Claude Skills

Simon Willison found the skills on 10 October 2025 by asking Claude to describe its own capabilities, then requesting a zip of the /mnt/skills directory. Inside were scripts Python for docx, pdf, pptx and xlsx, with pypdf behind PDF form filling, shipped as part of Code Interpreter. He published the contents on GitHub.

Six days later he argued that Claude Skills might be a bigger deal than MCP. The case rests on how little they require. A skill is a folder of instructions, scripts and resources, built on fichiers Markdown, and a single file can be the whole thing. Brief descriptions load at the start of a session and the rest loads when relevant, so nothing sits permanently in context. The only infrastructure dependency is a coding environment with filesystem access. Willison expected a cambrian explosion of them, and noted the format is not bound to Claude.

Anthropic applied the mechanism to its own weakness in November 2025. Frontend output kept converging on safe defaults drawn from training data: Inter fonts, purple gradients. A roughly 400-token skill covering typography, color, motion and backgrounds changed results across React, Tailwind CSS, shadcn/ui and GSAP work, without a persistent token cost.

By August 2026 Louis Claxton, of Anthropic's Applied AI team, had given skills a governance job. Brand, security, compliance and UX skills constrain the spec.md a team commits, with the skill as the advisory control and hooks as the deterministic layer that blocks or demands approval. Claxton's playbook advances no measured results.