Reinventing the Wheel (on Purpose)

There's a tool called OpenClaw that lets you build something pretty close to what my Discord bot does. I built mine from scratch anyway.

There are tools like Replit and Loveable that could have had my portfolio running in an afternoon. I did it in the terminal and deployed through Netlify CLI manually. I wanted to actually understand what I was building.

When you reach for a pre-built tool, a lot of decisions get made for you: which dependencies to pull in, how permissions are scoped, where credentials live, what happens when an API key expires. These decisions determine how a system fails, how it gets patched, and whether you can adapt it when requirements change. Reaching for an abstraction defers all of that until the worst possible time.

AI handles more of the execution now than it did a year ago. For most tasks I'm describing what I want and reviewing what comes back. But the judgment calls (which library, which permission model, what to do when the output is wrong) still require having done it before. The only way I've found to develop that is to actually go through it, including the failures. And especially because of them.

What I know about how these systems work didn't come from documentation. It came from breaking them. I don't think there's a shortcut to that.

There's also something else. Building your own version of an existing tool forces you to make every decision the original authors made, and understand why. The non-determinism of AI outputs, the retries, the edge cases, the moments where the agent does something unexpected and you have to decide whether to trust it or not. You only get calibrated on that by going through it.