Evaluating an LLM feature: start with 50 real examples, not a benchmark

Teams often pick a model from a public leaderboard and ship. Instead:

1. Collect 50 real inputs from your users.
2. Write down what a good answer looks like for each.
3. Run every candidate prompt and model against them and grade the results (by hand at first).
4. Keep the set and re-run it on every prompt change.

This catches regressions benchmarks never will.