| name | gemini-agent |
| description | Delegate tasks to Gemini CLI as an agent. Use when the user wants to offload token-heavy work to Gemini, get a second opinion from a different model, or run Gemini with a custom system prompt. Triggers include "use Gemini to review X", "ask Gemini to Y", "run Gemini on Z", "delegate to Gemini", "have Gemini check X".
|
Gemini Agent
Delegates tasks to Gemini CLI in headless mode using gemini_interface.py from this skill's scripts/ folder.
Prerequisite: Gemini CLI must be installed and on PATH (gemini --version should work).
Basic usage
python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py \
--prompt "<task prompt>" \
--cwd "$(pwd)"
With a system prompt file
Pass any markdown file as a persona/role for Gemini. YAML frontmatter is stripped automatically.
python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py \
--prompt "<task prompt>" \
--system-md /path/to/persona.md \
--cwd "$(pwd)"
With a context file
If you need Gemini to read a large existing file, attach it via the --context-file argument (which uses Gemini's @file syntax under the hood):
python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py \
--prompt "Review this file for security vulnerabilities." \
--context-file src/auth/login.py \
--cwd "$(pwd)"
Note: For short context or summaries (a few sentences), do not create a temp file. Just include the context directly inside your --prompt string!
Options
| Flag | Default | Description |
|---|
--prompt | required | Task prompt |
--system-md | none | Markdown file prepended as system instructions |
--context-file | none | File attached via @path in the prompt |
--cwd | current dir | Working directory for Gemini CLI |
--model | gemini-3.1-pro-preview | Gemini model ID |
--verbose | off | Stream all output; default is final response only |
Running in background (parallel with other work)
Bash("python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py --prompt '...' --cwd $(pwd)", run_in_background=true)
Constructing good prompts
Gemini has no conversation context. Include everything it needs:
- Which files or components to examine
- What specifically to look for
- What format to return (e.g., "list issues with file:line references")