Enter OpenClaw. Formerly known as Clawdbot, then Molt, before finally landing on OpenClaw. Watching all those name changes happen in such a short time was… something.

At first, I didn’t understand why OpenClaw was so hyped up. It just felt like a wrapper for Claude Code. Didn’t think much of it at the start.

But then reality hit: I wanted to start doing research work or plan some coding stuff while I was outside or far from my PC. At first, I set up a remote terminal. It worked, but it wasn’t worth the effort and felt really awkward to use.

So I came back to the idea of OpenClaw and thought:

“I can just use this as an interface to communicate with Claude Code.”

And that’s how it started.

When I knew it was more than a wrapper

To my surprise, the installation was actually really easy. Some stuff seemed weird, and security for sure was (and probably still is) one of the main concerns. Even when I installed it for the first time, the warning was clear: you’re going to let this thing go wild in your files, so be careful.

Since I was running this on my local PC, I set up a Linux sandbox so it had basically its own environment and wouldn’t mess with my stuff directly. That plus some configuration to secure what it can do and where, and boom, ready to go.

Once I saw what tools it had, I started to realize how much stuff this thing had built in, including the clever approach to memory that, even if not optimal, was better than some other things I had used before.

The optimization loop

At the start, it was really dumb and I was using it only to pass prompts to Claude Code. But then I started, little by little, giving it some more basic stuff. And it worked. Sometimes it remembered actually good stuff and reminded me where I’d left off something.

But the more I relied on the internal architecture, the more problems it raised. So I started researching how to optimize this thing the best way possible.

There are a lot of changes I made to the initial configuration, but I think the best part of this tool is mainly the ability to configure itself just by prompting it. It’s not perfect, it has a lot of failures, and I had to dig into logs more than once in certain situations. But once I got an optimal way to communicate with it, delegate coding tasks always to Claude Code, and just have it work mainly as an orchestrator for other agents and maybe some research, it became the best tool to level up my workflow.

My current setup

After a lot of tweaking, I can rely on Claudio (the name I gave the bot) to delegate really well. There are still a lot of things it struggles with regarding memory, even with a lot of optimization, but I can share what’s working right now:

Set up hard rules on AGENTS.md and SOUL.md. I’m not sure still where to put what, but anything you don’t want your clanker to do, put it here. These files load every session and act as the bot’s operating manual.

Try different techniques of classifying your memory. This varies from person to person. Personally, I divide everything by projects (similar to Projects on Claude Desktop) and give it a system prompt with the context for that project. Each project has its own memory and lessons learned.

Set up a way for auto-improving. This is a big one. There are a lot of approaches to this, but for me I set up an instincts system that basically ranks certain intents with confidence. If something lands a lot on the same use case, it starts to rank that with more confidence, so it doesn’t have to guess what I want and knows exactly what to use and how. Also, having a TOOLS.md file to reference important tools works really well paired with the instincts pattern.

These are the ones I know work and have improved my clanker a lot. I can share more after I optimize it further.

So now I have a really good way to interact with Claude Code from afar, have information stored on my PC, and have context for each project. That’s when I started thinking:

“If I can do this from anywhere, I want to start experimenting with some projects to see how far I can push this. This architecture I built for OpenClaw should work for an autonomous project, right?”

But that’s a story for another post.

This is part 2 of a series about my journey with AI-assisted development. Next up: how I set up my first autonomous agent-driven project where I didn’t touch a line of code, only doing end-to-end tests and behaving like a user.