Descriptions:
Dylan Davis, who runs an AI consultancy, exposes a widely overlooked problem with Claude’s skills feature: every active skill loads its title and description into the model’s context on every interaction, and stacking too many causes Claude to route to the wrong skill in ways that are difficult to diagnose.
The video provides a systematic breakdown of when to use Claude projects versus skills, starting with two diagnostic questions around task recurrence and required consistency. Davis explains project architecture — scoped workspaces focused on a single activity rather than a catch-all file dump — and demonstrates why projects scale cleanly while skills in the Claude browser interface do not. The key technical distinction is that Claude Cowork on the desktop allows skills to be scoped to specific subfolders, meaning only the relevant skills are active in a given context, whereas the browser loads all active skills simultaneously regardless of the task at hand.
Davis also shows the internal structure of a skill (title, description, detailed prompt, and optional subfolders), using Anthropic’s built-in skill creator as a real example, and walks through his own personal skill library. For practitioners building multi-skill Claude workflows — especially consultants, operations teams, or power users managing dozens of automations — the browser versus desktop architectural difference has direct implications for reliability and scalability.
📺 Source: Dylan Davis · Published April 25, 2026
🏷️ Format: Deep Dive
