Issue 01 › Method

Without this you are not improving it, just moving it.

A fixed set of cases with known-good answers is the dullest part of the work and the only thing standing between you and a year of confident guessing.

We read the footnotes · 5 min read

A grid of bordered cells, some marked with a navy tick, some with a red cross. — The same cases, scored on every change. The crosses are not a failure of the exercise — they are the entire reason for it.

The uncomfortable question

After a change, how do you know it is better? Most teams answer by trying a few prompts and forming an impression. That impression is shaped by whichever examples they happened to pick, and it is almost always flattering.

What an eval actually is

Thirty to fifty real cases, each with the answer you would accept, written down and frozen. Every change gets scored against the same set. The number moves or it does not.

It is not sophisticated. The difficulty is entirely in writing it before you are emotionally invested in the thing working.

Write it first

The rest is for subscribers.

It carries on in the same plain language, with the same drawn diagrams.

One week free, then ₹25