3 Traps I Found When Building with AI

Starting some new projects from scratch changed how I build products with AI

During 2025, I dove into some new projects with no existing codebase, going 100% AI-first. This was very different from my experience adding AI to existing projects that already have a solid foundation, such as documentation and a well-defined codebase.

I relied heavily on my engineering background to keep things on track when building from scratch. But for non-technical builders, it’s a whole different ball game: you’re figuring out what to build and how to build it.

Here are three traps to look out for. One can break your project, another will slow you down, and the third will cost you time and money.

Trap 1: Trusting and Not Verifying

This trap can break your project. AI tends to be a people pleaser: it will tell you everything worked, even when it didn’t really check to confirm that it did.

Why it’s a trap: AI sometimes skips over important verification steps, even if you tell it not to. I’ve seen it tell me that tests passed or tasks were completed, even when things were actually failing behind the scenes. When you’re building from scratch, you might not even know what “working” should look like, which makes it easy to overlook small, hidden issues.

How to avoid it:

• Make documentation a prerequisite. Before the AI implements ANYTHING, have it document the plan. After implementation, ask: “Does the documentation still match what you just built?”

• If you don’t have tests yet, ask the AI to generate them, using well known testing frameworks. Even better, have it write the tests first BEFORE implementation.

• Learn one simple command to run tests manually. It’s important to verify the changes it says are actually there and working as expected to avoid compounding problems making your app unstable.

Trap 2: Including More Than You Need

This trap will slow you down. When you add AI to existing projects, you can spot over-engineering because the AI’s additions feel out of place with your setup, but starting from scratch can make it harder to spot because so much is often happening at once.

Why it’s a trap: Over-engineering is often the result of unclear goals, leading you to implement unnecessary features. If things take off later, you can add scalability, but at this stage, you don’t need Netflix-level complexity. It’s easy to get caught up because something “seems like the right thing to do” or because the AI offers it. But these extra features slow you down and complicate your project.

How to avoid it:

• Break work into small, tracked pieces you understand. Instead of “build authentication,” try “create a login form” then “add password validation.” Have the AI create a task list upfront so you can reference it when you encounter limits or need to resume.

• Ask for the simplest solution that works. Always ask: “Is there an established tool or framework for this?” When established solutions exist, use them, they’ve been debugged by thousands of developers.

• Use version control. Ask the AI to automatically commit its changes when it has completed specific tasks, including clear messages about what changed. These checkpoints let you roll back specific changes without losing too much if something breaks.

Trap 3: Using One Model for Every Task

This trap can cost you time and money. Behind every request you make is a trained AI model trying to understand and execute what you need. It’s tempting to use the same model for everything, but not all models are suited for all tasks, which is why they vary in cost and capabilities.

Why it’s a trap: Lower-cost models often lack the reasoning power required for complex decisions, leading to inefficient code and poor planning that delays work. Meanwhile, using expensive models for routine tasks can drive up your costs without providing any real benefit. When building from scratch, it’s crucial to use the right model for the right task.

How to avoid it:

Think of it like a beehive: you need both a queen bee and worker bees.

• Queen bees (advanced models) should handle the big-picture tasks like planning and architecture, deciding what to build and how it should work (e.g., Claude Opus). Save the plan so both you and the AI can refer back to it.

• Worker bees (simpler models) can handle the implementation, writing the actual code (e.g., Claude Sonnet, GPT-3.5).

• Take breaks. Like human workers, worker bees can get stuck. AI can also get overwhelmed with too many requests. Resist the urge to keep pushing; sometimes taking a break will help you (and the AI) get back on track.

The Bottom Line

AI makes it easy to move fast, but moving fast comes with its own potential challenges. This is how you end up in the 80% trap: your prototype feels almost done, but foundational problems make the final 20% seem never ending.

Bonus: If you’re working with engineers instead of building everything yourself, here’s how to make that collaboration more effective.

If nothing else, make this habit stick: the next time you ask AI to implement something, have it fully document the plan first. This reference will guide both you and the AI moving forward.

Products are a collection of decisions made over time. Written plans help your future self understand the decisions you made, and why.

---