Prompt complexity in production environments can completely neutralize structured reasoning frameworks like STAR, dropping accuracy from 100% to 0%.
arXiv · March 17, 2026 · 2603.13351
The Takeaway
This is a critical warning for practitioners: structured reasoning prompts that work in isolation fail when surrounded by other instructions (like style guides). It proves that directive-heavy system prompts force 'conclusion-first' outputs, which prevents the model from utilizing its internal reasoning tokens effectively.
From the abstract
In a previous study [Jo, 2026], STAR reasoning (Situation, Task, Action, Result) raised car wash problem accuracy from 0% to 85% on Claude Sonnet 4.5, and to 100% withadditional prompt layers. This follow-up asks: does STAR maintain its effectiveness in a production system prompt?We tested STAR inside InterviewMate's 60+ line production prompt, which had evolved through iterative additions of style guidelines, format instructions, and profilefeatures. Three conditions, 20 trials each, on Claude