AI stays a pilot when nobody defines what working means

Most AI projects fail before they start. Not because the model is wrong — because nobody wrote down what success looks like.

I’ve watched teams spend months picking between GPT and Claude while the actual problem — what does the system need to produce, and how will we know it got there — stays undefined. The model is the easy part. The specification is the work.

Spec-first isn’t a methodology. It’s the same discipline that made the service business work: before you touch anything, write down what done means. Requirements. Criteria for success. The edge cases that matter. Then build to that, and verify against it.

Everything else is decoration.