An agent answers engineering questions across a repository of thousands of design documents and source files. The first version loads every possibly relevant file into the prompt up front, and answer quality drops as the prompt grows even though everything still fits in the context window. Which redesign best follows Anthropic's context engineering guidance?