Build a golden test set instead of eyeballing prompt changes

You keep saying a prompt looks better without ever checking it against the same real questions twice.

BeginnerRead 5 minDoing: two to three hours, mostly gathering questions and writing answers6 steps

What you'll have at the end

A single file holding 25 real questions, each paired with a checked, correct answer and, where one exists, the source page it came from.

Before you start

You need

A running LLM-powered assistant or prompt the reader already talks to, plus somewhere real questions already exist: a support inbox, a team chat log, or your own early manual tries at it.

Not covered

This does not cover automatically grading the answers or comparing two prompt versions against each other, only building the trusted question-and-answer file those tasks would run against.

Leans on

Collect real questions

Pull questions from where people actually ask them
Pick a spread across your real failure types

Write and source the answers

Ask the model to draft, then correct it
Note the source page for grounded answers

Turn it into one trusted file

Save every row in one fixed shape
Have someone else check a sample

Where to go next

Write your first LLM-as-a-judge grading script

for turning this finished file into an automatic pass or fail score instead of reading every answer by hand

Checked 15 Aug 2026

Part of the AI-LLM App Evaluation cookbook