openclaw-gemini-cli（OpenClaw Gemini CLI 搜索）

Gemini CLI

Use the locally installed gemini command through exec.

What this skill is for

Use this skill when you want Gemini CLI as a second brain on the same Linux machine. Typical uses:

• Get a second opinion from another model
• Ask Gemini to draft code, UI text, page copy, SQL, shell, or content
• Ask Gemini for structured JSON that OpenClaw can post-process
• Let OpenClaw inspect/edit files while Gemini generates plans, code, or rewrites
• Run quick one-shot prompts without spawning a separate ACP harness

This skill is for local CLI usage, not browser automation. If the user wants deep coding delegation over many steps, prefer ACP sessions or coding-agent.

Safety boundary

Gemini CLI can influence what gets executed, but it does not get direct tool access by itself. Keep the control loop like this:

1. OpenClaw reads files and inspects the environment
2. OpenClaw sends a bounded prompt to Gemini CLI
3. OpenClaw reviews Gemini output
4. OpenClaw decides what to execute, edit, or reject

Do not blindly pipe Gemini-generated shell into execution. Review before running anything destructive or external.

First-time setup and preflight

If this is the first time using Gemini CLI on a machine, do not assume install alone is enough. The user should complete Gemini CLI authentication first, then run a smoke test.

Basic checks:

which gemini
gemini --help
gemini -p "hello"

Interpretation:

• If gemini --help works but gemini -p "hello" fails, auth is usually missing or expired.
• If gemini -p "hello" returns a normal answer, the CLI is ready.

Practical rule:

1. Install Gemini CLI
2. Complete Gemini CLI login/authentication if prompted by the CLI
3. Run gemini -p "hello"
4. Only then rely on this skill

Known note on this machine:

• Gemini CLI may warn that keytar is missing and fall back to FileKeychain.
• That is not fatal if the prompt still completes.

Core usage patterns

1) Simple one-shot prompt

gemini -p "Summarize the tradeoffs between Astro and Next.js for a landing page site."

Use for quick answers, rewrites, summaries, naming, content ideas, or code snippets.

2) Save output to a file for later processing

gemini -p "Generate 5 hero headline options for a SaaS analytics site." > /tmp/gemini-headlines.txt

Then read or parse that file with OpenClaw tools.

3) Request strict JSON

Ask Gemini to return only JSON.

gemini -p 'Return only valid JSON: {"pages":[{"path":"/","title":"...","sections":["..."]}]}'

Best practice:

• Explicitly say "Return only valid JSON"
• Give the exact schema shape
• Avoid markdown fences
• Validate before trusting

4) Feed file context safely

Prefer bounded context, not giant dumps.

cat ./src/app/page.tsx | gemini -p "Review this file and suggest a cleaner component structure. Return concise bullet points only."

Better yet, use OpenClaw read first and then craft a focused prompt.

5) Multi-step AI-to-AI workflow

Recommended pattern:

1. OpenClaw inspects repo/files
2. Gemini produces plan/spec/UI copy/code draft
3. OpenClaw applies edits locally
4. Gemini reviews diff or rewritten file
5. OpenClaw finalizes

This is the right way to get “AI同士が会話して自律的に作る” behavior without surrendering execution control.

Good prompts for website generation

Site spec prompt

gemini -p 'You are helping design a small website. Return only JSON with keys: stack, pages, components, copy_style, seo_keywords, risks. Product: local Instagram analytics service for Japanese small businesses. Goal: conversion-focused landing page.'

Copywriting prompt

gemini -p 'Write Japanese landing page copy for a service that improves Instagram operations for stores. Tone: energetic, trustworthy, masculine, conversion-focused. Return sections for hero, pain points, benefits, CTA, FAQ.'

UI implementation prompt

gemini -p 'Generate a single-file HTML landing page with embedded CSS for a Japanese marketing service. Make it mobile-first, bold, readable, and production-lean. Output code only.'

Refactor/review prompt

gemini -p 'Review this React component for readability and conversion-focused UX. Return only: issues, fixes, rewritten component.'

When to use exec vs ACP/session delegation

Use Gemini CLI via exec when:

• The task is one-shot or short-loop
• You want another model's wording or plan quickly
• OpenClaw remains the main orchestrator
• You only need text/code output from Gemini

Do not force Gemini CLI through exec when:

• The task needs a persistent coding agent loop
• The task needs thread-bound collaboration in chat
• The task needs autonomous repo exploration over a long time
• The task needs tool use beyond shell text I/O

For those, prefer ACP sessions.

Recommended command patterns

Clean text response

gemini -p "<prompt>"

Model switching

If the installed Gemini CLI supports model selection flags in its current version, prefer explicit model selection when you need predictable behavior. Check supported options first:

gemini --help

Common practice:

• inspect help output for model-related flags or config options
• set the model explicitly when running important prompts
• reuse the same model during a multi-step workflow to reduce drift

Safe guidance for this skill:

• do not guess unsupported flags
• verify the exact syntax from gemini --help on the current machine
• if a default model changed unexpectedly, re-run with an explicit model setting supported by that CLI version

JSON capture

gemini -p 'Return only valid JSON with keys ...' > /tmp/gemini.json

Large prompt from file

Use a temp file when the prompt gets long or must be reproducible.

cat /tmp/prompt.txt | gemini -p "$(cat)"

If shell quoting gets ugly, write a temp prompt file first with write and use shell substitution carefully.

Operating rules for this workspace

• Use Gemini CLI as an advisor/generator, not an unsupervised executor
• Let OpenClaw own file edits, command execution, and validation
• For coding work, prefer this loop: inspect → ask Gemini → edit locally → test → ask Gemini to review
• If Gemini outputs code, save it to a temp file or inspect inline before applying
• If the output is meant to drive shell commands, require an explicit review step
• If authentication fails or Gemini hangs, fall back to OpenClaw-native work or ACP delegation

Optional helper scripts

If you need repeatable flows, use the bundled scripts in scripts/:

• scripts/gemini_json.sh — ask Gemini for JSON-only output
• scripts/gemini_review.sh — send a file to Gemini for review with a fixed prompt wrapper

Read those scripts before modifying them.

Troubleshooting

CLI exists but prompts fail

• Check auth state by running a tiny prompt like gemini -p "hello"
• If that fails, complete Gemini CLI login/authentication first
• Expect possible FileKeychain fallback warning
• If output still returns, it is usable

Need to change models but not sure how

• Run gemini --help
• Look for model-selection flags or config options supported by that installed version
• Use explicit model selection for important or repeatable workflows
• Do not rely on guessed flags copied from a different Gemini CLI version

Output is too fluffy

• Demand exact format
• Say Return only valid JSON
• Say Output code only
• Give explicit section names or schema

Shell quoting is breaking the prompt

• Write the prompt to a temp file first
• Avoid nested quote hell
• Keep prompts plain and deterministic

User wants full autonomous site building

That is possible only in a controlled loop. Use Gemini CLI to generate plans/code, but keep execution and editing in OpenClaw. For larger builds, combine this skill with ACP or coding-agent.