How to Fix AI Coding Agents That Reflexively Agree With Everything You Say

AI coding agents have a well-known habit of opening responses with affirmations like 'You're absolutely right' before actually checking any code, frustrating developers who rely on them. The instinctive fix is to add an instruction telling the agent to stop placating, but this rule then loads into the model's context on every single turn, costing tokens regardless of relevance. The article distinguishes between two types of rules: those tied to a specific file path and those governing general behavior, arguing each must be handled differently. As instruction files like CLAUDE.md grow over months, hundreds of unrelated rules end up loaded even for trivial tasks like fixing a README typo. The proposed solution involves smarter context engineering — loading rules only where they apply — rather than relying on blanket instructions that inflate cost and clutter every interaction.