They say don't build alone, so now it's me and Claude. Turns out two heads think better than one, even though one of them is technically just a huge spreadsheet. Which is which? I can't even tell

How I got here

A long time ago, in my place in this galaxy, I was just a developer with several ideas for side projects. Life always got in the way and I ended up starting a bunch of them, only for them to gather dust in my github unfinished and craving attention. Life continued happening; kids, dogs, friends and all of a sudden, I’m no longer a developer, and I've started feeling itchy to get some code done.

That’s when I met Claude. We did meet professionally, but I was very quick to adopt Claude in my daily life.

Yes, it’s true. I’m guilty of the vibe-coding sin (letting the AI run wild with a vague prompt while I nod along and sip on some drink), but I’m doing it responsibly (I think), and I want to share how I do it.

The main problem I started noticing

I would be very happy using /plan mode to be somewhat responsible and get a decent preview of what was about to happen and I’d get so lost in the technical details that I would end up losing track of some of the decisions I made before. Actual product decisions, that is. Things that would greatly impact the product quality. That’s where my professional experience came in clutch.

The files we love to hate, PRD

That's right. I started writing very minimalistic Product Requirement Document (PRD) files that live inside the project's repo. After ensuring that my CLAUDE.md file has enough context, I created a skill that debates the ideas with me. It will then look at every other existing PRD, debate whether what we're writing is a technical or a product decision, and log it. This gives me a sense that I actually have a roadmap. Yeah, the professional life is bleeding a little onto my personal life, but it's all right.

The nice part is that these aren't just notes to myself. Every decision gets written down with its reasoning. Not "we don't use WebGL," but "we removed WebGL for performance, and here's what we gave up." Six months from now the why is still sitting there, so I don't end up relitigating something I already settled, or worse, quietly undoing it.

That reasoning is exactly what stops things from sliding backwards. My CLAUDE.md has a little "do not do this without talking to me first" list, and it's not really for me. It's for Claude. It's a guardrail that stops an eager assistant from helpfully reintroducing the thing I deliberately killed last week.

And because updating the PRD is part of the work rather than a chore for later, it actually stays true. When I shipped the contact form, the decision flipped from "pending" to "accepted" in the same breath. The document reflects reality because I never let it drift.

What ends up happening is that now Claude knows what's got to be done and can take that into consideration when building the next thing. Solo work with an AI is fast, but it's forgetful. The PRD is the bit of process a team would normally get from meetings and old Slack threads, except there's no team. It's just me, a few markdown files, and a pair who needs re-briefing every single morning.

Conclusion

Anyway, I'm really happy with how my side projects have been going, and I'll write about one of them soon. Now I've got to go check on the picanha that's currently cooking away in the sous-vide.