January 29, 2026
Sticky notes beat plugins
Compressed Agents.md > Agent Skills
Tiny file humiliates fancy AI skills, commenters howl
TLDR: A tiny 8KB docs index hardwired into an AI’s memory beat fancy “skills,” which the bot often ignored. Commenters joked it’s classic dev behavior, predicted new features, and argued we need smaller models to route docs—until then, stick the instructions in the bot’s face.
An 8KB cheat sheet stuffed into AGENTS.md just outscored fancy "skills" by a mile, and the crowd is cackling. The team tested two ways to teach coding bots about new Next.js 16 features like “use cache” and “forbidden.” Skills are like a plug-in toolbox the bot can open on demand; AGENTS.md is more like a sticky note glued to its brain. The shocker: the bot ignored the skill in 56% of tests. Meanwhile, that tiny docs index sitting in its head? 100% pass rate. Cue the memes.
Commenters brought the popcorn. One joker said the bot “passes the Turing test” by acting like a confused junior dev who won’t read the manual. Another predicted the inevitable product launch: “Two months later: Anthropic introduces ‘Claude Instincts’”. A confused reader asked why this isn’t just how skills already work with “progressive disclosure,” while others argued the model simply isn’t trained to use skills yet—give it time. The real drama: subtle instruction wording swung results wildly, making the skill approach feel brittle. Techies pitched a fix: small routing models that pick the right docs before the big model does anything—it’s “silly” to waste tokens on dithering. For now, the community’s verdict? Glue the docs to the bot’s forehead and ship. skills directory
Key Points
- •An embedded 8KB Next.js docs index in AGENTS.md achieved a 100% pass rate in evals.
- •Agent skills were not invoked in 56% of cases, yielding no improvement over baseline and slightly worse test results (58% vs 63%).
- •Explicit instructions to use the skill increased invocation to 95%+ and improved the pass rate to 79%.
- •Agent behavior was highly sensitive to instruction wording, affecting coverage of required changes (e.g., next.config.ts).
- •The team began hardening the eval suite after identifying issues with ambiguous prompts and tests focused on implementation details.