Summary
David Ondrej argues that most people running AI coding agents have far too many skills installed, and that deleting 80 to 90 percent of them can make agents faster and cheaper to run. The video explains what an agent skill is, a reusable description and body of instructions that goes beyond a single prompt, and why hundreds of copied skills quietly consume tokens and context through their descriptions even when never invoked.
Ondrej criticizes the habit of copying viral skill repositories, pointing to a popular Andrej Karpathy skills repo that he says was not actually written by Karpathy. He explains that broad skill descriptions cause skills to trigger too often, filling the context window and pushing users into constant compacting or extra subscriptions. He recommends writing narrow, specific descriptions and building skills around your own workflow.
The second half covers the handful of skills he uses every day. These include a review skill that has a second frontier model from a different provider, such as OpenAI’s Astra, check work done by Claude’s Opus, which he says catches around 95 percent of bugs, and a Deep API skill that gives agents web search, scraping and research through a single interface. He demonstrates the setup using the open-source Cloudroom app.
📺 Source: David Ondrej · Published November 09, 2026
🏷️ Format: Workflow Case Study
