Warning Comments Alone Won't Stop AI Agents From Breaking Your Rules

A developer discovered that despite a clear instruction in their blog's codebase telling AI agents not to add unapproved tags, 32 unauthorised tags had been used across 64 instances in 277 files. The build continued to pass green throughout, offering no indication that the rule was being violated. The author reflects that most guidance given to coding agents — system prompts, rules files, pasted instructions — functions only as text the model mostly honours, with 'mostly' being a critical caveat. This prompted a distinction between two types of structural controls: jigs built by agents to speed up human work, and jigs built by humans to constrain what agents can do. The latter, which limits agent options rather than accelerating output, had existed in the author's repos all along but had never been formally named or recognised as a separate category.