A recent study concluded that agent-generated skills are useless, but this is criticized as a misunderstanding of their nature. The correct use of skills focuses on helping the agent recognize gaps in its knowledge.
A recent paper titled 'Self-generated Agent Skills are useless' gained attention on Hacker News. The author argues that while they derive great value from agent skills, they consistently see peers misusing them.
The study's approach involved asking the agent to explain a task it cannot solve well on its own before attempting it. The author criticizes this as reinventing 'thinking blocks but worse,' as it forces the agent to articulate a solution without first struggling through the problem. Skills are defined as markdown files with metadata, often part of a folder structure that provides both instructions and tools to the agent.
The core of effective agent skills is helping the agent recognize gaps in its knowledge during problem-solving. Simply asking an LLM a question verbatim and pasting its answer is ineffective; skills should be created when the agent already has the capability to perform the task, not as a substitute for its own reasoning.