Because: preregistered randomized tests of what makes a coding agent follow its instructions

People write instruction files for coding agents on folk advice: give a reason, write MUST, put the important rule first. I tested that advice on Claude Code in four preregistered randomized studies, with outcomes decided by code. Giving a reason came back inconclusive, and no wording factor survived correction on average. For skills, the largest effects came from the listing's character budget, which controls whether a description reaches the model. The numbers and their caveats are in the paper.