I vibe coded a recipe scaler in an afternoon. Here’s where it went wrong.
The premise was quite simple. Paste in the recipe, pick the number of servings, and obtain a scaled quantity. I told Claude Code about it in the morning and we had a prototype up and running by lunch.
And when I put my grandmother’s flatbread recipe through the test, everything seemed to go south fast. It would scale “a pinch of salt” into “three pinches of salt,” but that’s incorrect. Two eggs could be doubled to four, then tripled to six and halved to one, yet “half an egg” isn’t even a thing! And the AI had nothing to say about it. It would convert 3/4 cups into 0.75 cups.
It wasn't tricky prompting either. I made up the rules myself first: round eggs to integers, respect proportions of spices without falling below a pinch, preserve fractions as fractions, and anything that requires human sense-making. Then I handed the rules off to the agent and had it make up tests for each of them. It found another couple of bugs I hadn't considered.
So here's my lesson learned: vibe coding is quick in the happy path but silent about the edges. The edges contain the domain knowledge, and the AI lacks mine. Have you experienced something like that yourself, where everything worked great until the application met some actual data? How do you find those edges before your users do?
Replies
Be the first to reply
Have a question or a thought to share? Add a comment above to start the conversation.