Write the receipt first

Most people write the prompt, run it, and squint at the output to decide if it worked.

That squint is the problem. It feels like judgment. It's actually a coin flip you're calling after the fact — moving the goalposts to wherever the model happened to land. You ask for a customer email, you get something fluent, and fluency reads as correct. So you ship it. Then it goes out in your voice, on your work, and you find out three replies later that fluent and right were never the same thing.

The platform trap

The market has noticed. The whole field spent the first half of this year flipping its order of operations — define what good looks like, then write the prompt to hit it. They call it eval-first. The platforms have rushed in to sell it back to you: branching, dashboards, fifty metrics, drift alerts. For a team of forty, fine. For the rest of us, that's a cannon aimed at a coffee cup.

Here's the part the tooling crowd skips. An eval is not software. An eval is a receipt — a plain line, written before you prompt, that says what you'd accept and what you'd send back. You already do this for everything else you buy. You know a clean invoice when you see one. You're just not writing it down for the model, so the model never knows the floor you're holding it to. Neither, half the time, do you.

What to run Monday

Write the receipt first and three things change.

1. Name the floor before you prompt. One sentence per task — a good intake reply confirms the date, names the next step, and never quotes a price. Tape it above the prompt. Now good is a fact, not a feeling.
2. Keep five real examples, not a benchmark. Pull five messages from actual work — two you'd send, three you'd reject — and run any new prompt against all five. If it passes the three you'd reject, the prompt is too loose. Five takes ten minutes and catches more than any dashboard.
3. Save the wins as receipts, not screenshots. When a prompt clears the floor, paste it and its five examples into one file. That file is operating IP — it compounds. In a month you have a playbook nobody can poach, because it's built on your work, in your voice.

None of this needs a login. It needs a notebook line and the discipline to write it before you're emotionally attached to whatever the model gave you.

The craft moved. It used to be telling the model what to do. Now it's knowing whether it did it well — and being able to prove it to yourself before you prove it to a customer.

Stop grading on a curve you draw after the test. Write the receipt first.