With this Orange article, I will continue to share some content about ClaudeSkill and MCP to attract some traffic, haha. Analysis of the core features of Claude Skills. Claude Skills can be understood as a sophisticated and efficient "ability plug-in" system. Its core is to enable the Claude model to call pre-packaged professional knowledge and tools when performing specific tasks. 📷 Claude Skills: Modular and portable, its main features are as follows: Encapsulation and modularization: Skills will perform the instructions, scripts and even executables required to perform specific tasks Code is packaged into a separate "folder" or module. This is like preparing standardized tool kits for AI, such as "Excel report generation skill" or "company brand guideline following skill". It unifies the standards for Agent engineering and reduces the difficulty of implementation. --》In the context of Claude, an Agent can be composed of several Skills. Efficient: The system does not blindly load all capabilities, but automatically scans and calls only the most relevant Skills when processing tasks, loading only the minimum required information and documents. components, thus ensuring operation speed and efficiency. Determined context engineering paradigm: Through SKILL, FORMS, REFRENCE, script and the bound remote virtual machine environment, a complete context is formed from definition, transmission to execution declaration cycle (according to reliable sources, the implementation of Claude Skill deeply draws on Manus). Composable and stackable: Multiple Skills can work together, and Claude can automatically identify and coordinate their use to complete a more complex task, which embodies a "Lego-style" combination of capabilities. Reliable, more permissions: A groundbreaking feature of Skills is that it can contain executable code. For those tasks where traditional programming is more reliable than language model generation (such as precise mathematical calculations, complex data processing), Skills can ensure the accuracy and stability of the results by executing the code, which goes beyond pure token generation. Portability & Cross-Platform: Skills adopts a unified format and can be used seamlessly in Claude's various applications (web version, code assistant Claude Code) and API once built, which greatly improves development and deployment efficiency. MCP: “USB-C” connecting large models to applications Think of MCP as a USB-C interface for AI applications. Just like USB-C provides a standardized way for electronic devices to connect, MCP provides a standardized way for AI applications to connect to external systems. Although MCP is known as a protocol, it is actually a set of standards that connect traditional applications and large models. In-depth comparison and analysis of alternative possibilities Technical level MCP was launched by Claude in December
- After nearly a year of development, the MCP ecosystem has become quite prosperous. However, if you look at MCP from a technical perspective, its shortcomings are quite obvious.
- Most notably in terms of token usage: GitHub’s official MCP is famous for consuming tens of thousands of context tokens when used alone, and once a few more are added, there’s not much room left for LLM to actually perform useful work. Let's look at Claude Skills again. Since the context can be offloaded to the file system, this greatly optimizes the interaction between the application layer and the model layer in terms of consistency, performance, and cost.
- There are also obvious limitations in prompt words: the core of MCP is the definition of MCP Server, as follows: After Andrej Karpathy proposed context engineering, and we compare the core definition of MCP, we will find that the limitations of MCP are actually: Using the idea of prompt word engineering to connect applications and large models. This statement was true last year, but in 2025, it is obviously not true. 📷Peak Ji, the chief scientist of Manus, came to a conclusion after their extensive engineering practice: At present, contextual engineering is the clearest and most practical boundary between applications and models.
- Decreased appeal: After the emergence of capabilities like ClaudeCode and Agent SDK, there are not many MCPs that are really worth using, such as Context7, Playwright, Browser MCP, these MCPs that assist AI coding, and some common middleware official MCPs such as maps and databases. MCP is becoming less and less attractive to individual developers. Application perspective Especially in 2025, when AI Agent explodes, more and more Agent products appear, and major basic model manufacturers have begun to pay attention to Agentic capabilities, especially coding capabilities. There are actually very few AI agents or tools that are based on MCP as product selling points. And most of them serve various vertical fields of AI Coding. The threshold for developing an AI Agent is getting lower and lower. It is more attractive to use models to connect traditional applications than to directly use AI to recreate the tradition. Commercialization can also tell a story. Short-term integration, long-term replacement? So, will Claude Skill replace MCP? My answer is not in the short term and will take the route of integration. As the launcher of MCP, Claude is definitely not willing to slap himself in the face. So, let’s take a look at the official Claude Skill architecture: 📷Claude officials hope that MCP will be responsible for the connection with external applications, and Skills will be responsible for definable internal applications (or components) with independent operating environments. MCP, Skills and Claude Code together constitute Claude's Agent ecosystem. However, Skills can be shared, so it is foreseeable that there will be a gradually prosperous Skill Market, just like the MCP Market. Claude has officially given example Skills for PPT, Excel, Word, and PDF. In other words, there are these types of MCPs on the market now. Since Skill is more stable and reliable, why not use it? Another potential problem is that if Skill is not standard, the application layer is at risk of being bound by the Claude model, which is really expensive after all. However, if Skill is a good idea, I believe that other major prototype manufacturers will launch their own Skill-like capabilities in the next few months. In the medium to long term, the value of MCP will become smaller. With emotion, the context project went from being proposed to being concrete within one year, and the ecosystem was quickly established. This progress is really fast.