Claude Skills Explained: Ultimate Guide to Prompts, Projects, Subagents & MCP

高效码农

7 hours ago

★Claude Skills Explained: A Comprehensive Guide to Skills, Prompts, Projects, MCP, and Subagents★

Since the introduction of Skills, there’s been a growing interest in understanding how the various components of Claude’s agentic ecosystem work together. Whether you’re building sophisticated workflows in Claude Code, creating enterprise solutions with the API, or maximizing your productivity on Claude.ai, knowing which tool to reach for—and when—can fundamentally transform how you work with AI.
This guide breaks down each core building block of Claude’s ecosystem, explains when to use each component, and demonstrates how to combine them to create powerful, intelligent workflows that go beyond simple chatbot interactions. We’ll explore Skills, Prompts, Projects, Subagents, and the Model Context Protocol (MCP) in detail, providing you with the knowledge to architect AI solutions that are efficient, scalable, and tailored to your specific needs.

Understanding Your Agentic Building Blocks

Claude’s ecosystem is designed not as a single tool, but as a collection of specialized components. Each has a distinct purpose, much like different instruments in an orchestra. When used correctly, they harmonize to produce results that are far greater than the sum of their parts. Let’s start by examining each component individually.

What Are Skills?

Think of Skills as specialized training manuals or digital toolkits that you can equip Claude with. They are essentially folders containing instructions, scripts, and resources that Claude discovers and loads dynamically when a task requires that specific expertise. Instead of being a generalist, Claude can temporarily become a specialist in Excel spreadsheets, your organization’s brand guidelines, PDF manipulation, or any other domain you define.
How Skills Work: The Magic of Progressive Disclosure
The elegance of the Skills system lies in its efficiency, achieved through a mechanism called progressive disclosure. When you present Claude with a task, it doesn’t load every available Skill at once. Instead, it follows a three-stage process:

Metadata Scan (~100 tokens): Claude first scans the metadata (titles, descriptions, and summaries) of all available Skills. This lightweight scan provides just enough information for Claude to determine if a Skill is potentially relevant to the current task.
Full Instruction Load (<5k tokens): If a Skill’s metadata matches the task, Claude then loads the complete instructions contained within that Skill. This is where the detailed procedures and methodologies are stored.
Asset Loading (As Required): Only if the task specifically calls for it will Claude load any bundled files, scripts, or other resources associated with the Skill. This ensures that heavy assets don’t consume valuable context window space unless they are absolutely necessary.
This architecture means you can have dozens or even hundreds of Skills available without overwhelming Claude’s context window. It accesses exactly what it needs, precisely when it needs it, ensuring optimal performance.
When to Use Skills: The Expertise Multiplier
Skills are the ideal choice when you need Claude to perform specialized tasks consistently and efficiently across different conversations. They are perfect for:

🍄

Organizational Workflows: Standardize procedures like brand guidelines, compliance checklists, document templates, and employee onboarding processes. For example, a “Brand Guidelines” Skill can include your company’s color palette, typography rules, and layout specifications. When Claude creates presentations or documents, it automatically applies these standards without you needing to explain them each time.
🍄

Domain Expertise: Impart deep knowledge in specific technical or analytical areas. This could be a Skill for advanced Excel formulas and macros, another for PDF data extraction and manipulation, or a third for statistical data analysis using Python’s pandas library.
🍄

Personal Preferences: Codify your individual work style. This could include a Skill for your personal note-taking system (e.g., Zettelkasten), your preferred coding patterns and style guides, or your specific research methodologies.
The key differentiator is that Skills are proactive and persistent. Once created, they empower Claude to recognize when their expertise is needed and apply it automatically, saving you from repetitive explanations and ensuring consistent execution.

What Are Prompts?

Prompts are the most fundamental way to interact with Claude. They are the instructions you provide to Claude in natural language during a conversation. Prompts are ephemeral, conversational, and reactive—they exist for the moment, providing immediate context and direction to guide Claude’s response.
When to Use Prompts: The Art of the Moment
Prompts are your go-to tool for a wide range of immediate, one-off interactions. Use them for:

🍄

One-off Requests: Quick, singular tasks like “Summarize this article,” “Translate this paragraph into French,” or “Generate five ideas for a blog post about renewable energy.”
🍄

Conversational Refinement: Iteratively improving an output. For instance, after Claude generates a draft, you might prompt, “Make that tone more professional,” “Shorten this paragraph,” or “Add a more compelling introduction.”
🍄

Immediate Context: Providing the specific information needed for a single task. For example, “Analyze this sales data [paste data] and identify the top three trends.”
🍄

Ad-hoc Instructions: Giving specific formatting or structural requests like “Format this as a bulleted list,” “Create a table from this information,” or “Write this as an email to the marketing team.”
Example: A Comprehensive Security Review Prompt
To illustrate the power of a well-crafted prompt, consider this example for a code security review:

“

Please conduct a comprehensive security review of this code. I’m looking for:

Common vulnerabilities including:

🍄

Injection flaws (SQL, command, XSS, etc.)

🍄

Authentication and authorization issues

🍄

Sensitive data exposure

🍄

Security misconfigurations

🍄

Broken access control

🍄

Cryptographic failures

🍄

Input validation problems

🍄

Error handling and logging issues

For each issue you find, please provide:

🍄

Severity level (Critical/High/Medium/Low)

🍄

Location in the code (line numbers or function names)

🍄

Explanation of why it’s a security risk and how it could be exploited

🍄

Specific fix recommendation with code examples where possible

🍄

Best practice guidance to prevent similar issues

Code context: [Describe what the code does, the language/framework, and the environment it runs in – e.g., “This is a Node.js REST API that handles user authentication and processes payment data”]

Additional considerations:

🍄

Are there any OWASP Top 10 vulnerabilities present?

🍄

Does the code follow security best practices for [specific framework/language]?

🍄

Are there any dependencies with known vulnerabilities?

Please prioritize findings by severity and potential impact.
Pro-Tip: When to Upgrade a Prompt to a Skill
Prompts are your primary way of interacting with Claude, but they don’t persist across conversations. If you find yourself typing or pasting the same complex prompt repeatedly across multiple conversations, it’s time to create a Skill. Transform recurring instructions like “review this code for security vulnerabilities using OWASP standards” or “format this analysis with an executive summary, key findings, and recommendations” into a Skill. This saves you from re-explaining procedures each time and ensures consistent, high-quality execution every time.

What Are Projects?

Available on all paid Claude plans, Projects are self-contained workspaces that provide a persistent, centralized environment for focused initiatives. Each Project comes with its own chat history and a dedicated knowledge base, creating a bubble of context that informs all conversations within it.
How Projects Work: The Persistent Context Hub
Every Project includes a substantial 200K context window. Within this space, you can upload relevant documents, provide background information, and set custom instructions that apply to all conversations within that Project. Everything you upload to a Project’s knowledge base becomes instantly available to Claude across all chats in that workspace.
A standout feature is Claude’s intelligent context management. When your Project’s knowledge base approaches the context window limit, Claude seamlessly enables Retrieval Augmented Generation (RAG) mode. This automatically expands the effective capacity by up to 10x, allowing you to work with extensive document libraries without losing context or performance.
When to Use Projects: The Foundation for Focused Work
Projects are the ideal foundation when you need:

🍄

Persistent Context: Background knowledge that should inform every conversation on a topic. For example, a “Q4 Product Launch” Project containing market research, competitor analysis, and product specifications. Every chat in this project has access to this knowledge without you needing to re-upload or re-explain the context.
🍄

Workspace Organization: Separate contexts for different initiatives. You could have one Project for “Marketing Strategy,” another for “Software Development v2.1,” and a third for “Customer Support Escalations,” each with its own relevant data and instructions.
🍄

Team Collaboration: On Team and Enterprise plans, Projects allow for shared knowledge and conversation history, ensuring everyone on the team is working from the same information base.
🍄

Custom Instructions: Project-specific tone, perspective, or approach. You can instruct Claude to “Always adopt the persona of a senior financial analyst when responding in this Project” or “Focus on actionable recommendations for the European market.”

What Are Subagents?

Subagents are specialized AI assistants that operate as independent workers within the larger Claude ecosystem. Available in Claude Code and the Claude Agent SDK, each subagent has its own context window, a custom system prompt that defines its personality and purpose, and specific, restricted tool permissions.
How Subagents Work: Delegating to Specialists
Each subagent is a self-contained agent with its own configuration. You define:

🍄

What it does: Its core function and area of expertise.
🍄

How it approaches problems: Its system prompt, which acts as its “brain.”
🍄

Which tools it can access: A specific, limited set of tools (e.g., Read, Grep, Bash) to ensure it operates safely within its designated scope.
Claude can automatically delegate tasks to the most appropriate subagent based on their descriptions, or you can explicitly request a specific subagent for a task. They handle their assigned work independently and then return the results to the main agent (you).
When to Use Subagents: The Power of Specialized Execution
Subagents excel in scenarios requiring:
🍄

Task Specialization: Creating dedicated experts for highly specific functions like code review, test generation, or security audits. For example, you could create a code-reviewer subagent with access to Read, Grep, and Glob tools but not Write or Edit. When you modify code, Claude automatically delegates to this subagent for quality and security review without risking unintended code changes.
🍄

Context Management: Keeping the main conversation focused and clean while offloading complex, specialized work to a dedicated worker. This prevents the main context window from being cluttered with the intermediate steps of a sub-task.
🍄

Parallel Processing: Multiple subagents can work on different aspects of a larger task simultaneously, dramatically speeding up complex workflows.
🍄

Tool Restriction: Limiting specific subagents to safe operations. This is crucial for security and stability, as you can give a subagent read-only access to sensitive files or restrict it from executing system commands.

What is MCP (Model Context Protocol)?

The Model Context Protocol (MCP) is an open standard that creates a universal connection layer between AI applications like Claude and your existing tools and data sources. Think of it as a universal adapter that allows Claude to plug into the systems where your data and workflows already live.
How MCP Works: The Universal Connector
MCP provides a standardized way to connect Claude to your tools and data sources. Instead of building custom, one-off integrations for each new system (a Google Drive API, a Slack bot, a database connector), you build against a single, unified protocol.
The architecture is client-server based:

🍄

MCP Servers: These are applications that expose data and capabilities from a specific source (e.g., an MCP server for Google Drive, one for PostgreSQL, one for GitHub).
🍄

MCP Clients: These are AI applications like Claude that connect to these servers to access their functionality.
When to Use MCP: Unlocking Your Data Ecosystem
Choose MCP when you need Claude to interact directly with your existing digital infrastructure:
🍄

Access External Data: Seamlessly search and retrieve information from Google Drive, Slack channels, GitHub repositories, or various databases.
🍄

Use Business Tools: Interact with CRM systems like Salesforce, project management platforms like Jira, or communication tools without leaving the Claude interface.
🍄

Connect to Development Environments: Allow Claude to read and write to local files, interact with your IDE, and access version control systems like Git.
🍄

Integrate with Custom Systems: Build an MCP server for your proprietary internal tools and data sources, bringing them into Claude’s operational sphere.
For example, by connecting Claude to your company’s Google Drive via MCP, it can search documents, read files, and reference internal knowledge without manual uploads. The connection persists and updates automatically, creating a live link to your organization’s knowledge base.

How They Work Together: The Synergy of the Ecosystem

The real power of Claude’s ecosystem emerges when you combine these building blocks. Each serves a distinct purpose, and together they create sophisticated, multi-layered agentic workflows that are greater than the sum of their parts.

Comparison: Choosing the Right Tool for the Job

To make the right choice, it helps to see them side-by-side. This table breaks down the core differences:

Feature	Skills	Prompts	Projects	Subagents	MCP
What it provides	Procedural knowledge & expertise	Moment-to-moment instructions	Background knowledge & context	Task delegation & specialized execution	Tool connectivity & data access
Persistence	Across conversations	Single conversation	Within a project	Across sessions	Continuous connection
Contains	Instructions + code + assets	Natural language	Documents + context	Full agent logic	Tool definitions
When it loads	Dynamically, as needed	Every conversational turn	Always active within a project	When explicitly invoked	Always available in the background
Can include code	Yes	No	No	Yes	Yes
Best for	Specialized, reusable expertise	Quick, one-off requests	Centralized, persistent context	Isolated, specialized tasks	Connecting to external data sources

Example Agentic Workflow: Building a Comprehensive Research Agent

Let’s build a practical example to illustrate how these components combine. We’ll create a comprehensive research agent for competitive analysis that integrates all five building blocks.
Step 1: Set Up Your Project (The Knowledge Base)
First, create a “Competitive Intelligence” Project on Claude.ai. This will be our persistent context hub. Upload the following materials to its knowledge base:

🍄

Industry reports and market analyses
🍄

Competitor product documentation and pricing sheets
🍄

Customer feedback and support tickets exported from your CRM
🍄

Previous research summaries and internal strategy documents
Now, add custom Project Instructions to set the tone and focus:

“

“Analyze competitors through the lens of our product strategy. Focus on identifying differentiation opportunities and emerging market trends. Present all findings with specific evidence and actionable recommendations tailored to our product roadmap.”
Step 2: Connect Data Sources via MCP (The Live Data Feed)
Next, we need to connect Claude to our live data sources. Enable MCP servers for:

🍄

Google Drive: To access shared research documents, financial reports, and marketing presentations that are constantly being updated.
🍄

GitHub: To review competitor open-source repositories, track their technical development, and analyze their community engagement.
🍄

Web Search: To pull in real-time market information, news articles, and press releases as they break.
Step 3: Create Specialized Skills (The Expertise Playbook)
Now, let’s create a Skill that teaches Claude how to conduct the analysis according to our company’s specific methods. Create a “competitive-analysis” Skill with the following content:

# My Company GDrive Navigation Skill
## Overview
Optimized search and retrieval strategy for Meridian Tech's Google Drive structure. Use this skill to efficiently locate internal documents, research, and strategic materials.
## Drive Organization
**Top-level structure:**
- `/Strategy & Planning/` - OKRs, quarterly plans, board decks
- `/Product/` - PRDs, roadmaps, technical specs
- `/Research/` - Market research, competitive intel, user studies
- `/Sales & Marketing/` - Case studies, pitch decks, campaign materials
- `/Customer Success/` - Implementation guides, success metrics
- `/Company Ops/` - Policies, org charts, team directories
**Naming conventions:**
- Format: `YYYY-MM-DD_DocumentName_vX`
- Final versions marked with `_FINAL`
- Drafts include `_DRAFT` or `_WIP`
## Search Best Practices
1. **Start broad, then filter** - Use folder context + keywords
2. **Target document owners** - Sales materials from Sales/, not root
3. **Check recency** - Prioritize documents from last 6 months for current strategy
4. **Look for "source of truth"** - Files with `_FINAL`, `_APPROVED`, or in `/Archives/Official/`
## Research Agent Workflow
1. Identify topic category (product, market, customer)
2. Search relevant folder with targeted keywords
3. Retrieve 3-5 most recent/relevant documents
4. Cross-reference with `/Strategy & Planning/` for context
5. Cite sources with file names and dates

Step 4: Configure Subagents (The Specialist Team)
For this advanced workflow, we’ll use Claude Code or the Agent SDK to configure subagents. This allows for parallel, specialized processing.
Create a market-researcher subagent:

name: market-researcher
description: Research market trends, industry reports, and competitive landscape data. Use proactively for competitive analysis.
tools: Read, Grep, Web-search
---
You are a market research analyst specializing in competitive intelligence.
When researching:
1. Identify authoritative sources (Gartner, Forrester, industry reports)
2. Gather quantitative data (market share, growth rates, funding)
3. Analyze qualitative insights (analyst opinions, customer reviews)
4. Synthesize trends and patterns
Present findings with citations and confidence levels.

Create a technical-analyst subagent:

name: technical-analyst
description: Analyze technical architecture, implementation approaches, and engineering decisions. Use for technical competitive analysis.
tools: Read, Bash, Grep
---
You are a technical architect analyzing competitor technology choices.
When analyzing:
1. Review public repositories and technical documentation
2. Assess architecture patterns and technology stack
3. Evaluate scalability and performance approaches
4. Identify technical strengths and limitations
Focus on actionable technical insights that inform our product decisions.

Step 5: Activate Your Research Agent (Putting It All Together)
Now, you can simply ask Claude a high-level question: “Analyze how our top three competitors are positioning their new AI features and identify gaps we can exploit.”
Here’s the sophisticated orchestration that happens behind the scenes:

Project Context Loads: Claude immediately accesses the uploaded research documents and applies the “Competitive Intelligence” Project instructions, framing the entire analysis through your strategic lens.
MCP Connections Activate: Claude uses the MCP connection to search your Google Drive for the latest competitor briefs, pull data from GitHub repositories, and run real-time web searches for recent announcements.
Skills Engage: The “competitive-analysis” Skill is loaded, providing Claude with the precise methodology and search strategy for your internal documents, ensuring consistency and efficiency.
Subagents Execute (in Claude Code): Claude delegates the task. The market-researcher subagent gathers industry data and analyst reports, while the technical-analyst subagent simultaneously reviews technical implementations and open-source code. They work in parallel.
Prompts Refine: You can provide conversational guidance to steer the analysis: “Focus especially on enterprise customers in the healthcare sector and their data privacy concerns.”
The Result: A comprehensive, multi-faceted competitive analysis that draws from multiple data sources, follows your company’s specific analytical framework, leverages the expertise of specialized subagents, and maintains a persistent context throughout your research project.

Common Questions: Clarifying the Components

How do Skills work in practice?

Skills use progressive disclosure to keep Claude efficient. When working on a task, Claude first scans Skill metadata (descriptions and summaries) to identify relevant matches. This is a very low-cost operation. If a Skill matches the task’s requirements, Claude loads the full instructions. Finally, if the Skill includes executable code or reference files (like a Python script for data analysis), those are loaded only when the task explicitly calls for them. This three-tiered architecture means you can have dozens of Skills available without overwhelming Claude’s context window. Claude accesses exactly what it needs, precisely when it needs it.

Skills vs. Subagents: When should I use what?

This is a common point of confusion. The key distinction is portability versus isolation.

🍄

Use Skills when: You want capabilities that any Claude instance can load and use. Skills are like training materials—they make Claude better at specific tasks across all conversations and contexts. They are portable and reusable.
🍄

Use Subagents when: You need complete, self-contained agents designed for specific purposes that handle workflows independently. Subagents are like specialized employees with their own context, instructions, and critically, their own restricted tool permissions. They provide isolation.
🍄

Use them together when: You want subagents with specialized expertise. For example, a code-review subagent can use Skills for language-specific best practices, combining the independence and isolation of a subagent with the portable expertise of a Skill.

Skills vs. Prompts: What’s the right choice?

The choice here is between persistence and ephemerality.

🍄

Use Prompts when: You’re giving one-time instructions, providing immediate context for a single task, or having a conversational back-and-forth to refine an output. Prompts are reactive and temporary.
🍄

Use Skills when: You have procedures or expertise that you’ll need repeatedly across different conversations. Skills are proactive—Claude knows when to apply them based on the task—and persistent across all your interactions.
🍄

Use them together: Prompts and Skills complement each other perfectly. Use Skills to provide foundational, reusable expertise, then use prompts to provide the specific context and refinement for each individual task.

Skills vs. Projects: How do they differ?

This is about “knowing things” versus “doing things.”

🍄

Use Projects when: You need background knowledge and context that should inform all conversations about a specific initiative. Projects provide a static knowledge base that is always loaded within that workspace. They say, “Here’s what you need to know about this topic.”
🍄

Use Skills when: You need procedural knowledge and executable code that activates only when relevant. Skills provide dynamic expertise that loads on-demand, saving your context window for the task itself. They say, “Here’s how to do these specific things.”
🍄

Use them together when: You want both persistent context and specialized capabilities. For example, a “Product Development” project containing product specs and user research (the “what”), combined with Skills for creating technical documentation and analyzing user feedback data (the “how”).

Can Subagents use Skills?

Yes, absolutely. In Claude Code and the Agent SDK, subagents can access and use Skills just like the main agent. This creates powerful combinations where specialized subagents leverage portable expertise. For example, your python-developer subagent can use the pandas-analysis Skill to perform data transformations following your team’s conventions, while your documentation-writer subagent uses the technical-writing Skill to format API documentation consistently.

Getting Started: Your First Steps

Ready to start building with these powerful components? Here’s how to begin, depending on how you use Claude:
For Claude.ai Users:

🍄

Enable Skills in your Settings under the Features tab.
🍄

Create your first project at claude.ai/projects to establish a persistent knowledge base.
🍄

Try combining project knowledge with a custom Skill for your next analysis task to see the synergy in action.
For API Developers:
🍄

Explore the Skills endpoint in the official Anthropic documentation to learn how to create and manage Skills programmatically.
🍄

Check out the skills cookbook on GitHub (github.com/anthropics/claude-cookbooks/tree/main/skills) for practical code examples and implementation guides.
For Claude Code Users:
🍄

Install Skills directly through the integrated plugin marketplaces within Claude Code.
🍄

Visit the skills cookbook (github.com/anthropics/claude-cookbooks/tree/main/skills) for examples of how to build and integrate Skills into your development workflows.
By understanding the unique strengths of Skills, Prompts, Projects, Subagents, and MCP, you can move beyond simple question-and-answer sessions and begin architecting truly intelligent, automated workflows that leverage the full potential of Claude’s agentic ecosystem.