Blog / Build with AI /

When to stop vibing and start engineering

You can build a surprising amount by describing what you want to an AI, running what comes back, and pasting the error messages back in until it works. A new name for it has been doing the rounds this month: vibe coding. You go with the flow, accept the suggestions, and stop reading every line.

My argument is simple. Vibe coding is a fine way to start and a risky way to finish. The moment other people depend on what you’ve built, you need to change gears, and the hard part is noticing when that moment arrives.

Why vibing works so well at the start

Let me be fair to it first. When the stakes are low, going by feel is a great way to learn and to move fast:

  • A tool only you will use, such as a script that renames files or a page that converts units.
  • A prototype to find out if an idea is even worth building.
  • A weekend experiment where the worst outcome is a wasted afternoon.

In those cases, reading every line is a waste. The point is to see something working and learn from it. I’ve written before about building a tiny tool with AI, from idea to working page in an afternoon, and the same spirit applies. Speed beats polish when you don’t yet know what you’re making.

The five signs it’s time to engineer

1. Someone else is now using it

The moment a colleague, client or customer relies on your tool, a bug stops being your private annoyance. It’s their lost work. That changes how careful you need to be, even if the code hasn’t changed at all.

2. It touches money, personal data or logins

Anything involving payments, customer details, passwords or private files needs a human who understands what the code does. AI-written code can look right and still leave a door open. Security basics for AI-built web apps lists the usual gaps.

3. You can’t explain what it does

If the AI changed ten files and you can’t say what any of them is for, you’re no longer in control of the project. You’re a passenger. That’s fine for a toy. It isn’t fine when something breaks at 11pm and the AI’s fix breaks something else.

4. Each fix breaks something new

This is the classic spiral. You ask for one change and a different feature stops working. It usually means the structure has gone wrong, and more prompts on top will only deepen the mess. Stop adding and start tidying.

5. You’re afraid to touch it

If you avoid changing the project because you don’t know what will happen, it has already become fragile. That’s the clearest signal of all.

What “engineering” actually means here

It doesn’t mean becoming a senior developer overnight. For a small tool, it means a handful of boring habits:

  1. Write down what it should do. A short spec, a few lines, before the next prompt. See write a spec before you ask AI to code.
  2. Use version control. Save a snapshot you can go back to. Git basics for non-developers is enough to start.
  3. Test the things that matter. Even a short list of “if I do this, I should see that” catches a lot. See testing AI-written code when you’re not a developer.
  4. Make small changes. One request at a time, checked before the next one.
  5. Ask for a review. Either from a human developer or, at the very least, by asking the AI to explain its own code and look for weaknesses. Asking a developer to review AI-written code covers how to approach someone.
  6. Keep secrets out of the code. Securing AI-built tools goes through it.

None of that is glamorous. All of it is cheaper than the alternative.

The counter-argument, answered fairly

The strongest argument against me goes like this: “AI is improving so fast that the code will be better next year, so why bother engineering now? Just keep vibing and let the tools catch up.”

There’s something in it. The tools are getting better, and the amount of manual care needed for small projects will probably shrink. But the problem was never only the quality of the code. It’s about responsibility. Someone has to decide what the tool should do, notice when it’s wrong, and take the call when a customer complains. A better model doesn’t take that job. It only makes the first draft cheaper.

There’s also a practical point. Even if the AI writes perfect code, you still need to describe what you want precisely, and you still need to check that what you got is what you asked for. That’s the engineering mindset, whether a person or a model types the semicolons.

A website example

Say you’ve used AI to write a snippet that adds a custom banner to your WordPress store. Perfect for you. Then a client asks for the same thing, and then three more do. Now it runs on other people’s live shops, through updates you don’t control. At that point it needs a changelog, a test on a staging copy, a backup before each update, and a way to turn it off. Same code, different responsibility. For anything that touches a real site, see theme tweaks with AI: child themes and safe edits.

A quick test

Ask yourself three questions:

  • If this broke tomorrow, who would be affected?
  • If it broke, could I fix it without asking the AI?
  • If I left it for six months, could I understand it again?

If the honest answers are “other people”, “no” and “no”, it’s time to engineer.

Key takeaways

  • Vibe coding is great for throwaway projects, prototypes and personal tools.
  • Switch gears when other people rely on it, when money or private data is involved, or when you can’t explain it.
  • Engineering for a small tool means a spec, version control, tests, small steps and a review.
  • Better AI makes the first draft cheaper. It doesn’t take over responsibility for the result.

Where’s your own line? Is there a project you’ve built by feel that’s quietly become something people depend on?