The same skill in Claude, Cursor & ChatGPT

One of the quieter advantages of a well-built skill is portability: the same markdown file works in Claude, Cursor, Windsurf, GitHub Copilot, and ChatGPT, with no separate "versions" to maintain. You buy a skill once and run it wherever you already work. But "works everywhere" is only true if you load it the right way in each tool — and each tool has exactly one gotcha worth knowing.

Here is how the same skill drops into each of the major tools, and the single thing to watch for in each.

Portability comes from the format, not from adapters

The reason a skill is portable is that it is plain markdown — a structured document, not code tied to any one platform. There is no Claude version and a separate Cursor version; there is one file, and each tool simply needs it placed where it reads its instructions. That is the whole trick, and it is why "works in every tool" is a real claim rather than marketing.

The tools split into two groups: those that read a file from your project on disk, and those that load instructions through their interface. The mechanics differ slightly, but the skill itself never changes.

Claude Code, Cursor, Windsurf, Copilot, Codex

Claude Code

Claude Code reads a CLAUDE.md file at your project root and imports referenced files automatically. Keep your skills in a folder like .claude/skills/ and reference them with an import line in CLAUDE.md. The gotcha: CLAUDE.md applies to every session in that project, so keep one project per skill or workflow to stop skills bleeding into each other.

Cursor

Cursor loads rules from a .cursor/rules/ folder. The gotcha: the file has to end in .mdc and live in that exact folder, or Cursor will not load it. Rename the skill's extension when you save it there.

Windsurf

Windsurf's Cascade reads workspace rules from .windsurf/rules/. The gotcha: rules files have a size limit, so paste one skill at a time rather than dumping a whole bundle into a single file.

GitHub Copilot

Copilot reads .github/copilot-instructions.md for each repository, automatically. The gotcha: there is only one file, so swap or append its contents when you change which skill you are using.

OpenAI Codex

Codex reads an AGENTS.md file at the project root. The gotcha: like Claude Code, it is project-scoped — paste a skill in directly, or keep skills in a folder and point to them from AGENTS.md.

Claude app and ChatGPT

The Claude desktop and web app, and ChatGPT, do not read files from disk. They load a skill through their interface instead, and the choice is the same in both: a persistent home for reuse, or a paste for one-off use.

• For repeat use: in the Claude app, add the skill to a Project's knowledge; in ChatGPT, paste it into a Custom GPT's instructions. Either way it persists and gates for inputs every time you open it.
• For a one-off run: paste the skill's markdown as your first message, then add "use the skill above; ask me for any inputs you need first." The gotcha: a skill pasted into a plain chat only lasts that conversation. For anything you will run more than once, set up the Project or Custom GPT.

The behavior is identical everywhere

Wherever you load it, a well-built skill behaves the same way: it asks for the inputs it needs before it runs, follows its defined method, and returns structured output. The tool is just the room the skill works in. That consistency is the point — you can switch tools, or use different tools for different clients, without relearning anything or maintaining parallel copies.

The takeaway

Portability is not an accident; it is a property of building skills as plain, structured markdown. The same file runs in every major AI tool — you just place it where that tool reads its instructions, and mind the one gotcha each has. Buy the skill once, run it anywhere, and never maintain five versions of the same thing.