Do Claude Skills Work With GPT and Gemini?

The Open Standard Promise: Built Once, Run Anywhere?

Imagine writing an onboarding guide once and handing it to every new hire at every company in the world. That's the pitch behind Agent Skills — and on December 18, 2025, Anthropic made it official by publishing the format as an open standard for cross-platform portability.

Anthropic describes skills as "organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks" — like an onboarding guide for a new hire. The official standard's site goes further, promising you can "build a skill once and use it across any skills-compatible agent," and notes the format has been adopted by a growing number of agent products.

The momentum is real. The anthropics/skills repository passed 62,000 GitHub stars within four months of launch, and an academic survey notes that structurally identical architectures have appeared at other frontier model providers. So the format is genuinely open, and the ecosystem is genuinely exploding.

But here's the honest answer to this article's question: no source confirms that GPT or Gemini have adopted the standard. Neither Anthropic's announcement nor the agentskills.io site names OpenAI or Google as adopters, and the survey lists cross-platform skill portability as one of the field's open challenges rather than a solved problem.

Even where a skill does run, results aren't guaranteed to transfer. Anthropic's own best-practices documentation warns that "effectiveness depends on the underlying model" and advises testing every skill on all models you plan to use it with — even within Anthropic's lineup, what works for Opus may need more detail for Haiku.

So what should you actually take away?

• The SKILL.md format is open by design — anyone can implement it.
• Adoption by GPT and Gemini specifically is unconfirmed in available sources.
• Portability is a working goal, not a verified reality.

That gap between "open" and "proven everywhere" is exactly why testing matters. SkillzBench is being built to test every skill in a sandbox with the same model, with and without the skill, and publish the resulting Lift — so you know what a skill actually does on your stack, not just what its README promises.

Why Compatibility Is Harder Than the Format

A skill file that loads cleanly on one model can still fail to help on another — and that gap is where most cross-platform optimism breaks down. Anthropic's own documentation is blunt about this: skills "act as additions to models, so effectiveness depends on the underlying model," and the company explicitly recommends testing a skill with every model you plan to use it with.

The dependency runs even within a single provider's lineup. Anthropic notes that what "works perfectly for Opus might need more detail for Haiku," meaning a skill tuned for a frontier model may underperform on a smaller, faster one. If that variance exists between two Claude models, imagine the gap between Claude, GPT, and Gemini.

There's also an adoption question hiding beneath the format question. A recent academic survey observes that "structurally identical architectures were independently adopted by other frontier model providers" — but it names none of them. The same survey lists cross-platform skill portability as an open challenge, which suggests the industry itself doesn't consider this solved.

Here's what model dependency means in practice:

• A skill's instructions are interpreted, not executed — weaker models may miss nuance or skip steps entirely.
• Porting a skill to a new model requires re-testing, not just re-installing.
• Format compatibility (the file loads) and behavioral compatibility (the skill helps) are separate questions.

And "the skill helps" is not a given on any model. The SkillsBench benchmark of 86 tasks across 11 domains found that curated skills raise average pass rates by 16.2 points — yet 16 of 84 tasks showed a negative delta. Roughly one in five measured tasks got worse with a skill attached, despite the strong average.

That's why aggregate claims about skill performance deserve skepticism at the individual level. A 16-point average gain tells you skills help in general; it tells you nothing about the specific skill you're about to load onto a specific model. This is precisely the problem SkillzBench is being built to address: benchmarking every skill in a sandbox with the same model, with and without the skill, so the Lift number reflects actual measured performance rather than marketing claims.

The takeaway is straightforward. Even if GPT and Gemini fully adopt the skills format tomorrow, per-skill, per-model verification would still be the missing step between "it loads" and "it works."

The Risk You Inherit When You Install a Skill

Installing a skill feels harmless — a few lines of instructions, a folder of scripts — but that folder runs with your agent's full authority. Before you copy one from a community repository, it's worth knowing what the security research actually found.

The numbers are sobering. According to an academic survey of the skills ecosystem, 26.1% of community-contributed skills contain at least one vulnerability. Snyk's audit went deeper: after scanning 3,984 skills, researchers found that 13.4% carry a critical-level security issue and 76 confirmed malicious payloads are live in the wild. As Snyk put it bluntly: if you installed a skill in the past month, there's a 13% chance it contains a critical flaw and "a non-zero chance it's actively exfiltrating your credentials right now."

What makes this worse than the npm problem it resembles is access. Snyk notes the skills ecosystem mirrors the early days of npm and PyPI — except with unprecedented access to credentials, file systems, and APIs. A malicious package on npm steals code. A malicious skill directs an agent that holds your logins, your documents, and your payment tools.

The attack patterns are consistent and worth knowing:

• Prompt injection: 91% of confirmed malicious skills use it, layering hidden instructions on top of malicious code (100% of confirmed malicious skills contained malicious code patterns).
• Hardcoded secrets: 10.9% of ClawHub skills embed credentials directly in the package.
• Untrusted fetches: 17.7% of skills pull third-party content at runtime, opening a channel for payload swaps after review.

The scale amplifies the risk. The anthropics/skills repository passed 62,000 GitHub stars within four months of launch, and daily submissions to community hubs jumped from under 50 to over 500 in a matter of weeks — a 10x surge that no human review process can keep pace with. Anthropic itself advises installing skills only from trusted sources and auditing anything less trusted, warning that malicious skills can introduce vulnerabilities or direct data exfiltration.

OWASP has responded by launching the first security framework for AI agent skills, covering all major agent platforms. That's a strong signal this is a permanent category of risk, not a launch-phase hiccup.

This is exactly the problem SkillzBench is being built to solve: an independent directory that pins every skill version to a commit SHA, security-scans each one, and re-scans when upstream code changes — all free, with trust scores that money never touches. Until tooling like that is widely available, treat every community skill the way you'd treat an unvetted npm package with root access: verify before you install.

How to Test Skills Across Agents Safely

Testing a skill across agents isn't optional diligence — it's the difference between a skill that helps and one that quietly hurts. Anthropic's own documentation is blunt on this point: skills act as additions to models, so their effectiveness depends on the underlying model, and you should test your skill with all the models you plan to use it with. Even within one provider's lineup, what works perfectly for a frontier model may need more detail for a smaller one.

That guidance matters more than ever as the open standard spreads. The format was published as an open standard for cross-platform portability, but actual compatibility varies by agent — so treat every model switch as a fresh test, not a given.

Before you install anything, keep three practical rules in mind:

• Test on every target model. A skill that boosts one agent can degrade another; benchmarking found 16 of 84 tasks showed a negative delta even as curated skills raised average pass rates by 16.2 points overall.
• Keep SKILL.md files under 500 lines. Anthropic's best practices recommend this length for optimal performance — bloated instruction files waste context and dilute behavior.
• Install only from trusted sources. Anthropic explicitly warns that malicious skills could introduce vulnerabilities or direct data exfiltration — and the risk is real, with 13.4% of scanned skills containing at least one critical-level security issue.

That last rule deserves emphasis. The skills ecosystem has a supply chain problem that mirrors the early days of npm and PyPI, except agents have unprecedented access to credentials, file systems, and APIs. Roughly a quarter of community skills carry at least one vulnerability, so verification — commit-SHA pinning, security scanning, license checks — should precede any install, not follow it.

Finally, pair skills with MCP rather than treating them as rivals. The emerging agentic stack is elegantly divided: skills supply the "what to do" and MCP supplies the "how to connect". A skill tells your agent how to handle a quarterly report; MCP wires it to the systems where the data lives.

This is the gap SkillzBench is building to close. Its free directory will continuously crawl and verify skills from GitHub, pin every version to a commit SHA, run security scans and license checks, and benchmark each skill in a sandbox — with and without the skill — publishing the resulting Lift in percentage points so you can see whether a skill actually helps your workflow before you ever install it. A free MCP connector will then let you toggle verified skills and tools per agent, so each agent only sees what you've chosen for it. Everything — directory, connector, benchmarks, alerts — will be free, with a trust score from 0 to 100 built on security, quality, license, maintenance, and community signals, and Lift always shown separately so ratings and revenue never touch the numbers.

What to Watch as the Standard Spreads

The honest verdict on Claude Skills compatibility is open by design, adoption-dependent. Anthropic published Agent Skills as an open standard for cross-platform portability on December 18, 2025, and the standard's own site promises you can "build a skill once and use it across any skills-compatible agent." What no source confirms is that OpenAI or Google has implemented it — the academic survey instead lists cross-platform skill portability as one of seven open challenges. The format is open; real compatibility depends on vendors showing up.

A second caveat matters just as much: skills are model-dependent by design. Anthropic's own guidance warns that effectiveness depends on the underlying model, and that a skill tuned for Opus may need rewrites for Haiku. Even inside one vendor's lineup, portability is a testing problem, not a copy-paste problem.

Three signals are worth tracking as the standard spreads:

• OWASP's Agentic Skills Top 10 — the first comprehensive security framework for agent skills, covering all major agent platforms.
• Cross-platform portability itself — whether GPT, Gemini and other frontier providers formally adopt the format or keep building parallel, structurally identical systems.
• Supply-chain risk — 26.1% of community-contributed skills contain vulnerabilities, and a Snyk audit of 3,984 skills found 534 with at least one critical-level issue.
• Whether skills actually help — SkillsBench measured a 16.2-point average pass-rate gain, but 16 of 84 tasks got worse with a skill loaded.

That last point is the one most directories skip. A skill can be safe, well-written and still drag performance down on your model and your task. SkillzBench is building exactly that missing layer: a free, independent directory that pins every skill to a commit SHA, security-scans and license-checks it, then benchmarks it in a sandbox with and without the skill and publishes the Lift in percentage points — always separate from the trust score, because money never touches rankings. Everything stays free, including the MCP connector that lets any agent see only the skills you choose.

Join the SkillzBench waitlist to get verified, benchmarked skills and the free MCP connector the moment they launch.