Claude Skills vs MCP: The Difference, the Confusion, and Why HN Can't Agree
The short version: Skills package workflows — "here's how to do my REST-client routine" — as reusable markdown plus optional code. MCP is a protocol for tools — a standard way for any app to expose actions to any model. They overlap constantly (a skill can call an MCP tool; an MCP server can be wrapped as a skill), and that's exactly why Hacker News spent ten months arguing about it. If you only learn one, learn Skills (easier, higher leverage for personal use). If you're building products for other people, MCP matters more (cross-tool standard).
The one-line difference
| Skills | MCP | |
|---|---|---|
| What it is | Reusable instruction packages (SKILL.md + optional code) loaded into context | A protocol spec for connecting external tools to AI apps |
| Purpose | Reuse your workflow — stop re-explaining how to do things | Standardize how tools talk to models — one connector, many apps |
| Portability | Anthropic-flavored (works in Claude Code) | Cross-vendor (OpenAI, Google, and others adopt it) |
| Mental model | "A trained intern's cheat sheet" | "A USB-C port" |
Or as the community keeps putting it: MCP is what lets you reuse tools; Skills are what let you reuse procedures. A skill often contains MCP instructions; an MCP server can be configured through a skill. Different layers, same stack.
Why HN can't agree (the confusion is real)
The threads on Claude Skills have 127 comments mentioning MCP, split roughly into three camps. The first camp says the two are basically the same thing — and they have receipts:
"Claude Skills just seem to be the same as MCP prompts. I don't really see why they had to create a different concept. Maybe it makes sense 'marketing-wise' for their chat UI, but in Claude Code? Especially since we already had…"
— JyB, HN
The second camp — led by Simon Willison, whose post on the subject has its own HN thread — says Skills are the bigger deal:
"Claude Skills are awesome, maybe a bigger deal than MCP."
— simonw, HN, linking his analysis
The third camp just wants the taxonomy to stop growing — plugins, skills, MCP servers, subagents, marketplaces:
"I also see a burgeoning class of tools emerging: MCP servers, Skill managers, Sub-Agent builders. The patterns and protocols need more explainability — how do they synthesize into a practical dev toolkit?"
There's also a practical skeptic camp pointing out that neither solves the "will the model actually bother?" problem:
"How is it not going to have the same problem of maybe using MCP tools? A skill just sounds like more markdown fed into context, but with…"
What we can confirm from the ground
Strip the noise and the practical consensus is actually stable:
1. Skills are easier to start with. A skill is a folder with a markdown file. No server, no protocol ceremony. The barrier is "write a doc" vs "build a connector". One of the biggest Skill successes on HN was a guy who made Claude build complete Godot games from a skill — that's pure procedure re-use, no new protocol involved.
2. MCP wins for ecosystem plumbing. When vendors expose APIs (Notion, Figma, browsers), MCP gives them one standard instead of N custom integrations. One commenter noted MCP's side effect is real: it pushed vendors to expose APIs they previously kept internal.
3. They compose, and that's the real pattern. The practical stack in 2026 is: MCP servers for capabilities, Skills for workflows built on those capabilities, CLAUDE.md for project memory. The three aren't competitors — they're layers. The people confused by the overlap are usually people who haven't hit the case where all three earn their place.
The honest verdict
Start with Skills if:
You're a Claude Code user who keeps re-explaining workflows. Write one SKILL.md for the thing you repeat most. It'll pay back within a week — no infrastructure required.
Go deep on MCP if:
You're building tools, connectors, or products for other developers. MCP is the cross-app standard; being on it means working with every client that adopts it, not just Claude.
If HN's confusion is your confusion:
Don't pick a side. Use CLAUDE.md as memory, MCP for tools, Skills for procedures. The overlap complaints mostly come from people evaluating the edge cases; the middle is clean and practical.
Method: 913 comments read across the three largest Claude Skills HN threads (816p / 738p / 337p, Oct 2025 – Aug 2026); 127 comments specifically discuss MCP. Quotes verbatim with links. Repo stars live via GitHub API 2026-08-27. We don't take payments from vendors — this site survives on being useful.