How I actually use AI to build at night (not the hype version)

I keep seeing videos like "I built a full SaaS in 20 minutes with AI" and every time I watch one I'm like... okay but then what. Then what did you do for the next six weeks when the thing needed to actually work. That's not what building is. I want to talk about what it actually looks like when you use AI to build real apps and you have three kids and you can maybe get two focused hours on a good night.

The workflow I use is built around one core idea: I need to be able to kick something off, go handle bedtime, come back, and see real progress instead of a blinking cursor. That's the actual requirement right. Not "ship fast," not "10x productivity" — just, I have a finite window and I need every minute of it to count.

Claude Code is how I do that. But it's not magic and it doesn't think for you, and I want to be specific about what it actually does because I think people either oversell it or dismiss it and both are wrong. What it does is replace the typing. It replaces the "okay I know what I need to build now let me write all the boilerplate" part. It replaces the "let me look up the exact Supabase RLS syntax again" part. It does not replace knowing what you're building or why or what the architecture should look like. If you don't understand what you're doing, you're going to get AI slop — code that looks like it works, passes a quick test, and then quietly breaks in production in a way that takes you twice as long to debug. I've seen it happen, I've done it to myself early on. The AI is only as good as the thinking you do before you hand it a task.

So here's the actual setup. Every project has a CLAUDE.md file at the root. That's like a brain dump for the AI — it tells Claude what the project is, what the tech stack is, what patterns we use, what we never do, which files are off limits, all of it. When I start a session I don't have to re-explain anything, it already knows. Then I have skills — these are reusable instructions for common tasks I run across projects, like "how I set up auth" or "how I write API routes" — and I have agents for different types of work, like a content agent, a code review agent, a database agent. They each know their job.

The part that makes the two-hour window actually work is the hooks and agents together. I can describe a feature — like "add a filtering system to the content bank that lets me filter by hook type and platform and date range" — and kick off an agent to build it. I go do bedtime, which is a whole production you know, soccer bags, teeth brushing, someone always needs water, the usual. I come back maybe 40 minutes later and I have a diff to review. I'm not starting from nothing, I'm reviewing and steering. That's a completely different use of my time and it fits how my nights actually work.

I built Elevar this way — that's my interview prep app, 105 LeetCode-style problems with spaced repetition built in. And I built Fokas this way. Those are not toy projects. They have real architecture, real security, real data flows. What AI did was compress the time between "I know what to build" and "there is working code in front of me." It did not compress the time I spent thinking about the data model or how RLS should work or what the user flow needed to be. That thinking is still yours. You can't outsource that and you shouldn't want to.

The other thing I'll say is this right — AI makes you faster proportionally to how much you already know. The more you understand the stack, the better your instructions are, the better the output is. This is not a tool that levels the playing field for people who don't know how to code, I'm sorry if that's not what you wanted to hear. It's a tool that amplifies what you already have. If you're learning, it can help you learn faster — I actually think it's incredible for that — but you have to be in it, reading the code, asking why, not just accepting the output and moving on.

Claude Code. It's free to start — you can use it with the Claude.ai subscription you might already have. The thing that makes it different from just using Claude in a chat window is that it has actual file access and can run commands in your terminal, so it's operating inside your real project, not just describing what code might look like. The gap between "here's a code snippet" and "I'm actually making changes to your files and running your tests" is significant. That's the whole thing.

If you want my exact setup — the CLAUDE.md template I use, how I structure agents and skills, the hooks I have running — that's what I built the Claude Code Kickstart course to cover. It's 27 dollars and it's the actual workflow, not a watered-down version of it.

The Claude Code Kickstart is at yessieperez.com — 27 dollars, my exact setup, agents and hooks and CLAUDE.md and all of it. If you're building nights and weekends with limited time, this is the place to start.

Yesenia M. Perez Lead Backend Engineer. Mom of 3. Building at 11pm.

yessieperez.com | IG @yeseniavperez | TikTok @iamyessieperez | LinkedIn /yessiemalone