Skip to content
Use cases
Studioplanned
Studioplanned

Score a prompt change before you ship it

Run the same set against the old and the new prompt on identical hardware, then read the delta rather than trusting a hunch.

Everything below describes what this is being built to do. It needs Studio, which you can look around today but cannot yet use for the whole of this job. The problems are real now; the answers are the part still being built.

Everyone can feel that a prompt change helped. Almost nobody can show it.

Why it is hard

  • A before/after is usually a vibe

    One run on the old prompt and one on the new, read side by side, is a comparison of two anecdotes, not two samples.

  • Hardware drifts between runs

    Comparing runs from different sessions on different cards adds a second variable to the one actually being tested.

  • A single score hides where it changed

    An aggregate number can rise while it gets worse on the cases that matter most, and nothing below it says which.

What answers it

  • Two arms, one set, identical hardware

    The same input set runs against both prompts on the same card class, so the prompt is the only thing that changed.

  • One scoreboard, not two logs

    Both runs land in one place built to be read against each other, rather than two exports somebody diffs by hand.

  • The delta is the answer

    The comparison reads as a delta per case, not one aggregate number, which is the part not built yet.

The two-arm shape is real today in the Cookbook. Scoring one automatically is next.