Vercel’s AGENTS.md beats skills in evals

This is less a win for markdown than a win for removing an unreliable agent decision point. Vercel’s result matters because it turns the AGENTS.md vs. skills debate into a measurable reliability question instead of a vibes-only workflow preference.

• –The key failure mode was invocation: Vercel says the skill was never triggered in 56% of eval cases, which erased most of its theoretical benefit.
• –The winning setup was not dumping full docs into context, but a compressed index that points the agent to local, version-matched `.next-docs` files when needed.
• –Vercel found prompt wording was brittle: “explore project first, then invoke skill” beat more forceful phrasing, a sign that current agent behavior is still fragile.
• –The broader takeaway for framework authors is to optimize for retrieval-friendly context that is always present, not just tools that agents are supposed to remember to call.
• –Hacker News discussion around the post zeroed in on the tradeoff: passive context improves adherence now, but skills still matter for larger toolchains where context budgets are tight.