Use generation to accelerate structure, not judgment
Local models are useful for first-draft normalization. They can cluster raw engineering notes into headings, infer repeated themes, and surface missing fallback rules faster than a human starting from zero.
What they cannot do reliably is decide which instruction should become policy. That remains an editorial job.
Prompt the model with hard boundaries
The fastest way to get bad rule files is to ask for 'best practices' without repository context. Give the model known directories, coding stack, testing stance, and banned patterns so its output has sharp edges.
Review for collisions before adoption
Generated rule files often look coherent while hiding subtle contradictions. Before adoption, scan for duplicated policies, vague style guidance, and rules that are really task-specific but accidentally written as repository-wide constraints.