Do Claude Skills Work in Cursor and Codex?

One Skill Format, Five Tools: Why Portability Now Matters

Your team probably runs three AI coding agents before lunch. That's not a workflow problem anymore — it's the new normal, and it changes what you should expect from your skill library.

According to the JetBrains Developer Ecosystem Survey, 90% of professional developers now use AI coding agents at work at least weekly, with 68% using them daily. And the same analysis finds that senior engineers increasingly run multiple tools side-by-side — Cursor or Copilot for daily editing, Claude Code for complex delegated work — matching specific tools to specific workflows rather than crowning a single winner.

That multi-tool reality is exactly why portability matters. If your team builds a library of skills — reusable instruction packages that teach an agent how to handle a task — you don't want to rewrite that library three times.

So what is a skill, in plain terms? A skill is just a SKILL.md file: a folder of markdown instructions, plus optional scripts and references, that an agent loads only when a task matches its description. Anthropic recommends keeping them under 500 lines, and unlike MCP server definitions, they consume context only on match — Codex reads roughly 2% of its context budget for skill metadata at startup, per Firecrawl's technical breakdown.

Here's the good news: the format has outgrown Claude. Multiple independent sources confirm the same SKILL.md file works across Claude Code, Cursor, Codex CLI, Gemini CLI, and GitHub Copilot — one source states it works without modification across all five. OpenAI even ships a migrate-to-codex skill explicitly built to move Claude Code skills into Codex, which is about as clear a vendor endorsement of portability as you'll get.

The caveats come down to two things:

• Install paths differ. Cursor reads skills from `.claude/skills/`, while Codex and Copilot use the open-standard `.agents/skills/` directory, per technical documentation.
• Descriptions are the routing layer. A vague description is the most common reason a skill never triggers at all — and each agent's matcher reads the same text.
• Compatibility isn't safety. Published research found roughly 26% of skills carry at least one vulnerability, and the OWASP Agentic Skills Top 10 exists precisely because skills are now a cross-platform attack surface.

That last point is why SkillzBench is building an independent directory that security-scans and benchmarks skills before you install them anywhere — because a skill that runs in five tools is also a risk that runs in five tools. The short answer to "do Claude skills work in Cursor and Codex?" is yes. The honest answer is: they work everywhere, so vet them like it.

How Compatibility Actually Works: Install Paths, Not Rewrites

A SKILL.md file written for Claude Code can land in Cursor or Codex without rewriting a single line. The format has become an open standard, not a Claude-proprietary one, which means the compatibility question is really a question about where to drop the file.

Two directory conventions now coexist. Cursor reads skills from .claude/skills/, the same path VS Code uses, while Codex and GitHub Copilot read from the open-standard .agents/skills/ directory. Some clients scan both paths for pragmatic compatibility, so a single skill can hit multiple agents if it lives in the right place.

For developers running more than one coding agent, the practical rule is straightforward:

• Author in `.claude/skills/` when the skill is Claude-only and uses Claude-specific features like full `allowed-tools` support.
• Target `.agents/skills/` when the skill needs to run across Claude Code, Cursor, Codex, and Copilot.
• Keep the file content identical across both paths — the SKILL.md format itself does not change.
• Treat the two paths as related, not identical, per practical cross-platform guidance.

OpenAI has shipped its own confirmation that this works. A Codex skill called migrate-to-codex exists explicitly to convert Claude Code skills, agents, and config into Codex-compatible form, which is direct vendor evidence that Claude-authored skills are expected to function in Codex. The Codex skills ecosystem itself launched to 40,285 publicly listed skills within weeks of the SKILL.md spec becoming open, an 18.5× growth spike that signals how quickly developers adopted the shared format.

The routing layer that decides whether any of this works is the description field. Codex reads only skill name and description at startup, roughly 2% of the context budget, then loads full instructions only when the description matches the task. A vague description that would also match a generic Google query is the most common reason a skill fails to trigger at all, and that problem compounds when the same skill is meant to fire across multiple agents.

Compatibility is not the same as safety. Published research cited by the broader community found roughly 26% of agent skills carry at least one vulnerability, and the OWASP Agentic Skills Top 10 exists because skills are now a cross-platform attack surface. A skill that works in Cursor and Codex is also a skill that runs wherever you install it, so verifying what is inside the file matters as much as placing it in the right directory. SkillzBench is building toward exactly that gap, with plans to scan, version-pin, and benchmark skills before they reach an agent.

Browse the SkillzBench skills directory to find skills that have been independently verified, or join the waitlist to get notified when benchmarks and the free MCP connector go live.

Compatible Isn't Safe: The 26% Vulnerability Problem

So your favorite Claude skill runs fine in Cursor and Codex. That's the good news — and it's exactly where the danger starts.

Cross-platform compatibility means a skill's instructions now execute inside multiple agents, each with its own permissions, tool access, and file system reach. Published research found roughly 26% of skills carry at least one vulnerability, and the same risky SKILL.md file that misbehaves in one agent gets three more chances to misbehave when you install it across Claude Code, Cursor, and Codex. Compatibility multiplies convenience; it also multiplies attack surface.

The security community has noticed. The OWASP Agentic Skills Top 10 exists precisely because skills have become a cross-platform attack vector — instruction packages that agents follow with little scrutiny. Security vendors like Semgrep now publish guidance on scanning agentic skills before deployment, treating them like any other third-party code you'd pull into a production repo.

The ecosystem's sheer scale makes manual vetting unrealistic. One industry report counted over 40,000 publicly listed skills within weeks of the SKILL.md spec becoming an open standard — an 18.5× growth spike. Nobody reads that much unvetted instruction text by hand, and skills that load only when a task matches can sit dormant until the wrong moment.

Before trusting any skill across your agents, check four things:

• Security scanning — run the skill through static analysis for prompt injection, data exfiltration patterns, and unsafe tool calls.
• Version pinning — pin every skill to a specific commit SHA so upstream changes can't silently alter behavior.
• License checks — confirm the license permits your intended use, especially in commercial environments.
• Behavioral benchmarking — test the skill in a sandbox with the same model, with and without it, so you know what it actually changes.

This is the gap SkillzBench is being built to close: an independent directory that dedupes skills from GitHub, pins each version to a commit SHA, security-scans and license-checks every listing, and benchmarks skills in a sandbox to measure real Lift. A trust score from security, quality, license, and maintenance signals — with benchmark results always shown separately — so "does it work everywhere" and "is it safe everywhere" get answered independently.

Because in a market where 90% of professional developers use AI coding agents weekly, the question is no longer whether skills run across tools. It's whether you can prove the ones you're running deserve the access you've given them.

Practical Playbook: Making One Skill Work Everywhere

Knowing a skill can run in Cursor and Codex is one thing; making it run reliably across all three is where most teams stumble. The good news is that the path conventions and failure modes are well documented, so portability becomes a set of deliberate design decisions rather than luck.

Target .agents/skills/ as your canonical path. Cursor reads skills from .claude/skills/, while Codex and GitHub Copilot read from the open-standard .agents/skills/ directory, which some clients also scan alongside the Claude path for pragmatic compatibility, according to a technical deep-dive on skill authoring. If a skill is Claude-only and leans on Claude-specific features like full allowed-tools support, keep it in .claude/skills/; everything intended to travel goes in the open location.

Treat the description as a routing layer, not documentation. Codex reads only a skill's name and description at startup — roughly 2% of its context budget — and loads the full instructions only when a task matches, as vendor documentation on Codex skills explains. Practitioners report that bad descriptions are the most common reason skills never trigger at all, and offer a simple test: if your description would match a generic Google query, it is not specific enough.

When moving an existing Claude skill library into Codex, don't rewrite by hand. OpenAI ships a migrate-to-codex skill built specifically to migrate Claude Code skills, agents, and config — a documented migration path that confirms cross-platform intent at the vendor level.

For teams running multiple agents side by side, design for the lowest-common-denominator format. That's not a niche concern: adoption research shows 90% of professional developers use AI coding agents weekly, and senior engineers increasingly run several tools simultaneously, matching each to a specific workflow.

A few decisions to lock in before you author anything:

• Pick the destination first: `.agents/skills/` for portable skills, `.claude/skills/` for Claude-only features like hooks and subagents.
• Write a description specific enough that no other skill or generic query could match it.
• Keep SKILL.md under the recommended ~500 lines and ~5,000 tokens so it loads cleanly everywhere.
• Verify every skill before installing it across agents — published research found roughly 26% of skills carry at least one vulnerability, and compatibility does not equal safety.

That last point deserves emphasis. Cross-platform portability multiplies a skill's reach, which also multiplies its blast radius if it is risky. SkillzBench is being built to close exactly that gap — an independent directory where every skill gets security-scanned, license-checked, and benchmarked in a sandbox, so you can equip agents with skills that are proven, not just portable.

Verify Before You Install: Testing Skills Across Agents

Here's the uncomfortable truth behind every "it works everywhere" claim you've read so far: the evidence is documentation, not demonstration. No independent hands-on benchmark has yet proven that the same SKILL.md file runs identically in Claude Code, Cursor, and Codex. The strongest source in this space cites Anthropic's official Agent Skills quickstart and OpenAI's Codex docs — but citing docs and watching a skill behave the same way in three different agents are two very different things.

That gap matters because compatibility and trust are separate questions. Published research found that roughly 26% of agent skills carry at least one vulnerability, and the OWASP Agentic Skills Top 10 exists precisely because skills are a cross-platform attack surface. A skill that loads in Cursor, Codex, and Claude Code simultaneously loads its risks in all three.

This is exactly the gap SkillzBench is being built to close. Rather than asserting that skills work, the platform is designed to verify it, pillar by pillar:

• Sandbox benchmarking: every skill is tested with the same model, with and without the skill, and the difference is published as a percentage-point Lift — so you can see what a skill actually contributes, not just what its description promises.
• Version pinning: every skill is pinned to a specific commit SHA, so what you install is exactly what was tested — not a silently updated version.
• Security scanning: each skill is validated, license-checked, and scanned, feeding a 0–100 trust score built from security, quality, license, maintenance, and community signals.
• A free MCP connector: one server connects any agent — Claude Code, Cursor, Codex, or others — to only the skills you explicitly choose, with per-agent loadouts so each tool sees just what you selected.

The Lift metric deserves special attention. Vendor blogs in this space routinely rank their own skills first and assert cross-platform compatibility without showing test results. A published percentage-point Lift, measured in a controlled sandbox, is the antidote: it separates what a skill does from how well it's marketed.

For teams running multiple agents — now the norm, with 90% of professional developers using AI coding agents at least weekly — this verification layer is what turns "should work everywhere" into "proven to help everywhere." SkillzBench is free, everything from the directory to the benchmarks to the MCP connector, and it's actively being built right now.

If you want to know the moment verified, benchmarked skills become available for your stack, join the SkillzBench waitlist. Skills, proven — before you install them, not after something breaks.