// 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".
| 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". |
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).
python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py \
--prompt "<task prompt>" \
--cwd "$(pwd)"
Pass any markdown file as a persona/role for Gemini. YAML frontmatter is stripped automatically.
python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py \
--prompt "<task prompt>" \
--system-md /path/to/persona.md \
--cwd "$(pwd)"
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!
| 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 |
# Launch as a background task so Claude can continue working in parallel
Bash("python ~/.claude/skills/gemini-agent/scripts/gemini_interface.py --prompt '...' --cwd $(pwd)", run_in_background=true)
Gemini has no conversation context. Include everything it needs:
End-to-end orchestrator for multi-task parallel work. Phase 1 (setup) gathers code intelligence, analyzes merge-conflict risks, groups tasks into maximally-parallel phases, writes enriched self-contained agent prompts, and generates PLAN.md + STATE.md. Phase 2 (execution) dispatches Opus agents in worktrees with a per-task review-gauntlet sub-skill pass and auto-fix commits, then a doc-update agent at completion. Asks explicit permission before starting autonomous execution. Resilient to /clear via file-backed state. Triggers on /parallel-phases, "parallelize these tasks", "run these in parallel phases".
Use when errors occur deep in execution and you need to trace back to find the original trigger - systematically traces bugs backward through call stack, adding instrumentation when needed, to identify source of invalid data or incorrect behavior
Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.
