System prompts are constraint specs, not descriptions: a 4-step framework to stop LLM degradation in production

ClickOk5811 · reddit · 2026-07-31

Based on experience building multiple LLM-backed systems, the author argues system prompts should be treated as constraint specs, not role descriptions. Common failure: passes demos but breaks in production when users phrase things unexpectedly. Four elements: 1) narrowly define scope; 2) verbatim decline behavior; 3) output format as a rule, not an example; 4) explicit ambiguity handling. Most 'model got worse' complaints stem from undefined implicit decisions. Includes sample prompt diff and discussion on versioning/testing prompts like code.

Original post →

More from coding & agent

coding & agent channel →