AI coding tools
Gemini CLI Operating Card
Choose and operate Gemini CLI with an explicit authentication path, tool boundary, sandbox decision, and recovery plan.
- Format
- Reference
- Level
- Beginner
- Audience
- Practitioner, Developer, Operator
- Owner
- Project42 Editorial
- Review cadence
- Every 30 days
- Prerequisites
- A supported local runtime and operating system; A Google-supported authentication path
Use it when a terminal agent fits the task and identity boundary
Gemini CLI is an open-source terminal agent with file, shell, web, memory, MCP, and other tool surfaces. Choose it after confirming that its supported platform, authentication method, model access, and tool set fit the repository and organizational policy.
Authentication paths have different account and cloud-project requirements. Decide the identity path before work begins, then keep credentials outside prompts and repository files. Review mutating tool confirmations and use sandboxing when host isolation is required.
Control context, tools, and recovery
Give the tool a bounded directory and explicit GEMINI.md instructions. Identify which file, shell, web, MCP, or extension tools the task needs and disable or deny unrelated capability. Treat extensions and MCP servers as code and services requiring separate review.
Preserve a version-control checkpoint before mutation. If a tool executes against the wrong directory, produces an unverifiable result, or encounters an authentication mismatch, stop and recover before changing configuration.
Task: [one bounded outcome]
Scope: [working directory and allowed files]
Permissions: [tools, sandbox, shell, network, and MCP limits]
Authentication: [approved account path; no credential value]
Expected evidence: [diff, output, and sources]
Verification: [tests and independent review]
Stop conditions: [wrong directory, auth mismatch, unreviewed extension]
Recovery: [checkpoint, preserved logs, and safe retry]Expected evidence and verification
Expected evidence includes the selected authentication mode, effective tool boundary, repository diff, and test output. Verify that ignored files and secrets remain excluded, commands ran in the intended directory, and cited web evidence comes from the expected source.
For recovery, stop the current tool loop, preserve diagnostic output, restore the affected files from the checkpoint, and retry with fewer tools or a sandbox. Do not switch authentication methods or cloud projects merely to bypass an authorization failure.