Claude Code vs. Gemini CLI: Agent Skill Portability in 2026

The comparison between Claude Code and Gemini CLI has fundamentally changed. For a long time, the discussion centered on context window size and which agent had the better native capabilities. But two recent shifts have made that debate obsolete. First, both agents now support the same open standard for skills. Second, Gemini CLI is no longer a consumer-facing tool for individual developers.

That reframes the whole Claude Code vs Gemini CLI question. The focus is no longer the proprietary features of each harness, but the portability of the skills themselves. What does it take for a skill to work reliably on any platform that supports the standard? And what does our test data—from running 2116 skills in a Claude Code environment—tell us about why they break?

The Shared Foundation: The Agent Skills Standard

Before comparing the harnesses, it’s critical to understand what they have in common. The query “do skills work in gemini cli” has a simple, direct answer: yes. This is because both Claude Code and Gemini CLI, along with a growing list of other clients like Codex, Cursor, and OpenCode, have adopted the open Agent Skills format.

Originally developed by Anthropic and now maintained as an open standard at agentskills.io, this format defines a skill as a self-contained directory. At its core is a SKILL.md file containing YAML frontmatter for metadata (like name and description) and a markdown body with instructions for the agent. This directory can optionally bundle its own resources, such as scripts in a scripts/ folder or reference documents in references/.

This shared standard means that, in principle, a skill written for Claude Code should be readable by Gemini CLI, and vice versa. The skill itself is portable. The divergence happens in how each agent harness discovers, loads, and executes it.

Harness Mechanics: How Claude and Gemini Run Skills

While the skill format is the same, the operational mechanics of Claude Code and Gemini CLI differ in key ways. These differences affect how skills are managed, how they’re loaded into context, and how they interact with the agent’s security model.

Claude Code Skill Execution

Claude Code looks for skills in two primary locations: a global directory at ~/.claude/skills/<name>/SKILL.md for personal skills, and a project-specific directory at .claude/skills/<name>/SKILL.md. It also discovers skills in nested .claude/skills/ directories as it traverses a project.

When a task triggers a skill, Claude Code renders the entire SKILL.md file into the conversation as a single message. This context persists for the rest of the session; the file is not re-read on subsequent turns. This is a simple, file-based approach. The frontmatter in SKILL.md also allows for some control over the agent’s toolset during the skill’s execution. The allowed-tools key can pre-approve specific tools for the turn that invokes the skill, while disallowed-tools can temporarily remove tools from the agent’s reach while the skill is active.

Gemini CLI Skill Execution

Gemini CLI uses a more structured, multi-tiered discovery system and a two-stage loading process. It searches for skills in a specific order of precedence: built-in skills, extension skills, user skills (~/.gemini/skills/ or the ~/.agents/skills/ alias), and finally workspace skills (.gemini/skills/ or its .agents/skills/ alias).

Unlike Claude Code, Gemini CLI does not immediately load the entire skill. At the start of a session, it injects only the name and description of each enabled skill into the system prompt. This keeps the initial context lean. Only when the model determines a task matches a skill does it call the activate_skill tool. This triggers a consent prompt for the user, detailing the skill’s name, its purpose, and the directory path it will be granted access to. Upon user approval, the full SKILL.md body and its folder structure are added to the conversation history, and the skill’s directory is added to the agent’s list of allowed file paths.

This system is managed via a dedicated set of commands, like gemini skills install <git-url-or-dir> --consent and /skills link <path> --scope user|workspace.

Here is a summary of the key differences:

Feature	Claude Code	Gemini CLI
Skill Path	~/.claude/skills/, .claude/skills/	~/.gemini/skills/, .gemini/skills/ (plus .agents/skills/ alias)
Invocation	Full SKILL.md injected on first use, persists	Name/Description in system prompt; SKILL.md injected on activation
User Consent	Implicit on install/placement	Explicit prompt on first activation per session
Management	File system operations	CLI commands (gemini skills ..., /skills ...)

The Real Story for 2026: Gemini CLI Stopped Being a Consumer Tool

The technical comparison is now largely academic for most developers. On May 19, 2026, Google announced it is transitioning Gemini CLI to a new tool called Antigravity CLI. As of June 18, 2026, Gemini CLI and the associated Gemini Code Assist IDE extensions stopped serving requests for individual Google AI Pro and Ultra subscribers, as well as the free tier (which was formerly 60 requests/minute and 1,000/day).

Service continues for enterprise customers with Gemini Code Assist Standard and Enterprise licenses and for those using the Gemini API via direct billing. The gemini-cli repository remains public under an Apache-2.0 license and is still in active development.

Crucially, Google has stated that the core agent technologies—Agent Skills, Hooks, Subagents, and Extensions—will carry over to Antigravity CLI, which is described as a rewrite in Go for performance. The announcement did not address the license for Antigravity CLI. That shift changes what choosing between Claude Code or Gemini CLI in 2026 even means. For an individual developer without an enterprise licence or direct API billing, it is no longer a choice between two consumer tools. The practical concern is whether the investment in building a personal library of skills will be portable to future agent harnesses.

Context Window vs. Skill Depth

Much has been made of Gemini’s advertised 1M token context window. While large-context configurations exist on both sides, fixating on the token count misses the point. Context size and skill depth solve different classes of problems.

A massive context window is powerful for tasks that require a holistic understanding of a large codebase. It allows the agent to hold more of your repository in its immediate attention, which can be invaluable for complex refactoring or analysis. However, it does not solve for action.

A 1M token window cannot install a missing binary. It cannot provision an API key. It cannot fix a SKILL.md file that is written too vaguely to trigger reliably. These are problems of skill depth—the ability of an agent to execute a precise, repeatable, environment-aware task. That is the domain of a well-written skill, and its success often depends less on the size of the context window and more on the environment in which it runs.

What Our Data Shows About Skill Failure (On Any Platform)

This brings us to the core of the issue: why do skills fail? At SkillProof, we install and run skills to publish an honest verdict. Of the 2116 skills we have tested, 1307 (62%) pass, 704 are marked “Works with setup” — they do the job, but only after configuration the listing has to spell out — and 105 land on a fails verdict: some ran and left you worse off than no skill at all, others could not run at all.

To be perfectly clear: we run these tests exclusively in a Claude Code harness. We have never run a skill on Gemini CLI or Antigravity CLI, and we make no claims about their pass rates. Our full process is detailed in our methodology.

However, our data on why skills fail is still highly relevant to any harness. We scanned the test notes for every skill that didn’t pass cleanly. The results are telling.

Of the 704 skills requiring setup:

• 195 mentioned a missing CLI or binary.
• 99 depended on a separate MCP server being configured and running.
• 92 required an API key or other credential.
• 64 needed a paid account or subscription.

Of the 105 skills that landed on a fails verdict:

• 27 involved a missing CLI or binary.
• 12 depended on an MCP server.
• 10 required a paid account.
• 7 needed an API key.

These are counts from a keyword scan of our own test notes, not an exhaustive taxonomy of causes, and the categories overlap: a single skill might need both a CLI and a key. But the pattern is hard to miss. The overwhelming reason a skill underdelivers is a dependency on the external environment. A skill that requires a binary you haven’t installed will fail whether you run it via Claude Code or Gemini CLI. A skill that needs an API key will not work without one, regardless of the agent’s context window.

Case Studies: Where Skills Pass and Fail

Looking at specific examples makes this clear. The difference between a passing skill and one that needs setup often has nothing to do with the agent harness.

• Sync Agent Instructions: This coding skill passed our tests. Its purpose is to synchronize instructions across CLAUDE.md, GEMINI.md, and AGENTS.md files. It works because it has no external dependencies; it only uses basic shell commands like cp, git diff, and git add. Its logic is entirely portable. The value it provided over a naive baseline was its precision—it correctly staged the destination files via git add while leaving the source file unstaged.

• AI Multimodal: This data skill landed on “Works with setup.” Its scripts are clean and use the official google-genai Python SDK. However, it requires that SDK to be installed and, more importantly, a live GEMINI_API_KEY in the environment. We could not provision a key for the test, so its output was never measured. This is a Gemini-powered skill that fails for a reason that would block it just as effectively inside Gemini CLI itself.

• Amazon Alexa QA: Another “Works with setup” example, this seo skill illustrates the hidden dependency problem. Its instructions never mention that it silently requires a separate browser-act-cli tool, which must be installed via uv tool install browser-act-cli. It also needs a live, logged-in Amazon account. This failure pattern is entirely independent of the agent.

• AI Search Hub: This passing data skill is an example of a tool that just works. We ran its documented command, and it returned a correct answer on the first try. The baseline—the same model without the skill—took three attempts and produced zero results.

Comparing Gemini CLI vs Claude Code for coding, then, is less a question about the two tools than about the skills you run on them. A shared standard is a real step forward, but as our data shows, portability comes down to whether a skill is honest about what it needs from the machine underneath it.

Related reading: we ran the same portability question against OpenAI’s agent in Claude Code vs Codex CLI, and covered how instruction files differ from skills in AGENTS.md vs Claude Skills.

This is the problem SkillProof was built to solve. We install each skill, run it on real work, and publish the verdict — including the failures. The full catalog of 2116 tested skills is free to browse; if you would rather start from a shortlist, we also sell role-based packs of ten skills for $10 each.