Google Cloud Live: Unlocking Gemini CLI with Skills, Hooks & Plan Mode

Key Concepts

• Gemini CLI Extensibility: Gemini CLI’s power lies in its ability to be extended through Hooks, Skills, and MCP servers, allowing for tailored workflows.
• Skills vs. Extensions: Skills enhance Extensions by providing bundled expertise and custom commands, while Extensions handle installation and sharing.
• Plan Mode: A read-only mode for research and planning before execution, offering a balance between automation and user control.
• Automated Workflows: Gemini CLI can automate tasks like UI testing (with Playwright) and deployment (to Firebase).
• Community & Open Source: A growing community actively contributes to Gemini CLI’s development through skill creation and open-source contributions.

Gemini CLI: Advanced Features & Extending Functionality

The Google Cloud Live session detailed advanced features of the Gemini CLI, a terminal-based agent powered by Gemini models, and explored methods for extending its functionality. The Gemini CLI provides access to local files, tools, and MCP servers, catering to both developers and non-developers.

Hooks: Deterministic Behavior in an Agentic Loop

Hooks are scripts executed at specific points within Gemini CLI’s agentic loop – for example, before a tool call or at session start. They enable deterministic behavior within the often-indeterministic agent loop. Hooks are configured in the settings.json file, specifying a name, type (command), and the command to execute. A demonstration showcased a hook checking for a running development server for the “Memory Wall” application (built with 3JS and React, deployed to Firebase). A simple “hello world” example involved running a linter on code as it’s being written.

Skills: Bundled Expertise & Progressive Disclosure

Skills represent bundled expertise that Gemini CLI can activate on demand. Unlike custom commands (pre-packaged prompts), skills utilize progressive disclosure – loading context only when needed, keeping the overall context cleaner and improving performance. Skills are standardized for portability across different agent platforms. A 3JS expert skill was demonstrated, and a live skill creation process for documentation writing was showcased using the built-in skill creator. Skills are defined using YAML/Markdown front matter and reference files. Jack Weatherspoon described skills as “bundled expertise…like a library book.”

Plan Mode (Preview) & Automation vs. Control

Plan Mode, currently in preview (estimated general availability Tuesday), is a read-only mode where Gemini CLI focuses on research and planning before executing actions. It addresses scenarios where a clear workflow isn't immediately apparent and supports read-only MCP servers. The demo showed Gemini CLI planning a feature for the Memory Wall application (linking notes) and then offering to execute the plan with or without automatically accepting edits. This feature offers a balance between automation and user control.

Extending Gemini CLI: Skills, Extensions & MCP Servers

The session further explored extending Gemini CLI’s functionality, outlining three primary methods: Extensions, Skills, and MCP servers. Skills enhance Extensions, rather than replacing them. Extensions package functionality for easy installation and sharing (e.g., checking code into GitHub), while Skills are bundled within extensions, providing custom commands and context. MCP servers provide a runtime environment for specific tasks.

Practical Applications: UI Testing & Deployment

Gemini CLI can automate UI testing using Playwright, an end-to-end testing framework. The process involves instructing Gemini CLI to test the UI, receiving feedback on failures (e.g., broken links), and iteratively improving the application. Deployment to Firebase, including database setup, web hosting, and authentication, can also be automated through prompting. A live demonstration showcased adding a linking feature to an application, testing it with Playwright, and deploying it to Firebase. The Conductor extension was highlighted as a tool for “spec-driven deployment,” enabling iterative planning and context preservation.

Tool Selection & Community Contributions

The session emphasized that the choice between Gemini CLI and tools like Anti-gravity is largely preference-based, depending on project complexity. Anti-gravity is a more visually-oriented, IDE-based tool suitable for complex tasks and parallel orchestration, while Gemini CLI is minimalistic and command-line focused. The open-source nature of Gemini CLI encourages community contributions, with “hundreds” of external contributors on GitHub. A personalized coaching hook was brainstormed as an example of community-driven customization.

Technical Terms

Key technical terms discussed included: Gemini CLI, Hooks, Skills, MCP (Model-Controlled Process), Agentic Loop, Progressive Disclosure, Playwright, YAML/Markdown Front Matter, Context Window, Auto-Accept Edits, Extensions, Firebase, Anti-Gravity, and Plan Mode.

Conclusion

The Google Cloud Live session demonstrated the powerful extensibility of Gemini CLI through Hooks, Skills, and MCP servers. The introduction of Plan Mode offers a valuable balance between automation and user control, while practical applications like automated UI testing and deployment showcase the tool’s potential for streamlining development workflows. The thriving community and open-source nature of Gemini CLI promise continued innovation and customization, making it a versatile tool for both developers and non-developers. The key takeaway is that Gemini CLI is not just a tool, but a platform for building tailored AI-assisted workflows.