AGR
Menu

Writing

writing

Case studies and field notes on evals, guardrails and the unglamorous engineering that makes AI features safe to ship.

All posts1 entries
Placeholder hero: a pixel-art deployment scene, to be replaced with a regression-guard graphicB-01
Dev LogOct 26, 2026

A Prompt Edit Broke 9 Test Cases. The Merge Gate Caught It.

A one-paragraph prompt edit looked harmless and broke 9 test cases. How I built a CI gate that treats prompts like code and blocks the merge.

Read →