Skills compares to prompts, Projects, MCP, and subagents (www.claude.com)

Anthropic published a practical guide mapping Claude’s agentic building blocks—Skills, prompts, Projects, subagents, and the Model Context Protocol (MCP)—so developers and power users know which tool to use and how to combine them into reliable workflows. The announcement clarifies roles: prompts are ephemeral conversational instructions; Skills are persistent, discoverable folders of procedures, scripts, and assets that load progressively (metadata ~100 tokens, full instructions <5k tokens, files on demand); Projects provide isolated workspaces with a 200K-token context window and automatic RAG expansion up to ~10×; subagents are independent assistants with their own context, system prompts, and tool permissions for parallelized, sandboxed tasks; MCP is an open protocol to connect Claude to external data sources and tools. Projects, Skills, and subagents are available across Claude’s product ecosystem (Claude Code, Claude.ai, and the Agent SDK) with team/enterprise collaboration features. Technically, the guide highlights complementary use patterns: convert recurring, procedural prompts into Skills for portability; use Projects for persistent, project-scoped knowledge and workflows; delegate specialized or risky tasks to subagents with constrained tool access; and wire in live data via MCP so Skills teach “how” to use data while MCP supplies the “where.” For AI/ML teams, this modular stack enables reproducible automation (e.g., standardized security reviews or competitive-intel agents), safer tool access through permissioned subagents, and scalable context management for complex pipelines.