This article examines how instruction templates and system prompts function as a regla compilada that redistributes decision rights in downstream model outputs. Focusing on five institutional domains, we construct a controlled synthetic corpus of 1000 prompts, each executed under five template conditions, and quantify shifts in authority-bearing constructions across 5000 generated texts. A syntactic authority inventory operationalizes seven indicators, including delegation ratio, veto path count, default scope strength, agent deletion rate, deontic stack depth, escalation clause presence and override friction. Using intra-prompt comparisons and mixed-effects models, we estimate the causal impact of template variants on the locus of control, robustness of exceptions and balance between model-proposed and human-confirmed decisions. Robustness checks include template ablations, adversarial prompts and placebo reformulations to disentangle authority effects from style. The article delivers a Template Authority Index, a public test battery and a de-identified dataset to support external replication. The central claim is that seemingly minor changes in instruction templates produce systematic and measurable reallocations of authority in institutional texts, with direct implications for governance, accountability and the design of human–AI decision pipelines. DOI Primary archive: https://doi.org/10.5281/zenodo.17879314 Secondary archive: https://doi.org/10.6084/m9.figshare.30849092 SSRN: Pending assignment (ETA: Q4 2025)