gemini

name	gemini
description	Codex skill alias for /gemini. Direct Gemini invocation with model selection. Delegates a prompt to Gemini CLI and returns the response. Use for ad-hoc tasks that benefit from Gemini's search grounding, large context, or code execution. Use when this capability is needed.
metadata	{"author":"Design-Machines-Studio"}

Codex Command Alias: /gemini

• This generated skill exposes the Claude slash command /gemini to Codex.
• Treat the user's message that invoked this skill as the command arguments.
• When the embedded command body says $ARGUMENTS, substitute those user-supplied arguments.
• Follow the embedded command body as the workflow source of truth.
• Translate Claude-only tool names to Codex equivalents when needed: use update_plan for TodoWrite-style ledgers, request_user_input for AskUserQuestion-style gates when available, and normal chat questions when that tool is not available.
• If the command body says to launch an agent but no direct subagent tool is available, use the best available Codex delegation tool or execute the referenced agent protocol yourself while clearly noting the degradation.

Embedded Claude Command

/gemini

Delegate a prompt directly to Gemini CLI.

Usage

/gemini What are the current best practices for Datastar SSE performance?
/gemini --model pro Analyze this complex architectural decision...
/gemini --model flash-lite What version of Go is latest?

Process

Step 1: Parse Arguments

Extract the prompt and optional --model flag from the user's input.

• If --model is specified, use that model
• If --model is not specified, auto-select based on prompt length:
  • <500 characters: flash-lite (fast factual lookup)
  • 500-5000 characters: flash (balanced)

  • 5000 characters: pro (deep reasoning)

Step 2: Select Timeout

Based on model:

• flash-lite: 10s
• flash: 60s
• pro: 180s

Step 3: Invoke Gemini

Use heredoc to avoid shell injection from user input:

RESULT=$(timeout ${TIMEOUT} cat <<'GEMINI_INPUT' | gemini -m ${MODEL} --yolo --output-format json --raw-output 2>/dev/null
${USER_PROMPT}
GEMINI_INPUT
)

Never use -p "${USER_PROMPT}" — user input may contain shell metacharacters ($(...), backticks, double-quotes) that break out of quoting. Always use the heredoc pattern with a quoted delimiter.

Step 4: Handle Errors

Check for the four failure modes (timeout, rate limit, empty, malformed).

On rate limit, fall back through the model chain: pro → flash → flash-lite → report failure.

On other failures, report the error type to the user.

Step 5: Present Response

Extract the response field from Gemini's JSON output and present it to the user.

If Gemini used google_web_search (check stats.tools.byName), note: "Sources cited via Google search grounding."

If Gemini ran code (check for code execution in stats), note: "Code executed in Gemini's Python sandbox."

---

Source: Design-Machines-Studio/depot — distributed by TomeVault.