Another reason not to front-load a bunch of AI-generated general instructions. > Developer-written context: +4% avg improvement on AGENTBENCH, marginal > LLM-generated context: -3% average, net negative Recent papers show that the declarative context the model can discover itself is redundant and costly. Debating whether AGENTS[.]md files help or hurt seems like the wrong question. A flat, monolithic context file is always going to be redundant or distracting. I think the right architecture is a protocol layer that routes to focused, persona/focus-specific context. The main file should describe how agents should think about your repo. Layer 1: Protocol file (CLAUDE[.]md / AGENTS[.]md) Not a repository overview. Not a style guide dump. A routing document that declares: Available personas/agents and when to use each - Available skills and what task classes they cover - Available MCPs and what they connect to - Minimal essential repo info (only what the model cannot discover) - The dependency graph between these components Layer 2: Focused persona/skill files Each persona (UX designer, information architect, SWE) gets its own scoped context with domain-specific procedural knowledge. The model loads only the relevant one based on the current task, not all of them at once. Layer 3: Maintenance agent (exploratory) A subagent whose role is keeping the protocol file updated as the codebase evolves, so it doesn't go stale. The right comparison would be a hierarchical system that loads the right context for the right task versus dumping everything upfront.