Medical AI Agent Skills Library: Practice of Agent Skills for Radiology and Clinical Workflows

This article introduces a medical AI agent skills library designed specifically for medical imaging and clinical workflows, aiming to integrate large language model capabilities into clinical workflows. The library supports mainstream AI programming assistants such as Claude Code, Codex, Cursor, and Windsurf, covering fields like image analysis, clinical document processing, AI integration, and radiological research, providing a complete AI-assisted toolset for radiologists, medical IT personnel, and researchers.

Agent Skills is an emerging AI application model that provides AI agents with domain-specific expertise and workflows in the form of Markdown files. Compared to traditional prompt engineering, it features structured knowledge, context awareness, composability, cross-platform compatibility, and can automatically identify task types and apply corresponding frameworks.

The skills library adopts a layered architecture, with radiology-context as the foundation. The core skills layer includes environment configuration and modality detection; the clinical document layer covers report analysis, structured report generation, etc.; the patient communication layer supports result letter generation and care gap closure; the AI integration layer involves PACS interaction, DICOM query, and AI detection pipeline; the research layer includes PubMed search, guideline integration, etc.

Skills form an interconnected knowledge network (e.g., closed loop between report analysis and structured report generation, collaboration between AI detection and quality review). Practical scenarios include: radiology report analysis (from environment configuration to standardized report generation), AI-assisted detection integration (algorithm routing to follow-up management), and patient communication (converting reports into plain-language letters and arranging follow-ups).

The sensitivity of medical data requires strict safety measures: protecting patient privacy (prohibiting PHI input, using de-identified data); AI-generated content must be reviewed by clinicians; the llm-radiology-use skill provides hallucination mitigation strategies (prompt engineering, verification processes).

The skills library follows the Agent Skills Specification and is compatible with tools like Claude Code and Codex. Skill files are in Markdown format (including YAML metadata), which are human-readable and AI-parsable. The project is maintained by Corpus Analytica, and community contributions are welcome (PRs, new skills, issue feedback). It uses the MIT license, and the current version is 1.0.1 (26 skills, 34 integrations, 85% test coverage).

With the development of multimodal AI and medical large models, the prospects of the skills library include: image-text fusion analysis, real-time reading assistance, personalized workflow customization, and cross-department expansion (pathology, cardiology, etc.). This skills library represents a new model of AI-assisted clinical practice, which is expected to improve efficiency and ensure quality and safety.