2025-03-10 · Min-jun Park

Measuring Prompt Quality Without Guesswork

Teams often debate prompt changes in chat threads. We teach a simple harness: frozen inputs, expected constraints (JSON schema, word limits), and human-labeled severity for failures.

The lab exports CSV templates compatible with spreadsheet review and a lightweight Python runner. You tag regressions before merging prompt version bumps.

Safety filters get the same treatment—blocked topics, PII patterns, and escalation paths are test cases, not policies buried in PDFs.

This article mirrors week three content. It is not a substitute for mentor review of your domain-specific edge cases.