Why Loading All AI Instructions at Once Hurts Coding Agent Performance

Developers using AI coding agents have long debated how to structure instruction files, with early approaches using nested AGENTS.md files placed at different folder levels to load rules only when relevant. Tools like Cursor and Claude later introduced finer-grained, path-based configuration allowing instructions to be scoped down to individual files rather than entire directories. However, even these improvements do not fully solve the core problem: loading all instructions simultaneously forces the model to spread its attention across every rule, diluting the effectiveness of any single directive. An analysis of 28,721 real-world repositories found that the median instruction file contained around 50 items, but only about a dozen were actual directives, with the rest being scaffolding the model still processes every turn. The concept of progressive disclosure — loading only the instructions relevant to the current task — is presented as the more effective solution to managing AI agent context and attention budgets.