Why Plan Mode Changes the Game
Every developer knows the pain: you start refactoring, and two hours later you've broken three unrelated features. The new plan mode in Gemini CLI addresses this by forcing a research-first, read-only workflow. Instead of jumping straight into code generation, the agent first analyzes your request, explores the codebase, and even asks clarifying questions before proposing a strategy.
This is not just a safety net—it's a fundamental shift in how AI-assisted development should work. By restricting the agent to a subset of tools (read_file, grep_search, glob), plan mode eliminates the risk of accidental file modifications or executions. You can ask it to "research how to migrate this database" or "plan a new feature" without worrying about eager code changes.
For more context on building culturally-aware AI datasets, check out our previous article on Nemotron-Personas-Brazil: The Open Dataset for Building Culturally-Grounded AI.

How Plan Mode Works in Practice
Plan mode is now enabled by default for all users. You can enter it by:
- Typing
/planin the input box - Pressing
Shift+Tabto cycle through approval modes - Simply asking the agent to "start a plan for..."
The ask_user Tool: No More Guessing
One of the standout features is the ask_user tool. Instead of making assumptions about your intent, the agent can pause its research and ask targeted questions. For example:
# Example: Planning a database migration
# User asks:
> Plan a migration from PostgreSQL to MySQL
# Agent response (plan mode):
I need to understand your schema. I have a few questions:
1. Are you using any PostgreSQL-specific data types (e.g., JSONB, ARRAY)?
2. Do you have stored procedures or triggers?
3. What is the expected downtime window?
# Once clarified, the agent produces a step-by-step plan
This bi-directional communication ensures the final plan aligns with your vision before you commit to implementation.

Conductor: Orchestrating Complex Workflows
For larger projects, the Conductor extension takes plan mode to the next level. Conductor enables Context-Driven Development by orchestrating multi-step development tracks. It uses plan mode to perform exhaustive pre-flight checks and research with zero risk, and leverages the ask_user tool to confirm critical decisions at each milestone.
Key Capabilities of Conductor:
- Multi-step migration planning: Breaks down complex migrations into manageable sub-tasks
- Dependency mapping: Understands system dependencies before any code changes
- Milestone-based approvals: You remain in control of the high-level direction
The team is working on bringing Conductor into Gemini CLI as a built-in mode soon.
Limitations and Caveats
- Plan mode is read-only only within the CLI context—it cannot modify files, but external tools (e.g., GitHub, databases) accessed via MCP still require their own permissions.
- The
ask_usertool can slow down workflows if you prefer fully automated planning. - Currently, plan mode may not handle extremely large monorepos efficiently due to tool call limits.

Conclusion: A Safer Path to AI-Assisted Development
Plan mode is a welcome addition for teams that value safety and collaboration. By separating research from execution, it reduces the risk of accidental damage and encourages better planning practices. Whether you're migrating a database, adding a new feature, or refactoring a legacy system, starting with a plan is always a good idea.
If you prefer to keep your terminal lean or have your own planning process, you can disable plan mode via /settings. The choice is yours—Gemini CLI is designed to fit your workflow.
Next Steps
- Try plan mode on your next refactoring task
- Explore the Conductor extension for complex migrations
- For more on high-performance GPU communication, read our article on RCCLX by Meta: Revolutionizing GPU Communication for AMD Platforms