The article outlines a surprising finding regarding how AI coding agents handle framework-specific knowledge, particularly with Next.js 16 APIs. Initially, the authors believed that using "skills," which bundle documentation and prompts for the agents, would help them generate accurate code. However, tests revealed that skills were often not invoked by the agents, achieving only a 79% pass rate compared to a 100% pass rate when a compressed documentation index was embedded directly in AGENTS.md, a markdown file that provides persistent context to the agents. The authors identified two main issues with skills. First, in 56% of cases, agents simply did not use the available skills. Even with explicit instructions added to encourage their use, the performance remained inconsistent, influenced by subtle changes in wording. In contrast, using AGENTS.md eliminated the decision-making process for agents, providing instant access to relevant documentation. The study showed that when the documentation was directly included, the agents performed flawlessly, highlighting how passive context can outperform more complex retrieval methods. To address concerns about context bloat, the authors compressed the documentation index from 40KB to 8KB, allowing agents to access necessary information without overwhelming their context window. The article concludes with a command that developers can run to set this up in their Next.js projects, streamlining the integration of version-matched documentation for coding agents. This approach simplifies the interaction between the agents and the documentation, ultimately leading to more reliable code generation.