Agent Skills: Where They Help and Where They Hit a Wall
Princeton and UC San Diego ran 8,000 tests and found that skills for AI agents work as structured playbooks, not knowledge bases. With 100 skills in the library, the right one is retrieved only 3.3% of the time.
A team of researchers from Princeton University and UC San Diego conducted over 8,000 experiments to understand why AI agents require “skills” — a set of predefined instructions for certain actions. They found that having a well-structured sequence of actions within a skill can help solve a task in 65.7% of cases, whereas transferring knowledge only works in 4.5% of situations.
However, when you increase the size of your skill library (e.g., going from 5 to 100 skills), the retrieval rate decreases significantly (from 29.6% to 3.3%) and there’s a chance (one in ten) that the agent will automatically apply the incorrect instruction. So skills are essentially action recipes rather than knowledge repositories, and this is why building large libraries of skills has always been challenging.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles