Fixing one prompt failure can silently break another: output budgets make prompt patches zero-sum

ClickOk5811 · reddit · 2026-09-26

Adding a security instruction to a review prompt fixed missed hardcoded credentials — but two weeks later style feedback nearly vanished on large PRs. The cause: model output has a fixed budget under 'be concise', so weight added to security was silently taken from elsewhere, with no error or diff signaling the tradeoff. Natural-language prompts are one blended paragraph, not independent rules. The proposed habit: keep a small fixed set of test inputs and rerun all of them after every prompt change — most prompt regressions ship invisibly because only known-broken cases get checked.

Original post →

More from coding & agent

coding & agent channel →