Claude Skills

Original Article ↗ Hacker News Discussion ↗

What Skills Are (As Interpreted in the Thread)

  • Many see Skills as “structured system prompts”: bundles of markdown instructions plus optional scripts that get injected into context on demand.
  • Technically, they’re viewed as a more organized version of CLAUDE.md / “folder of .md guides” plus helper scripts that Claude is trained to discover and use.
  • Some compare them to Alexa Skills, but emphasize these are about context/instructions, not third‑party cloud APIs.

Relationship to MCP, Subagents, Tools, Projects

  • Strong sense of overlap and confusion: Skills vs MCP prompts, subagents, plugins, commands, hooks, projects, slash commands.
  • Emergent consensus:
    • Skills = data/instruction bundles loaded selectively into the current agent’s context.
    • Subagents = separate Claude instances with isolated context; can themselves use Skills.
    • MCP = protocol for exposing external APIs/tools; Skills can describe how to use those MCP tools.
  • Some argue Anthropic could have extended subagents or MCP instead of introducing a new primitive.

Perceived Benefits / Good Use Cases

  • Helps with “context bloat” in large repos where one CLAUDE.md becomes huge and noisy.
  • Lets users define narrow, reusable workflows (PDF/XLSX handling, CI workflows, REST calls, refactoring, etc.) and only load them when relevant.
  • Viewed by some as a natural evolution of hand‑rolled “skill folders” they were already using with Claude Code.

Skepticism & Critiques

  • Many say this is “just prompt stuffing with marketing”: prepackaged instructions that could already be done with docs or MCP prompts.
  • Concerns that models still inconsistently follow instructions/workflows, so Skills may not solve reliability issues.
  • Worry about vendor‑specific features increasing lock‑in; advice from some: focus on raw model APIs and model‑agnostic tooling.
  • Some see it as investor‑facing “feature churn” rather than fundamental capability gains.

Complexity, Churn, and Developer Fatigue

  • Strong theme: too many overlapping concepts (skills, agents, tools, MCP, plugins, apps, rules, etc.) echoing frontend or cloud “framework sprawl.”
  • Several suggest ignoring most of it until the ecosystem stabilizes, or only learning concepts “on demand.”
  • Others argue there are stable underlying patterns (agent loop + context management) and Skills are just one more wrapper around that.

Context, Learning, and Model Limits

  • Discussion about Skills exposing the fact that LLMs don’t truly learn or build evolving “skills”; they repeatedly “start from zero” with a description.
  • Some tie this to broader debates: RL vs pure imitation, lack of goals/consequences, and why this may cap progress toward AGI.
  • Counterpoint: context engineering and better orchestration (like Skills) are where practical gains now come from, even as raw model capability plateaus.

Security & Operational Concerns

  • Skills can execute code; users warn about supply‑chain risk (future “skill package managers”), data exfiltration, and need to trust sources.
  • Clarification that Skills run locally in Claude Code but in sandboxed containers in the cloud UI, with limited network access there.