How to use OpenClaw well: don’t — use it badly

A lot of people are asking my advice on how to use OpenClaw well to help them automate a lot of their administrative tasks.

Don't.

Use it badly.

There's a lot of advice floating around out there right now. Probably some of it is even good. I've seen some things I've found helpful.

But even if the perfect setup was out there and you managed to wade through all the dreck to find it, you'd be missing the point. The power of OpenClaw is that it's yours. And the only way it becomes yours is if you build it.

I've nuked my whole setup 4 times before I found the "right" way to do it. I'm constantly trying things and throwing them away. Don't treat OpenClaw like a magic system you put into place—treat it like an intern you're going to grow into a senior employee (but compress your timeline from 15 years to a few weeks). Don't expect perfection. Don't expect an instant click. And don't let someone else do the training for you.

Go in messy and break things, push too far, hit bottlenecks, find hallucinations. And then constantly, relentlessly, rebuild around that. Around your actual needs and your actual experience.

I really liked the framing of Nathaniel Whittemore (if you don't listen to The AI Daily Brief, you should). He said people have been spending the past month "chewing glass". It's crunchy, it's sharp, it's uncomfortable. You'll inevitably bleed a little. But at the end you have something that works for you (but doesn't work for this metaphor, unfortunately).

So don't borrow or buy someone else's perfect system.

Chew glass.

That's my advice for OpenClaw.