Your system prompt collapses as you stack constraints: follow rate falls from ~96% to 20%, and a single "output JSON" instruction conflicts with nine others

arXiv 2608.02639·high signalXBlueskyLinkedInCopy link

A benchmark stacking 24 verifier-checked instructions across three production-tier LLMs shows instruction-following degrades non-linearly as constraints accumulate — not randomly, but through identifiable pairwise conflicts that are structured and reproducible. The worst offender is the most common one in production: a single "output valid JSON" requirement conflicts with nine other instructions in the stack. A training-free instruction compiler recovers up to +11 points for weaker models but leaves strong models essentially unchanged, so the fix is capability-dependent — worth deploying exactly where cheap models are doing bulk work at scale.