First-run with agent skills from Anthropic (macwright.com)

Anthropic’s new “Agent Skills” are lightweight, file-based plugins for Claude that bundle instructions, optional code, and metadata in a Markdown file with a YAML header. The YAML tells the model how to use a particular command or tool so the LLM can load just the skill it needs instead of bloating prompts with long guidance—a practical workaround for current LLM context limitations. Skills can execute code either locally (via Claude Code, which runs on the user’s machine and can call any binary in PATH) or remotely on an Anthropic agent VM (a minimal sandboxed Python environment with only a few preinstalled packages). A first-hand test showed the feature is promising but rough: Claude generated a nonfunctional skill (wrong file path and missing YAML), requiring manual fixes. More broadly, the approach raises reproducibility, dependency-management, and security questions: local skills can access arbitrary system binaries while remote skills can’t install dependencies, making a single portable “skill” hard to guarantee. Robust solutions (Nix, Docker, or stronger sandboxing) would improve safety and reusability but reduce accessibility. For the AI/ML community, Agent Skills point toward useful modular tooling for agent workflows—but teams should treat them cautiously until Anthropic tightens UX, packaging, and security controls. Simon Willison’s deeper writeup is recommended for practical details.