AGENTS.md outperforms skills in our agent evals

What this post added

This post introduces and validates a new method for providing framework-specific knowledge to AI coding agents: embedding a compressed documentation index directly into `AGENTS.md`. This approach achieved 100% pass rates in evaluations for Next.js 16 APIs, outperforming the 'skills' abstraction which maxed out at 79%. The post details the problem of outdated training data, the two approaches tested (skills vs. `AGENTS.md`), the development of a hardened eval suite, and the surprising results. It explains the benefits of passive context (no decision point, consistent availability, no ordering issues) and addresses context bloat concerns through aggressive compression (8KB). A codemod is provided to automate this setup for Next.js projects.