There's a pattern I keep seeing: developer shows off a complex feature they "built in 10 minutes with AI." What they don't show is the three hours they spent debugging the output, untangling hallucinated dependencies, and explaining to the AI — three times — what the actual requirement was.
That's not fast. That's just moving the work around.
I've flipped my approach. AI handles the boring 80% of what I build. The interesting 20%? That's still mine.
What I mean by "boring 80%"
The boring work isn't bad work — it's just predictable work. Writing boilerplate CRUD routes. Setting up test fixtures. Drafting the error handling logic for a form. Writing database migrations. Building out utility functions I've written a dozen times across different projects.
This stuff needs to exist. It needs to be right. But it doesn't need my full brain.
When I feed AI a well-scoped prompt — "here's my schema, write the TypeScript service layer for this entity with proper error types" — it nails it 90% of the time. I skim the output, run the tests, move on. Fifteen minutes of work becomes three.
The trap is thinking you should apply this same leverage to the interesting 20%.
Where AI makes you slower, not faster
The interesting 20% is where product decisions live. It's the part where you're deciding what to build, how it should feel, and why this architecture over that one.
If you're using AI to design your data model, you're outsourcing judgment. The AI will give you something plausible-looking — maybe even technically correct — but it won't know that you're planning to add multi-tenancy in six months, or that your users care deeply about offline mode, or that you have a performance constraint nobody has written down yet.
That context lives in your head. The moment you hand the wheel over, you're building on top of someone else's guess.
I've shipped features three times only to throw them away because I let AI do the thinking. Fast to build, wrong to ship.
The workflow that actually works
My current split looks like this: I think through the architecture in a notebook or scratchpad. I make the product decision. I figure out the interface — what goes in, what comes out. Then I write a detailed spec as a prompt and let AI generate the implementation.
If the output is off, that's usually a signal my spec was vague — which means my thinking was vague. AI is actually good feedback on the clarity of your own ideas.
The meta-skill here isn't "use AI better." It's knowing when your own judgment is the bottleneck versus when the keyboard work is the bottleneck. AI solves the second problem beautifully. The first one is still yours.
One last thing
I'm a solo builder. I don't have a team to rubber-duck ideas with or a senior engineer to sanity-check my architecture. AI fills some of that gap — but only when I treat it like a fast, tireless junior dev, not an oracle.
Give it the work that doesn't require understanding. Keep the understanding for yourself.