Evals before vibes: how you actually know an AI feature works
The dangerous moment in any AI project is the demo that goes well. It feels like proof, so you ship-and then a prompt tweak, a model upgrade, or an unusual input quietly breaks something you can't see. The fix isn't more careful eyeballing. It's evals: a repeatable way to measure whether the feature does what it's supposed to, on examples that matter.
Turn quality into a number
An eval is just a set of inputs with a way to judge the output. For a classifier, that's labeled examples and an accuracy score. For something open-ended like grading or summarization, it's a rubric-and often another model scoring against that rubric, or a human spot-check on a sample. The point is to replace "looks good" with a number you can track across changes. Once you can measure it, you can improve it on purpose instead of by luck.
If you can't measure an AI feature, you can't safely change it. Every prompt edit and model swap is a blind bet until you have an eval.
Run them when things change
Evals earn their keep at the moments most likely to break you: a new prompt, a new model version, a new kind of input. Run the eval set, compare the score, and you catch the regression before it ships. It's the AI equivalent of a test suite-and just like tests, the goal isn't perfection, it's not shipping a known regression.
How to start without overbuilding:
- Collect real examples-including the weird ones that already broke once.
- Define "correct" concretely: a label, a rubric, a must-contain check.
- Score automatically where you can; sample for human review where you can't.
- Gate changes on it: re-run before every prompt or model change, compare the number.