Claude Skills, Prompts, Projects, Subagents, and MCP: A Complete Guide to Differences and Selection¶
To master Claude Code, understanding the differences between the five mechanisms (Skills, Prompts, Projects, Subagents, MCP) is essential. This guide explains each role and implementation steps for engineers tasked with implementing Claude in their organizations.
Target Audience
- Individual developers and tech leads starting with Claude Code
- Engineers and SREs tasked with Claude implementation in organizations
- Those using ...
Summary in One Sentence Each¶
| Mechanism | One-Sentence Description |
|---|---|
| Prompts | "Verbal instructions" given on the spot |
| Skills | "Standard procedures" used repeatedly |
| Projects | "Project folders" with documentation and history |
| Subagents | "Specialist team members" with roles and permissions |
| MCP | "Wiring" to connect internal systems and SaaS |
Decision Criteria When in Doubt¶
- One-time consultation → Prompts
- Explaining the same procedure every time → Skills
- Ongoing discussion about specific product/project → Projects
- Want to separate permissions by role → Subagents
- Need direct access to GitHub or databases → MCP
What is Skills: Procedural Manuals and Knowledge Folders¶
A mechanism for organizing repeatable work procedures and specialized knowledge in reusable form. It packages instructions, metadata, scripts and other resources, loading them progressively only when needed, enabling context conservation (progressive disclosure).
Use cases: Security check procedures, deployment steps, coding standard verification
First Steps¶
- Create a
.claude/skillsdirectory in your Claude Code repository root - Create a file like
SECURITY_REVIEW.skill.mdand template "when, what code, how to check" - Verify the log on the Claude Code side until the Skill is actually invoked
Common Mistakes
Creating too many Skills becomes unmanageable. Start with 3-5 and only promote procedures you actually repeat
What is Prompts: One-Time Instructions¶
One-time instructions given directly in conversation. Execute immediately with flexible instructions, but aren't saved.
Use cases: "Rewrite this code for Python 3.12 compatibility", "Create 3 test cases for this function"
First Steps¶
- Start with Prompts only, and promote to Skills when you notice repeating the same explanation
Key Point
Consider promoting to Skills or Projects when repeating the same task 3+ times
What is Projects: Workspace with Knowledge Base¶
A workspace that maintains project-specific knowledge and context. Preserves conversation history and context for referencing past decisions in long-term development. Features 200K context + RAG integration to extend project knowledge.
Use cases: API specification management, design document storage, implementation discussion history
First Steps¶
- Create a Project from Claude's left menu
- Upload only 3-5 specifications/README/design documents used in that project
- Add just one instruction like "This project prioritizes ◯◯ perspective" and test behavior
Common Mistakes
Throwing everything into Projects reduces searchability. Focus only on highly relevant materials
What is Subagents: Mechanism for Task Division as Specialist Teams¶
A mechanism for combining multiple AI agents specialized in specific domains. Each has independent context and tool permissions, functioning not as mere "role labels" but as units of permission design.
Use cases: Frontend/backend/testing division of labor, parallel development of complex features
First Steps¶
- When using Claude Code, create one Subagent just for test execution or static analysis
- Set limited permissions to prevent unexpected behavior
- Review operation logs and adjust permission design
Common Mistakes
Creating too many granular Subagents makes behavior unpredictable. For simple tasks, a single Prompt is more efficient
What is MCP: Connection Layer to External Systems¶
A connection layer for integrating with external databases and APIs. Enables integration with external data and systems Claude alone cannot access. As an open standard, it can be used from clients other than Claude and consolidates integration with multiple tools into a single protocol.
Use cases: GitHub API integration, database access, cloud service integration
First Steps¶
- Install just one official MCP server (e.g.,
@modelcontextprotocol/server-filesystem) - Add server information to Claude Code configuration file
- Start with simple file operations
Common Mistakes
Connecting everything to MCP causes governance breakdown. Start with read-only access
Comparing the Five Mechanisms¶
What They Preserve / How Long They Persist¶
| Mechanism | What It Preserves | How Long It Persists |
|---|---|---|
| Prompts | One conversation turn of context | Within one chat only |
| Skills | Procedures, know-how, scripts | Reusable across all projects |
| Projects | Documents, history, settings | Continuous per project |
| Subagents | Role prompts + permissions | Continuous per agent |
| MCP | Connection definitions | Continuous between server/client |
Selection by Scenario¶
| Scenario | Recommended Mechanism |
|---|---|
| One-time question | Prompts |
| Repetitive task | Skills |
| Long-term project | Projects |
| Complex task division | Subagents |
| External system integration | MCP |
These mechanisms can be combined. You can incorporate Skills into Projects, fetch external data with MCP, and execute parallel work with Subagents.
Where to Start: Implementation Steps¶
Step 1: Start with Prompts only Use chat + Projects to identify repetitive tasks
Step 2: Promote repetitive tasks to Skills Identify long prompts you type repeatedly and promote them to Skills
Step 3: Separate long-term/team projects with Projects Consolidate specifications, README, design documents to maintain project-specific context
Step 4: Implement permission separation/parallelization with Subagents Create one Subagent just for test execution or static analysis to learn permission design
Step 5: Connect internal systems/external SaaS with MCP Connect existing GitHub/DB/SaaS via MCP and integrate into "production workflow"
Summary¶
By understanding when to use each of the five mechanisms, you can efficiently leverage Claude Code. Start with simple Prompts, promote repetitive tasks to Skills, and utilize Projects and MCP as your project grows.
Reference Links¶
Official Documentation¶
- Agent Skills - Anthropic Docs
- Projects - Anthropic Support
- Model Context Protocol (MCP)
- Skills explained - Anthropic Blog
Related Articles¶
- Claude Skills Complete Guide - Detailed Skills implementation guide
- Claude Code Complete Guide - Complete Claude Code overview
- Claude MCP Implementation Guide - MCP server implementation methods