SkillsBench shows curated skills beat self-generated

This paper lands a clean shot on a big 2026 assumption: agents do not magically write the playbooks they need to follow. For AI builders, the win is not “more autonomous prompting” but higher-quality, modular procedural knowledge.

• –SkillsBench frames skills as a separate abstraction layer above models and agent harnesses, which is a useful mental model for anyone building agent systems
• –The strongest practical result is cost leverage: smaller models with the right skills can match or beat larger models without them
• –Gains vary sharply by domain, from modest improvements in software engineering to huge jumps in healthcare, which suggests skills matter most where tacit workflow knowledge dominates
• –The paper’s finding that focused 2-3 module skills beat broad documentation is a warning against dumping giant SOPs into context windows and calling it memory
• –Theo’s takeaway is directionally right: if you want better agents, invest less in self-generated “meta-prompting” and more in curated, testable operating procedures