domains / agent-tooling / gemini-cli-invocation / gemini

gemini

Skill card

combined

The three copies describe different procedures for invoking Gemini. The first is a cross-model second-opinion workflow for Google Ads/SEO that delegates to a separate WORKFLOW.md. The second is a detailed guide for running the Gemini CLI in background/automated contexts, with specific model selection, approval-mode rules, and timeout handling. The third is a minimal one-shot Q&A guide that explicitly warns against using --yolo. Because the second and third copies directly conflict on the use of --yolo, and the first copy targets a different use case (Google ecosystem review) than the CLI execution guides, crowning one would hide real alternatives.

Repositories

Other copies

Combined skill

# Gemini Skill

## 1. Determine Use Case
- **Google Ecosystem Review**: If the user requests a "gemini review", "challenge", or "consult" specifically for Google Ads, SEO, or campaign strategy, follow the NotFair cross-model workflow. Read the associated WORKFLOW.md and use Gemini as a second opinion with native Google knowledge.
- **CLI Execution**: If the user requests code review, plan review, or large-context processing, use the Gemini CLI.

## 2. CLI Execution Guidelines

### Model Selection
Ask the user which model to use if not specified. Options include:
- `gemini-3-pro-preview`: Flagship, best for complex reasoning/coding.
- `gemini-3-flash`: Sub-second latency, speed-critical.
- `gemini-2.5-pro` / `gemini-2.5-flash`: Legacy options.

### Approval Mode & Safety
- **Interactive Terminal**: You may use `--approval-mode default` or `auto_edit`.
- **Background/Automated (e.g., Claude Code tool calls)**:
  - **Conflict Note**: `softaworks/agent-toolkit` recommends `--approval-mode yolo` to prevent hanging. `LeoYeAI/openclaw-master-skills` advises avoiding `--yolo` for safety.
  - **Resolution**: If using `--approval-mode yolo`, ensure the task is strictly read-only or sandboxed (`-s`). If safety is a priority, use `timeout` wrappers and avoid `yolo` if possible, or use `auto_edit` if the environment supports non-interactive approval for edits only. **NEVER** use `--approval-mode default` in non-interactive shells as it will hang.

### Command Assembly
- Use `-m <model>` to specify the model.
- Use `--include-directories <DIR>` for multi-file analysis.
- Use `--output-format json` for structured output if needed.
- For one-shot Q&A, use positional prompts: `gemini "Prompt..."`.

### Execution
- **Background**: `timeout 300 gemini -m <model> --approval-mode <mode> "Prompt"`
- **Interactive**: `gemini -m <model> -i "Prompt" --approval-mode <mode>`

## 3. Post-Execution
- Inform the user that the analysis is complete.
- If the process hangs (0% CPU, sleeping state), kill it with `pkill -9 -f "gemini"`.

## Conflicts
- **Repository**: `softaworks/agent-toolkit` vs `LeoYeAI/openclaw-master-skills`
- **Step**: Use of `--approval-mode yolo` / `--yolo`.
- **Detail**: `softaworks` requires it for background tasks to prevent hangs. `LeoYeAI` advises against it for safety. The agent must weigh the risk of hanging against the risk of unauthorized actions, potentially using timeouts or sandboxes as mitigations.