gemini

name	gemini
description	[Hyper] Use when the user wants to invoke Google Gemini CLI (`gemini`) for reasoning, research, or AI assistance. Trigger phrases: "use gemini", "ask gemini", "run gemini", "call gemini", "gemini cli", "Google AI", "Gemini reasoning", or when users request Google's Gemini models, research with web search, plan-mode review, or want to resume a previous Gemini session. Do not use for generic writing, runbook cleanup, or local edits that do not require the Gemini CLI.
compatibility	Requires Google Gemini CLI (`gemini`) v0.21.1 or later. Targets v0.40.0 behavior. Works only in environments where that CLI is installed and authenticated.

@rules/routing.md

Gemini Skill

<output_language>

Default all user-facing deliverables, saved artifacts, reports, plans, generated docs, summaries, handoff notes, commit/message drafts, and validation notes to Korean, even when this canonical skill file is written in English.

Preserve source code identifiers, CLI commands, file paths, schema keys, JSON/YAML field names, API names, package names, proper nouns, and quoted source excerpts in their required or original language.

Use a different language only when the user explicitly requests it, an existing target artifact must stay in another language for consistency, or a machine-readable contract requires exact English tokens. If a localized template or reference exists (for example *.ko.md or *.ko.json), prefer it for user-facing artifacts.

</output_language>

Wrap the official Google Gemini CLI for reasoning, research, plan-mode review, and session continuation. The CLI must be invoked headlessly with -p / --prompt — positional prompts open the interactive REPL.

<instruction_contract>

Field	Contract
Intent	Build, run, or explain a safe gemini CLI invocation when the user explicitly asks for Gemini CLI or Gemini models.
Scope	Owns Gemini CLI command shape, headless prompt usage, approval mode choice, model selection only when requested, resume flows, preflight checks, and result summarization.
Authority	User intent and local project rules come first; installed CLI help and official Gemini CLI docs govern flags when current behavior matters.
Evidence	Ground the answer in preflight checks, command shape, CLI output, approval mode, and observed warnings/errors.
Tools	Use shell execution only when the user wants the CLI run; otherwise provide the command or route away.
Output	Return the exact command used or recommended, summarize output/warnings, and preserve CLI/model tokens verbatim.
Verification	Check gemini is installed before command construction, non-interactive runs use -p, edit/yolo modes are explicit, and model overrides are user-requested.
Stop condition	Stop after the requested Gemini command is constructed or run and its result, warnings, or blocker are reported.

</instruction_contract>

Defaults

Parameter	Default
Model selection	Use the Gemini CLI default (-m auto) unless the user explicitly asks for -m pro, -m flash, -m flash-lite, or a specific id
Approval mode	--approval-mode default (prompts on every tool call)
Headless mode	-p / --prompt
Resume target	gemini --resume latest (or gemini -r for the most recent)
Output format	text (use --output-format json only when the caller needs structured output)

Do NOT specify -m unless the user explicitly requested a model. Auto routing picks Pro for reasoning and Flash for execution.

Routing

Use this skill only when the request actually needs the Gemini CLI or a Gemini session.

• Read rules/routing.md before building a command when the request might be out of scope.
• Route away to direct editing or another skill when the user wants generic writing, runbook cleanup, or local edits without needing the Gemini CLI.

Examples

Positive examples:

• "Use Gemini to review this architecture and list risks."
• "Run gemini -p and research the current API behavior."
• "Ask Gemini Pro to reason about this design tradeoff."
• "Resume the latest Gemini session and continue the analysis."

Negative examples:

• "Rewrite this runbook for readability."
• "Make this local code edit directly."
• "Create a new skill for this repo."

Boundary examples:

• "Research Gemini CLI approval modes and explain them." Use this skill only if the user wants the gemini CLI involved; otherwise route to research or direct documentation work.

CRITICAL: Headless Mode

Always use -p / --prompt for non-interactive Gemini runs. A positional prompt drops into the interactive REPL and blocks automation.

# Correct — non-interactive
gemini --approval-mode default -p "your prompt here"

# Interactive (do not use for automation)
gemini "your prompt here"

If you need to seed the prompt and then keep talking interactively, use -i / --prompt-interactive instead. -p is for one-shot, automated runs.

Running a Task

Read references/recipes.md for concrete command recipes before changing approval modes, adding sandboxing, picking a model, or resuming a session.

Approval Mode Selection

Flag	When to use
--approval-mode default	General research, reasoning, and standard Gemini usage
--approval-mode auto_edit	Only when the user explicitly asks Gemini to edit files
--approval-mode plan	Read-only planning or analysis with no code modifications (write is restricted to Markdown in the plans directory)
--approval-mode yolo	Only after the user explicitly approves a full-auto run. Prefer this spelling — the bare --yolo flag is deprecated.

Command Discipline

• Start from gemini --approval-mode default -p "your prompt here".
• Add -m <pro|flash|flash-lite|auto|model-id> only when the user explicitly asks for a specific Gemini model.
• Use --approval-mode auto_edit for normal edits instead of full auto-approve.
• Ask before using --approval-mode yolo. Prefer --approval-mode auto_edit or --sandbox first.
• Add --sandbox (-s) when the task is higher-risk or you want tighter local constraints.
• Pass extra workspaces with --include-directories /path/a,/path/b when Gemini needs context outside the current folder.
• Use --output-format json only when the caller will parse the response programmatically.

Resuming a Session

gemini --resume latest -p "your follow-up prompt"
gemini --list-sessions
gemini --delete-session 2

--resume accepts latest, an index from --list-sessions, or a full session UUID. Inside an interactive session, use /resume to open the Session Browser and /rewind to step backward.

When resuming, keep the existing session behavior and do not change model or approval mode unless the user asks.

After Completion

• Summarize the result, including any warnings or partial output.
• Tell the user they can resume with gemini --resume latest (or gemini -r).
• Ask whether to continue, adjust the prompt, or switch back to direct work.

Critical Evaluation

Treat Gemini as a colleague, not an authority.

• Trust your knowledge when confident. Push back on incorrect claims directly.
• Research disagreements using current sources or docs before accepting Gemini's claims.
• Remember knowledge cutoffs — Gemini may not know about recent releases or API changes.
• When freshness matters, ask Gemini for current sources and verify important claims against docs or primary sources.
• Let the user decide when there is genuine ambiguity.

Preflight Check

Before building any command, verify the CLI is installed and on PATH:

command -v gemini >/dev/null 2>&1 || { echo "gemini CLI not found"; exit 1; }

If the check fails, route to the install handler in Error Handling below instead of constructing a gemini ... command.

Error Handling

• command not found: gemini: tell the user to install with npm install -g @google/gemini-cli (or brew install gemini-cli).
• Auth errors: run gemini and complete the sign-in flow, or set GEMINI_API_KEY. For Vertex AI, set GOOGLE_API_KEY and GOOGLE_GENAI_USE_VERTEXAI=true.
• 429 / rate limit: free tier is 60 req/min, 1000 req/day on Gemini 3 — wait or upgrade.
• 404 / model not found: retry without -m, or pick pro/flash/flash-lite from gemini --help and the official docs.
• 403 / access denied: check the active auth path and retry with a supported model or account configuration.
• Session not found: run gemini --list-sessions.
• Exit code 42 means input error (bad flag/prompt); 53 means turn limit exceeded.

Version Notes

• Targets Gemini CLI v0.40.0 behavior; minimum supported v0.21.1 (Gemini 3 routing).
• --yolo is deprecated; the canonical form is --approval-mode yolo.
• Plan Mode is enabled by default; trigger explicitly with --approval-mode plan or /plan in-session.
• Headless mode auto-activates in non-TTY environments and whenever -p is supplied.

<validation_checklist>

• The request explicitly needs Gemini CLI, Gemini models, or a Gemini session.
• command -v gemini or equivalent preflight is checked before constructing a runnable command.
• Non-interactive runs use -p / --prompt; positional prompts are not used for automation.
• Model flags are omitted unless the user explicitly requested a model.
• auto_edit and yolo approval modes are used only under their documented gates.
• Final output includes the command shape, observed result, warnings, and any install/auth/rate/model blocker.

</validation_checklist>