I use AI a lot. But if I’m honest, I’m less interested in AI in general and more interested in Claude specifically and more specifically still, in the way Claude works when you actually set it up properly.
Agents, skills, commands, CLAUDE.md files. It sounds fiddly. And compared to just opening a chat window and typing “build me a thing,” it is at first! But vibe coding has a habit of quietly going sideways. Security decisions get skipped because nobody told the model to care. Scope creeps because there’s nothing to stop it. You end up with something that roughly works and a vague sense that you’re not entirely sure what it actually did.
The markdown approach is different. You’re not fighting the AI or reining it in — you’re telling it what kind of collaborator you want it to be. Here’s the context. Here are the boundaries. Here’s the voice. Here’s what matters. The model doesn’t push back on that. It just works within it, and the output is noticeably more focused for it.
I think the distinction is between using AI as a shortcut and using it as a tool. Shortcuts are fine until they aren’t. Tools work the way you built them to work.
The guardrails aren’t the boring part. They’re kind of the whole thing.
Leave a Reply