Gemini CLI: Agent Skills Reach Stable

Agent Skills Reach Stable in Gemini CLI v0.26.0

After several weeks of iterative preview development beginning with v0.23.0, Google formally launched Agent Skills as a stable, production-ready feature in Gemini CLI v0.26.0, released January 28, 2026. The milestone marks a significant shift in how developers interact with and extend the CLI, moving from generic prompt engineering toward declarative, reusable expertise bundles.

What Are Agent Skills?

Agent Skills are self-contained directories that package specialized instructions, procedural workflows, and task-specific resources into a discoverable capability unit. Rather than injecting all context upfront, skills operate on a just-in-time basis — the agent activates only the skills relevant to the task at hand. This design keeps the context window leaner, improves response accuracy, and reduces latency and token costs compared to always-on system prompt bloat.

Skills follow a tiered discovery model. Project-scoped skills live in .gemini/skills within the repository and are typically shared with the team via version control. User-scoped skills live in $HOME/.gemini/skills and apply to personal workflows across projects. As of v0.28.0, skills are also discoverable from the .agents/skills directory alias, offering greater flexibility in how teams structure their agent configuration.

Built-In Skills at Launch

The v0.26.0 release ships three built-in skills available out of the box:

• code-reviewer — provides structured guidance for code reviews, including comment style, severity levels, and focus areas
• docs-writer — specializes in technical documentation tasks, enforcing consistent tone and formatting
• skill-creator — a meta-skill that helps developers author new skills by scaffolding the necessary directory structure and instruction files

These built-in skills serve both as practical utilities and as reference implementations for teams building their own custom skills. Prior to v0.26.0, skill management required explicit opt-in; as of v0.26.0, skills are enabled by default for all users.

Security and Consent

With skills now stable and enabled by default, Google introduced explicit security consent prompts for skill installation. Before a skill can affect the agent's behavior, users must authorize it through a confirmation dialog. This guard prevents malicious or misconfigured skills from silently altering the agent loop — a critical consideration for teams operating in enterprise environments or running Gemini CLI in CI/CD pipelines.

Conflict detection and override warnings were also added: if two skills define overlapping instructions, Gemini CLI surfaces a warning rather than silently applying one over the other.

Hooks Reach Stable Simultaneously

The v0.26.0 release also graduated Hooks to stable status, making it the first release where both of Gemini CLI's two primary extensibility systems — Skills and Hooks — are production-ready simultaneously. Hooks allow developers to intercept the agent loop at defined points (BeforeTool, AfterTool, model response, etc.) without modifying Gemini CLI's source code. Together, Skills and Hooks form a composable customization layer that rivals plugin systems in more established tools.

The community response was immediate: Supabase announced a dedicated Gemini CLI extension at the same time as the v0.26.0 release, and third-party skill collections began appearing on GitHub within days.