Claude Skills are awesome, maybe a bigger deal than MCP (simonwillison.net)

Anthropic announced Claude Skills, a lightweight, folder-based pattern for extending Claude’s abilities: each skill is primarily a Markdown file with frontmatter YAML that explains when it applies, plus optional documents and executable scripts the model can load and run. At session start Claude scans skill metadata (only a few dozen tokens per skill), and it will fetch the full skill contents only when relevant—making the mechanism token-efficient. Anthropic already implemented Claude’s document creation and code-interpreter features as skills (PDF, DOCX, XLSX, PPTX support) and published example skills like a slack-gif-creator that runs Python to build and validate a Slack-ready GIF (including size checks and automatic retries). Technically this is powerful because it couples a model with a filesystem and an execution environment (Claude Code) to create true general-agent behavior: skills let the model run scripts, validate outputs, and iterate—without stuffing huge amounts of context into the prompt. Compared with MCP and plugins, skills are simpler to write, share, and compose and avoid massive token costs, while making it easy to build domain-specific agents (e.g., a data-journalism skillset). The main caveat: skills require safe sandboxing and careful defenses (prompt injection, runaway code). If those safety challenges are solved, skills could enable a rapid ecosystem growth of sharable, executable capabilities.