Simon Willison explores in depth Claude's new "skills" system, revealing a sophisticated document manipulation architecture that transforms the AI assistant into a powerful tool for automated file processing. His methodical investigation demonstrates how savvy users can discover and exploit hidden but documented capabilities of modern AI systems.
Willison's central discovery concerns the /mnt/skills folder, a structured collection of pre-written Python scripts that give Claude specific "superpowers." These skills cover the most common professional document formats: Word documents (docx), PDF files, PowerPoint presentations (pptx), and Excel spreadsheets (xlsx). Each skill represents a specialized capability, implemented via optimized scripts and established Python libraries.
The discovery methodology itself is instructive. Willison did not dig through closed source code or use reverse-engineering techniques. He simply asked Claude to describe its own capabilities, then requested a zip archive of the /mnt/skills folder. This direct approach worked, revealing that Anthropic designed these skills to be discoverable by curious users.
Examination of the scripts reveals careful technical implementation. For example, the PDF skills use pypdf, a robust Python library for PDF manipulation. A particularly interesting use case demonstrated is automatic filling of PDF form fields - a common but tedious task in professional workflows that Claude can now automate seamlessly.
These skills are an integral part of Claude's Code Interpreter, the feature that allows the assistant to execute Python code in a secure environment. This integration means users can benefit from these advanced capabilities without installing additional software or managing complex dependencies. Availability on the iOS app extends these capabilities to mobile devices, making sophisticated document manipulation accessible on the go.
In a commendable act of transparency, Willison published the content of the skills on GitHub, allowing the community to examine, understand, and potentially draw inspiration from these implementations. This open approach contrasts positively with the opacity often associated with proprietary AI systems.
The article also reveals an interesting design philosophy at Anthropic. Rather than hiding these capabilities in a "black box," the company structures them in a discoverable and understandable way. This approach not only improves transparency but also allows advanced users to better understand how to formulate their requests to obtain the best results.
The notion of "skills as superpowers" represents an important conceptual shift. Instead of treating AI as a monolithic general-purpose system, this modular approach allows thinking of AI capabilities as a collection of distinct specializations, each with its own specific strengths and limitations.
The practical implications are considerable. Professionals who regularly work with documents can now automate repetitive tasks through natural language conversations with Claude, without requiring programming skills. This democratization of document automation could significantly improve productivity in many professional contexts.
In conclusion, Willison's exploration reveals that Anthropic is building a sophisticated and discoverable skills infrastructure, positioning Claude not only as a conversational assistant but as a robust and extensible document automation platform.