“Claude Skills很棒，可能比 MCP 更重要”

Core Insights - Anthropic has launched Claude Skills, a new mode that allows its model to acquire new functionalities through the use of Markdown files containing instructions, scripts, and resources [1][3][6] Summary by Sections Skills Overview - Skills are organized folders containing a SKILL.md file that provides instructions for agents to perform additional functions [3] - The new document generation feature of Claude is implemented through Skills, which now includes support for .pdf, .docx, .xlsx, and .pptx files [3][6] Practical Application - An example of a skill, slack-gif-creator, is designed to create GIFs optimized for Slack, including size validation [4] - The process of generating a GIF using the slack-gif-creator skill is straightforward, with the model checking file size to ensure it meets Slack's requirements [8] Technical Implementation - Skills rely on the model's ability to access the file system and execute commands in a coding environment, distinguishing them from previous large model extensions [9] - The implementation of Skills allows for easy iteration and improvement, making it a powerful tool for automating tasks [6][9] Comparison with MCP - Skills are seen as a more efficient alternative to the Model Context Protocol (MCP), which has limitations such as high token consumption [14] - Unlike MCP, Skills allow for direct task execution through simple Markdown files, reducing the need for extensive token usage [14][17] Future Potential - The potential for Skills is vast, with expectations for a significant increase in the number of Skills available, both as single files and more complex folders [15][16] - Skills can be integrated with other models, enhancing their functionality and usability across different platforms [15] Simplicity and Effectiveness - The simplicity of Skills is highlighted as a key advantage, allowing for easy implementation and execution without the complexity of traditional protocols [17] - Skills focus on providing text-based instructions that the model can interpret and execute, aligning with the essence of large models [17]