Don't Build Agents, Build Skills Instead – Barry Zhang & Mahesh Murag, Anthropic

The team discovered that code (bash, file system, Python) is not just a use case but the universal interface to the digital world (1:30). After building Claude Code, they realized a coding agent is actually a general-purpose agent capable of API calls, data analysis, file management, and synthesis — making domain-specific agent scaffolding largely unnecessary.

While agents are brilliant general reasoners, they lack the consistent, domain-specific expertise needed for real work (2:30). The analogy of a mathematical genius vs. an experienced tax professional illustrates that raw intelligence without procedural domain knowledge leads to inefficient, inconsistent results.

Skills are organized collections of files that package composable procedural knowledge for agents — essentially folders containing markdown instructions, scripts, and assets (3:15). This simplicity is deliberate: anyone (human or agent) can create them, they work with existing tools like Git and Google Drive, and they leverage the file system as a decades-old proven primitive.

Traditional tools suffer from poor documentation, ambiguity, and immutability within the context window (4:00). Skills solve this by including code scripts that are self-documenting, modifiable by the agent, and stored in the file system until needed — like a saved Python styling script that Claude previously rewrote repeatedly.