Building an OpenClaw Agent: Rudolf The Bookkeeper
The Recap and Materials of the AI Realist workshop
OpenClaw is not an easy tool to teach. It is unpredictable, unstable, and you never know what will break. If you run a hands-on session, you need to make sure participants have pre-installed everything. Installing OpenClaw is an adventure in itself: it requires the ability to use a terminal and to debug.
That is why I had no choice but to start a Telegram group and offer a setup troubleshooting session the day before the workshop.
This AI Realist workshop was probably the most time-consuming in terms of preparation.
So why bother teaching OpenClaw when you could just show Claude Cowork or Codex?
Because AI Realist does not teach applications, it teaches principles. Just as I never taught prompt engineering but rather the principles behind why LLMs behave the way they do, I do not want to teach agentic AI through sanitised commercial solutions. That approach teaches the application: click here, click there, write this prompt, enable this plug-in.
I want to show how desktop agents work, and using OpenClaw is perfect for this because it has no guardrails, no limitations, no blockers. Complete freedom to do what you want, and to learn.
With OpenClaw you will understand what an agent is: a massive prompt that calls tools.
You will see why you even need agents: because if you do not split your setup into agents and skills, the agent’s context window will be massive and the LLM will simply be too bamboozled to follow any critical instruction, like “do not delete all my files”.
You will see the real value of safety prompting: none. It does not matter how many exclamation marks, capital letters, or swear words you put before telling your agent not to send emails without your permission. It still will.
You will understand the real trade-off between safety and utility. You can put OpenClaw in a sandbox. You can cripple it to the same level that your company’s cybersecurity team has crippled your already not-particularly-capable MS 365 Copilot. You can make it even more useless than MS Copilot’s free version, and just as safe.
But the good news is you can also do the opposite. You can make it extremely useful, but you will need to navigate the art of making it safe(ish), accepting a certain amount of risk, and being ready to learn. OpenClaw in real production settings is a challenge that can give you a competitive advantage far beyond what paid solutions can, but there are no low-hanging fruit here. And if you learn to operate this one, you can switch to anything else. Your prompt “engineering” skills, without question, will be excellent.
Today’s workshop for paid subscribers of AI Realist had over 40 participants. It is structured so that even those who could not participate live can easily follow the recording and complete all the exercises. You can find all the materials for the workshop linked here.
The workshop materials below include:
The recording of the workshop
Slides for the workshop
Setup instructions
Exercise handout with prompts so that you can reproduce everything
GitHub repository where you can download the finished agent
A folder with supplementary materials
A Telegram group where you can ask your questions if you have problems with the setup or the exercises
Further in the article: Everything you need to do the workshop
Free subscribers get my public analysis. The stuff Forbes and Die Welt cite.
Paid subscribers get what’s too honest to say publicly: Vendor comparisons, AI announcements worth paying attention to, analytics of the latest events, myth debunking, and calling out hype and nonsense. And you can ask me anything about your AI decisions directly. And more workshops…
Founding members get all of that plus a 1:1 call with me. 45 minutes. 16 years in enterprise and research AI, from machine translation to production GenAI systems. Review your use case, validate your startup idea, or pressure-test your AI strategy. You decide how we spend this time.