From gemini
Parallel sub-agent backed by Gemini CLI with live Google Search grounding. Free of Claude's quota — default to calling it, don't deliberate. TRIGGER for anything with an information dimension: facts, current events, research, comparisons, versions, prices, news, trends, analysis, verification, "what's the latest", "is X still true". Even routine questions may have recent developments worth catching. Default is Gemini-only. Claude's own WebSearch only when explicitly needed for citable sources or when Gemini's result is insufficient. Skip for pure file edits, git ops, mechanical code changes.
How this skill is triggered — by the user, by Claude, or both
Slash command
/gemini:geminiThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Gemini CLI at `/opt/homebrew/bin/gemini`. Runs on the user's Google account — independent
Gemini CLI at /opt/homebrew/bin/gemini. Runs on the user's Google account — independent
of Claude's token budget. Internally uses gemini-3.1-pro-preview with thinking,
google_web_search, and a Flash sub-agent. One call = a full agentic pipeline.
bash ~/.claude/skills/gemini/scripts/call_gemini.sh "your prompt"
Always run_in_background: true. Launch it first, then do your own work in parallel.
If no result after ~45s, proceed without it.
Info query?
└─ Yes → Gemini only (default)
└─ Result sufficient? → Done
└─ Result thin / off-target?
├─ Refine prompt → call Gemini again
└─ Need citable URLs → add targeted WebSearch
└─ No (pure code/file/git) → Skip
Claude's own WebSearch is a fallback, not a default. Only reach for it when:
For everything else — single questions, current events, facts, prices, comparisons — Gemini alone is enough. Don't run WebSearch in parallel "just in case".
If the first result misses the mark, don't switch to WebSearch by default — try Gemini again with a better prompt first. Common fixes:
A second Gemini call is cheaper and often sufficient.
Only when the user explicitly asks for comprehensive research, analysis, or a report — and needs citable sources alongside breadth/recency.
Gemini framing for research:
"Claude is covering citable sources. You focus on: past 6 months news, real-world adoption, community sentiment, practitioner opinions."
Before calling Gemini, transfer everything you already understand — current date, user's situation, precise terms, what's already known. Gemini should start with the same picture you have, not reconstruct it from scratch.
Beyond that, just ask for what you need. No rigid template.
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub lcy000/claude-gemini-plugin --plugin gemini