How One Team's 1,000-Line AI Instruction File Reveals the True Cost of Model Guidance

A content pipeline team built with Claude Opus and Sonnet over seven months accumulated 1,042 lines of model instructions across three files, far exceeding the tool maker's recommended 200-line limit. The project began in January 2026 as a Hugo-based site and progressed through multiple stages of AI adoption, from basic assisted use to supervised autonomy, as described in Boris Cherny's five-step framework. The team's growing instruction files reflect each instance where the model lacked context, but the author argues the real problem is mechanical rules left as readable prose rather than enforced as code hooks or validators. Keeping unnecessary instructions wastes context in every session, while deleting the wrong ones risks silently removing a check the team didn't know they relied on. The article sets out to audit which instructions the model now handles natively, which belong in automated hooks, and which represent facts too project-specific to ever be inferred from code alone.