The Project Nursery

AI is pretty good at copying patterns. It's not as good at inventing them. Give it a blank slate and ask for "a service layer," and you'll get something generic... probably fine, but not how you'd do it yourself. Point it at code that already works the way you like, and it'll copy that instead.

This is obvious in hindsight, but it took me a while to discover. I kept starting new projects from scratch, describing what I wanted, getting okay-ish results. Meanwhile I had several working projects with patterns I'd refined over time just sitting there.

---

The Setup

The trick that made this click for me was symlinking my existing projects into the same parent folder as my new one:

mkdir ~/projects/nursery
cd ~/projects/nursery
ln -s ~/projects/flow-myna/mono flow-myna
ln -s ~/projects/leveloh leveloh

Then I opened that folder in Cursor. Now when I say "build the auth context like flow-myna does it" or "use the same alembic migration setup as leveloh," Cursor can see those folders/files. I'm not suggesting it copies and pastes... it reads the patterns and applies them to the new context.

One nice thing about symlinks specifically: Cursor will ask for confirmation before editing files in a symlinked folder. So your reference projects are protected... the AI can read them but won't accidentally modify them.

---

Your Projects as a Style Guide

I think of this folder as a nursery for new projects. The other projects aren't templates... they're just real working code that happens to be visible. They've diverged in different directions based on what each project needed.

This is different from maintaining a boilerplate repo. With a boilerplate, you're trying to anticipate what you'll need and abstract it into something reusable. That's a lot of upfront work, and six months later it's outdated anyway.

Here, there's nothing to maintain. Each project is just itself. When I start something new, I tell the AI "look at how this other project handles X and do something similar." It handles the copying, adapting as needed.

---

Staying in Your Own Stack

I've found this works better than pointing the AI at some popular starter template. When it's my code, I already know the patterns. I know why things are structured the way they are, where to look when something breaks.

With someone else's boilerplate, you're learning their conventions while trying to build something. The AI doesn't know which parts are load-bearing and which are just preference. It's easier when everything comes from your own brain.

My stack for context: Next.js 14 with Tailwind on the frontend, FastAPI with PostgreSQL on the backend. Nothing unusual... just things I've used before and have working examples of.

---

The Result

I'd been sitting on a domain for a simple metrics service idea for years. A service where users POST a chart name and value, and charts appear automatically. It felt like a multi-week project, so I kept putting it off.

With this setup it took about 8 hours spread across a few days. The service is at spikelog.com if you want to take a look.

I don't know if this generalises to everyone's workflow, but the main idea seems sound: the patterns you've built up over time are worth reusing, and making them visible to the AI is an easy way to do that.