Claude Skills: Simple Files, Big Agent Power

Claude Skills package task know-how as Markdown plus optional scripts that Claude loads only when needed, making them token-efficient and easy to share. They depend on a sandboxed coding environment, turning Claude Code into a general-purpose agent for real computer automation. Compared to MCP, Skills are simpler, lighter, model-agnostic, and likely to drive a rapid proliferation of community-built capabilities.

Key Points

• Skills are lightweight: a Markdown file with YAML metadata and optional scripts that Claude loads only when relevant, keeping token costs low.
• They require a coding environment (filesystem + command execution), which unlocks powerful automation but demands strong sandboxing and security.
• Claude Code functions as a general-purpose agent for computer automation; Skills make this explicit and extensible for real workflows.
• Compared to MCP, Skills avoid protocol complexity and token bloat—LLMs can learn tools at runtime and use simple, shareable instructions.
• Skills are portable and model-agnostic, enabling broad community sharing and a likely surge in practical, domain-specific capabilities.

Sentiment

The overall sentiment is engaged and intellectually serious, but decidedly mixed-to-skeptical rather than uniformly enthusiastic. The community largely accepted that Skills have real pragmatic value — especially the token-efficiency argument against bloated MCP setups — but resisted the framing that they represent a breakthrough or are bigger than MCP. Many experienced users felt they had already discovered this pattern independently, which simultaneously validated the idea and undercut the novelty claim. There was notable irritation at perceived hype cycles in AI tooling. Willison's active participation was viewed both positively and negatively. The discussion was more technically substantive than is typical for AI product announcements.

In Agreement

• Skills solve a real problem: MCP and AGENTS.md frontload too much context, degrading model quality on long sessions; Skills' progressive disclosure is genuinely more efficient
• Naming the pattern has value even if the technique is familiar — it enables community coordination, sharing, and higher-quality conversations about the approach
• Skills are dramatically simpler to create, iterate on, and share than MCP servers, lowering the barrier for non-engineers significantly
• The combination of markdown instructions plus executable scripts in a sandboxed environment is a uniquely powerful paradigm — the model can dynamically discover CLI tool usage via --help rather than requiring pre-specified schemas
• Context pollution is a genuine and serious problem; anything that reduces tokens loaded at session start improves outcomes
• Skills are vendor-neutral by design: other tools like Codex CLI can use skill files even without formal support
• The feedback loop for AI-context documentation is so tight that it's driving better documentation practices than any previous incentive structure
• Skills work well alongside MCP, not against it — a skill can even serve as a lightweight wrapper that loads MCP context only when needed

Opposed

• Skills are not meaningfully new: anyone competent with Claude Code has been doing this for over a year with CLAUDE.md table-of-contents patterns
• The automatic triggering is unreliable in practice — real-world testing shows Claude Code fails to invoke skills even with unambiguous trigger descriptions
• Skill is a misleading name that evokes model adaptation like fine-tuning, when what's actually happening is just instruction injection
• MCP's advantages are durable: it works without a sandboxed Linux environment, with smaller models, across different vendors, and provides authentication and structured API surfaces that CLI scripting cannot replicate
• The core claim that Skills are bigger than MCP is premature, made within hours of the feature's public release and based on limited data
• This is essentially repackaged RAG with a marketing layer; context engineering already described everything Skills does
• Writing documentation for AI rather than for humans is arguably regressive — we should have had good docs all along
• Skills require a full execution environment, making them impractical for consumer-facing or mobile use cases where MCP's server-side model excels
• Early-access relationships with Anthropic create implicit incentives toward positive coverage even without direct payment