A demo proves a thing can happen once. An eval proves it keeps happening. The gap between those two sentences is where most AI projects die, usually two months after the launch announcement, when nobody can say whether the output got worse or the users got pickier.
So we invert the order. Before a prompt is written, we collect thirty to fifty real cases from the client — the actual messy inputs, including the ones that made someone escalate to a manager. Each case gets an expected outcome, graded by whoever owns the process today. That set becomes the contract.
Once the harness exists, prompt work stops being taste and starts being measurement. You change one instruction, you re-run, you see 0.81 become 0.88, and you keep it. You see it become 0.74, and you throw it away without arguing about it in a meeting.
The second benefit is slower and more valuable: the eval set is the only artefact that survives a model change. When the underlying model is swapped for something cheaper or newer, the set tells you in ten minutes whether the swap was safe. Without it, you are shipping a rumour.