先说结论

If you mainly use Claude for chat and coding, this mistake costs you: you read every new skill repo as better prompt writing, assume models are interchangeable, and miss where reuse, switching cost, and lock-in actually live.

The slower cost is that you keep putting Claude in the wrong role, and your workflow gets messier over time.

You open alirezarezvani / claude-skills expecting prompt craft and instead find install paths, scripts, references, and plugins. That is the tell. A skill is not a prompt. It is an installable package that carries behavior across tools. The most interesting product shifts are often not about raw strength. They are about where the boundary tightens first.

为什么这次值得看

The clearest proof is the repo shape. The README frames the project as 354 production-ready skills, backed by 593 CLI scripts and 711 reference files, with one library spanning 13 AI coding tools [S002]. That is not "better wording." That is packaging and distribution.

The install guide reinforces the same point. Claude Code, Codex, Gemini CLI, and OpenClaw each have their own install path, and there is also an agent-skills-cli route for rolling skills out across multiple agents [S003]. Then the marketplace manifest breaks the repo into dozens of plugins [S008]. Once you see that, the reusable unit becomes obvious: not the sentence, the package.

关键证据

That is also the line people will actually share. It is rarely "the model got stronger." It is "why wasn't the strongest capability shipped as raw text at all?

" Here, the durable value seems to sit one layer lower: in the reusable layer around the prompt, not the prompt alone.

Boundary: this read is based on the public README, INSTALLATION.md, and marketplace manifest on the main branch as of July 6, 2026. It is not a production hit-rate test of individual skills.

If someone on your team still treats skills as prompt writing, share this with them. "13 tools, one skill" is a much better frame for what is being built.

适合谁 / 下一步怎么用

最后落到动作：share