Usually, yes. A Codex skill is a folder with a SKILL. file, and Claude Code reads that same format from its own skills folder. A skill that is only instructions often runs unchanged. What breaks is the part that names Codex: its commands, its config paths and the instruction file it expects. This guide covers what carries over, how to check a skill before you rely on it, and when the official Codex plugin for Claude Code is the better route.
Can I use a Codex skill in Claude Code without changing it?
If the skill is only instructions, usually yes. Both products build on the same open format. The Agent Skills specification requires a SKILL. file whose YAML frontmatter holds a name and a description, and the name must match the folder name (Agent Skills specification). Claude Code says its skills follow that open standard and extends it with features of its own (Claude Code skills docs). Codex says its skills build on the same standard: a skill is a directory with a SKILL. file, plus optional scripts and references (Codex skills docs).
The standard covers the file, not what the file tells the agent to do. A skill whose steps run codex exec or read ~/. assumes Codex is installed and set up. That part does not move with the folder.
Where do the two agents look for skills?
Personal skills are where most people start, because they apply in every project. The paths differ, but the folder shape is the same:
| Scope | Claude Code | Codex |
|---|---|---|
| Personal, all projects | ~/. | $HOME/. |
| Repository | . in the project, loaded from the starting directory up to the repository root | $CWD/., scanned from the working directory up to the repository root |
A copy from one personal folder to the other is a folder move. The SKILL. inside should read the same afterwards, until you edit it.
What breaks when a skill moves between agents?
Usually three things: commands that name one agent, the name of the instruction file, and the way a skill is called.
A developer who built a converter for turning Claude skills into Codex skills listed the same kinds of difference in one forum post: "The tool references are different, the frontmatter format is stricter, and CLAUDE.md becomes AGENTS.md" (OpenAI Community). That post runs in the opposite direction to yours, but the cause is the same. Each agent names its own tools and reads its own instruction file, so a skill written for one names things the other may not have.
How is a skill called once it is copied?
The calling syntax differs. In the Codex CLI and IDE extension, you invoke a skill explicitly by typing $ or running /skills; Codex can also pick a skill on its own when your request matches its description (Codex skills docs). In Claude Code, you type / followed by the skill name, or Claude loads the skill when it judges it relevant (Claude Code skills docs). Any line that says "type $name" should say /name in Claude Code.
What about scripts and file paths inside the folder?
Many skills ship scripts, references and templates beside the SKILL.. The specification puts optional files in scripts/, references/ and assets/, and it asks that scripts be self-contained or document their dependencies, since supported languages depend on the agent (Agent Skills specification). The paths that break on a copy are the absolute ones: a line that reads /Users/alex/tools/report. points at a machine that is not yours. Rewrite those as paths relative to the skill folder, so the same reference resolves after the move. A script that needs Python packages should say so in the body, because the copy brings the code but not the environment it expects.
Which frontmatter fields are safe to keep?
The specification defines six fields: name, description, license, compatibility, metadata and allowed-tools, the last marked experimental (Agent Skills specification). Claude Code adds fields of its own, such as controls over who can invoke a skill and whether it runs as a subagent. Its docs list which fields belong to the standard and which are Claude Code's (Claude Code skills docs). Keep to the standard fields and the file reads the same in both agents. When a copied skill misbehaves, a field one agent does not know is the first place to look.
Why doesn't a copied skill trigger on its own?
Both agents load skills in stages, and the description is the part they read first. The specification has agents load every skill's name and description at startup, the full SKILL. only when the skill is used, and other files only when a task needs them (Agent Skills specification). Codex works the same way: it starts each skill with its name, description and file path, and it loads the full instructions only when it selects the skill (Codex skills docs).
So the description is what decides whether a skill is ever used. A copied skill with a vague description sits unused, and that looks like a failed copy. Put the words your real tasks use in the description, say when the skill should not fire, and keep the body for the steps.
How do I check a Codex skill before I rely on it in Claude Code?
Run these steps on a copy, not on the only version you have:
- Copy the skill folder into ~/.claude/skills/ so the folder name is the skill name.
- Open SKILL.md and check the name: lowercase letters, digits and single hyphens, and the same as the folder name.
- Search the file for codex, AGENTS.md, $ mentions and any path under ~/.codex, and rewrite each one for the agent you run.
- Read the description and confirm it says when to use the skill, in words your task will match.
- Start a fresh Claude Code session if the skill does not appear, then invoke it with /skill-name on a small, low-stakes task.
- Read every step the agent takes, and every script it runs, before you let it touch real files.
The last step matters most. A skill's steps use the tools of the agent that loads it. Claude Code runs an allowed-tools grant through its normal permission flow (Claude Code skills docs), so read what the skill asks for before you let it touch real files.
Does the Codex plugin for Claude Code replace copying a skill?
For review and delegation, often yes. OpenAI publishes openai/codex-plugin-cc, an Apache 2.0 plugin that brings Codex into Claude Code. It adds /codex:review for a normal read-only review, /codex:adversarial-review for a more skeptical challenge, and /codex:rescue to hand a task to Codex, with /codex:status, /codex:result and /codex:cancel to manage background jobs (openai/codex-plugin-cc). It needs a ChatGPT subscription, including Free, or an OpenAI API key, plus Node.js 18.18 or later. Its runs count against your Codex usage limits.
OpenAI's launch thread explains the design: the plugin delegates through the local Codex CLI and Codex app server, so it uses the same local auth, config and MCP setup you already have with Codex (OpenAI Community). The plugin hands work to Codex. It does not load your skill folder into Claude Code, so if you want a particular Codex skill's behaviour inside Claude Code, you still need the copy.
Third-party bridges exist as well. tomc98/claude-code-codex-skill is an MIT-licensed Claude Code skill that runs the Codex CLI in the background. We have not run it, and it is third-party code, so it gets the same read-before-running check as any other skill.
Where do Tigerless Labs skills fit?
Labs projects that ship as skills are the easy case, because several are built to run in more than one agent. Tigerless Labs is our open-source work on agent tools, and this is how the skill-based ones reach each agent (Tigerless Labs):
design-harness is an Agent Skill for Claude Code, Codex and any SKILL.md-compatible agent. Claude Code installs it from a plugin marketplace, Codex from
codex plugin marketplace add tigerless-labs/design-harness, and other agents take a copy of the skill folder.paper-radar is a SKILL.md skill with no dependencies beyond the standard library, and its install line works in Claude Code, Codex or any agent with a shell.
auto-gtm is a Claude Code and Codex plugin that drafts X and Reddit posts. It only drafts; publishing stays manual.
autoharness is a Claude Code plugin that learns skills from your sessions. Its page lists only the Claude Code install.
Two more Labs projects cover both agents at the data level. agent-memory keeps one Markdown store that Claude Code and Codex CLI can share, and our article on giving Claude Code memory across sessions walks through its setup. cost-xray captures the local API traffic of both agents, so you can see what a skill adds to each one's requests; our earlier article on seeing what Claude Code sends to the API explains how to read that. Neither OpenAI nor Anthropic makes or endorses Tigerless Labs projects.
When is copying a skill the wrong way?
Copying is the wrong move in five cases.
The skill is tied to Codex's own settings. A skill that changes Codex's sandbox mode, names a config. profile or relies on Codex's own skill selection has nothing to attach to in Claude Code, so the behaviour you expect will not appear.
The skill calls tools the other agent does not have. Read its steps for tool names and compare them with what your Claude Code session offers. A step that names a missing tool cannot run, and the skill then does less than its description promises.
The skill depends on a model's habits. Instructions are text, but the agent that reads them is a model, and the same wording can produce different behaviour under a different model or a different default setting. A skill that was tuned for one model family may need a short test run before you trust it with a longer task. Treat the first few runs as a comparison, not a formality.
You want Codex to do the work. For a review or a delegated task, the plugin is the route, not a copied skill.
You cannot tell who wrote it or what it may do. The specification makes license optional, so a skill without one states no terms for reuse. Its scripts run when the agent runs them, so a skill from an unknown author deserves a full read before it is copied.
Conclusion
A Codex skill and a Claude Code skill are the same kind of file in two places, so most instruction-only skills move with a folder copy. The parts that break are the ones that name Codex: its commands, its config paths, its instruction file and its invocation syntax. Those are quick to find, and most are quick to fix.
The decision rule is short: copy the skill if it is only instructions and uses standard fields; use the plugin if you want Codex to do the work; leave a skill alone if it is tied to Codex's own settings.
The rule holds because the two agents share the part that matters most, the folder and the SKILL., and differ in the parts they own: the commands, the config, the instruction file and the model underneath. Move the first, check the second, and test the result on a small task before it goes near a real project.
A free next step: copy one instruction-only skill you already use into ~/., run it once on a small task, and read what it did. If you want a skill that already works in both agents, design-harness installs in both from its plugin marketplace.
FAQ
Can Claude Code read a skill written for Codex?
Yes, if the skill follows the Agent Skills format: a SKILL. with a name and a description, in a folder named after the skill. Claude Code reads those files from ~/.. The contents still have to make sense for Claude Code, so check any Codex-only commands before you rely on the skill.
Where does Claude Code look for personal skills?
In ~/.. Anthropic's skills documentation says personal skills are available across all your projects.
Do I need to change the SKILL.md file to use it in Claude Code?
Not if the skill is only instructions and uses the standard fields. Change it if it names Codex-only commands, config paths or instruction files, or if it uses frontmatter fields the specification does not list.
Is the Codex plugin for Claude Code free?
The plugin is open source under the Apache 2.0 license, but it needs a ChatGPT subscription, including Free, or an OpenAI API key, plus Node.js 18.18 or later. Its runs count against your Codex usage limits.
Is it safe to copy a skill someone else wrote?
Read the SKILL. and any scripts before you run them. A skill's scripts run when the agent runs them, and a missing license field means the skill states no terms for reuse.
