Skills vs. Prompts, Projects, MCP, and Subagents: Choosing the Right Claude Building Block

Understanding Claude’s Building Blocks

What are Skills? Skills are folders containing instructions, scripts, and resources that Claude can discover and load dynamically when a task matches their metadata. They act like specialized training manuals, enabling Claude to handle domain‑specific tasks such as Excel formulas, PDF processing, or brand‑compliant document creation.

How Skills work: When Claude receives a request, it scans available skill metadata (≈100 tokens). If a skill matches, the full instruction (under 5 k tokens) and any bundled files are loaded on demand.

When to use Skills: Use them for repeatable, procedural knowledge that should be consistent across conversations—e.g., brand guidelines, data‑analysis routines, or personal coding styles.

Example: A brand‑guideline skill stores company colors and layout rules; Claude automatically applies them when generating presentations.

What are Prompts?

Prompts are natural‑language commands you give Claude in a conversation. They are temporary, context‑specific, and do not persist beyond the current dialogue.

When to use Prompts: One‑off requests ("Summarize this article"), conversational refinements ("Make the tone more professional"), or ad‑hoc formatting instructions.

Example: "Please format this list as bullet points."

What are Projects?

Projects are self‑contained workspaces available in paid Claude plans. Each project has its own chat history, a 200 K token knowledge base, and can store uploaded documents and custom instructions that apply to every conversation within the project.

When to use Projects: When you need persistent context across multiple chats—e.g., a product‑launch project containing market research, competitor analysis, and specifications.

Example: A "Q4 Product Release" project holds all relevant files so you never need to re‑upload them.

What are Subagents?

Subagents are independent AI assistants with their own context windows, system prompts, and tool permissions. They run in Claude Code or the Agent SDK and can be delegated specific tasks.

When to use Subagents: Specialized tasks such as code review, test generation, or security audits that require isolated tool access.

Example: A "code‑reviewer" subagent granted only Read, Grep, and Glob permissions to safely audit code without risk of modification.

What is MCP (Model Context Protocol)?

MCP is an open standard that connects Claude to external data sources—Google Drive, Slack, GitHub, databases, IDEs, etc.—without building custom integrations for each source.

When to use MCP: Whenever Claude needs live access to external tools or data, such as pulling the latest market reports from Drive or querying a repository on GitHub.

Example: Connecting Claude to a company’s Google Drive enables it to search and cite internal documents automatically.

How They Work Together

Each component excels at a different aspect of an agent workflow. Skills provide procedural knowledge, Prompts supply immediate instructions, Projects supply persistent background context, Subagents execute isolated tasks, and MCP supplies live data connections.

Comparison summary (original table converted to text):

Feature            | Skills          | Prompts       | Projects      | Subagents    | MCP
-------------------|----------------|---------------|--------------|--------------|----------------
Provides           | Procedural know| Instant cmd  | Background   | Task delegation| Tool connectivity
Persistence        | Cross‑dialogue | Single turn   | Within project| Across sessions| Always available
Contains           | Commands+code  | Natural lang  | Docs+context | Full agent logic| Tool definitions
Load timing        | On‑demand      | Every turn    | Always in project| When called | Always available
Can include code? | Yes            | No            | No           | Yes          | Yes
Best suited for   | Specialized skill| Quick request| Centralized context| Specialized task| Data access

End‑to‑End Agent Example: Research Agent

The article walks through building a competitive‑analysis research agent:

Create a "Competitive Intelligence" project and upload industry reports, competitor docs, CRM feedback, and past research.

Enable MCP servers for Google Drive, GitHub, and web search.

Define a "competitive‑analysis" skill that outlines the analysis framework.

Configure two subagents— market-researcher (reads Drive, searches the web) and technical-analyst (reviews code repositories).

Activate the agent by asking Claude to analyze the top three competitors' new AI features.

The workflow proceeds as follows:

Project context loads the uploaded documents.

MCP fetches the latest briefings from Drive and GitHub.

The skill supplies the analysis template.

Subagents execute their specialized tasks in parallel.

Final prompt refinements add focus (e.g., "focus on the healthcare sector").

The result is a comprehensive competitive‑analysis report that combines persistent project knowledge, live data via MCP, procedural guidance from the skill, and specialized work performed by subagents.

FAQ Highlights

How do Skills work?

Skills use progressive loading: Claude first scans metadata, then loads full instructions only if the skill matches the current task, keeping the model’s context window efficient.

Skills vs. Subagents: When to use which?

Use Skills for reusable expertise across any Claude instance. Use Subagents when you need a self‑contained agent with its own tool permissions.

Skills vs. Prompts: When to use which?

Prompts are for one‑off, conversational commands. Skills are for repeatable processes that should persist across dialogues.

Skills vs. Projects: When to use which?

Projects provide static background knowledge for a set of conversations. Skills provide on‑demand procedural actions that load only when relevant.

Can Subagents use Skills?

Yes—subagents can access the same Skills as the main Claude instance, allowing them to combine specialized tool access with reusable expertise.

Getting Started

Claude.ai users: Enable Skills in Settings → Features, create your first project at claude.ai/projects, and try combining project knowledge with Skills in your next analysis.

API developers: Explore the Skills endpoint in the documentation and check out the Skills cookbook.

Claude Code users: Install Skills from the plugin marketplace and refer to the Skills cookbook for examples.