prototype sites by chat
Get your ClawdBot/OpenClaw set up with a single prompt. Task: Get your ClawdBot/OpenClaw set up with a single prompt 👋 Hey there! I’m trying something new with this post. From here on out, I plan to continue to post regularly on my Substack, but I will be making my content consumable by AI as well. That means that you can copy-paste a link to this post into your LLM app of choice (i.e. Cursor, Claude Code, or just Gemini and ChatGPT). Those agents should be able to consume this content to help you make sense of it. On to the post! tl;dr A huge part of getting OpenClaw set up is the security, prompts, and habits to using the repo. I’ve shared a setup prompt below that handles installation, security hardening, and a crash-recovery watchdog into a single copy-pasteable prompt below. This post explains what the prompt does, how to shape your agent’s personality through workspace files (SOUL.md, AGENTS.md, USER.md), and what your first week should look like. In Part 1, I walked through the hardware and setup: buy a Mac Mini, get an eSIM, install the software, scan a QR code, text your AI. If you followed along, you now have an agent that responds when you message it. Cool! But also... now what? 😅 Here’s what I found: getting it running is maybe 20% of the work. The other 80% is making it actually useful, and making sure it doesn’t accidentally publish your API keys to the internet or respond to strangers in your group chats. Part 1 was about the hardware. This post is about the software you can’t install with npm: the security, the prompts, and the habits that turn “I have a chatbot on WhatsApp” into “I have a chief of staff that knows how I think.” If you just want the setup prompt, I’ve put my battle-tested config into a single file you can paste into Claude Code (or any coding agent). It handles security hardening, workspace prompts, and a watchdog for crash recovery. Paste it below into your coding agent (i.e. a fresh claude code session on the machine you will install OpenClaw on): “Read https://raw.githubusercontent.com/amanaiproduct/openclaw-setup/main/PROMPT.md and follow every step. Ask me for my Anthropic API key when you need it.” AI Product Playbook is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. ⚠️ I highly recommend not installing Open Claw on your daily machine or work machine. Read more in my previous post about why I don’t recommend running this on a machine you use regularly. If you’re looking for a private agent sandbox to run Open Claw, check out my friends here at Runtime (not sponsored, I just like their product because it strikes the right balance of seeing the “guts” without being overly complicated) ⚠️ The repo is here: github.com/amanaiproduct/openclaw-setup. The rest of this post explains why each piece matters and how to get the most out of it. Lock It Down First I’m putting this first on purpose. Before you do anything fun with your agent, you need to spend 15 minutes making sure it’s not a liability. Remember, this thing has access to your filesystem. It can run commands. It can browse the web. That’s the whole point. But it also means a misconfiguration can expose your machine to anyone on your network (or worse, the internet). When I first set mine up, the gateway was running with default settings. No auth token. Bound to all interfaces. Basically an open door. I didn’t realize this until I ran the built-in security audit and it flagged everything red. The setup prompt handles all of this for you. Paste it into Claude Code and it configures everything. But here’s what it’s doing and why, so you understand what’s being locked down: Gateway binding. By default, the gateway listens on all network interfaces. That means any device on your WiFi can talk to it. The prompt sets it to loopback only, so only your machine can connect. If you’re on a shared network (coffee shop, coworking space), this is the difference between “my agent” and “everyone’s agent.” Authentication. The prompt enables token auth on the gateway. Without it, any connection is trusted. Never run without auth. Ever. File permissions. Your config directory has API keys and tokens in it. The prompt locks it down so only your user account can read those files. Basic hygiene.; tools: OpenClaw. Human role: The original post does not say how a person stayed involved.; limitations: Prerequisites include a dedicated isolated machine (e.g., Mac Mini) and an Anthropic API key. The source is a personal blog post, so reliability depends on the author's current setup. Security constraints are critical: never use Tailscale Funnel, and ensure gateway binding is loopback-only. Human approval is required for API key entry and final workspace configuration..
Tools
Published on Agent Work Atlas:
Overview
Get your ClawdBot/OpenClaw set up with a single prompt
Workflow
Get your ClawdBot/OpenClaw set up with a single prompt 👋 Hey there! I’m trying something new with this post. From here on out, I plan to continue to post regularly on my Substack, but I will be making my content consumable by AI as well. That means that you can copy-paste a link to this post into your LLM app of choice (i.e. Cursor, Claude Code, or just Gemini and ChatGPT). Those agents should be able to consume this content to help you make sense of it. On to the post! tl;dr A huge part of getting OpenClaw set up is the security, prompts, and habits to using the repo. I’ve shared a setup prompt below that handles installation, security hardening, and a crash-recovery watchdog into a single copy-pasteable prompt below. This post explains what the prompt does, how to shape your agent’s personality through workspace files (SOUL.md, AGENTS.md, USER.md), and what your first week should look like. In Part 1, I walked through the hardware and setup: buy a Mac Mini, get an eSIM, install the software, scan a QR code, text your AI. If you followed along, you now have an agent that responds when you message it. Cool! But also... now what? 😅 Here’s what I found: getting it running is maybe 20% of the work. The other 80% is making it actually useful, and making sure it doesn’t accidentally publish your API keys to the internet or respond to strangers in your group chats. Part 1 was about the hardware. This post is about the software you can’t install with npm: the security, the prompts, and the habits that turn “I have a chatbot on WhatsApp” into “I have a chief of staff that knows how I think.” If you just want the setup prompt, I’ve put my battle-tested config into a single file you can paste into Claude Code (or any coding agent). It handles security hardening, workspace prompts, and a watchdog for crash recovery. Paste it below into your coding agent (i.e. a fresh claude code session on the machine you will install OpenClaw on): “Read https://raw.githubusercontent.com/amanaiproduct/openclaw-setup/main/PROMPT.md and follow every step. Ask me for my Anthropic API key when you need it.” AI Product Playbook is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber. ⚠️ I highly recommend not installing Open Claw on your daily machine or work machine. Read more in my previous post about why I don’t recommend running this on a machine you use regularly. If you’re looking for a private agent sandbox to run Open Claw, check out my friends here at Runtime (not sponsored, I just like their product because it strikes the right balance of seeing the “guts” without being overly complicated) ⚠️ The repo is here: github.com/amanaiproduct/openclaw-setup. The rest of this post explains why each piece matters and how to get the most out of it. Lock It Down First I’m putting this first on purpose. Before you do anything fun with your agent, you need to spend 15 minutes making sure it’s not a liability. Remember, this thing has access to your filesystem. It can run commands. It can browse the web. That’s the whole point. But it also means a misconfiguration can expose your machine to anyone on your network (or worse, the internet). When I first set mine up, the gateway was running with default settings. No auth token. Bound to all interfaces. Basically an open door. I didn’t realize this until I ran the built-in security audit and it flagged everything red. The setup prompt handles all of this for you. Paste it into Claude Code and it configures everything. But here’s what it’s doing and why, so you understand what’s being locked down: Gateway binding. By default, the gateway listens on all network interfaces. That means any device on your WiFi can talk to it. The prompt sets it to loopback only, so only your machine can connect. If you’re on a shared network (coffee shop, coworking space), this is the difference between “my agent” and “everyone’s agent.” Authentication. The prompt enables token auth on the gateway. Without it, any connection is trusted. Never run without auth. Ever. File permissions. Your config directory has API keys and tokens in it. The prompt locks it down so only your user account can read those files. Basic hygiene.
Reported result
The author reports that initial setup is only 20% of the effort, with the remaining 80% focused on making the agent useful and secure. The setup prompt successfully configures gateway security and watchdogs, preventing common issues like open network exposure. The author notes that Claude Sonnet handles 90% of tasks at a lower cost, while Opus is reserved for complex reasoning.
Human role
The original post does not say how a person stayed involved.
Limitations
Prerequisites include a dedicated isolated machine (e.g., Mac Mini) and an Anthropic API key. The source is a personal blog post, so reliability depends on the author's current setup. Security constraints are critical: never use Tailscale Funnel, and ensure gateway binding is loopback-only. Human approval is required for API key entry and final workspace configuration.
Try it
I want to set up OpenClaw on my [replaceable: dedicated machine name] using a coding agent. Please execute the following steps: 1. Open a fresh terminal session on the target machine. 2. Run the command: `claude "Read https://raw.githubusercontent.com/amanaiproduct/openclaw-setup/main/PROMPT.md and follow every step. Ask me for my Anthropic API key when you need it."` 3. When the agent asks for the API key, provide it securely. 4. After installation, verify that the gateway is bound to loopback only and token auth is enabled. 5. Review the generated SOUL.md, AGENTS.md, and USER.md files. Do not send any messages or publish any content until I have manually reviewed and approved these files. 6. Report the status of the watchdog script and any security audit results. If any step fails or requires permissions I haven't granted, stop and ask for clarification. Do not proceed with sending messages or modifying external systems without my explicit confirmation.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.