gemini-delegate

Gemini Delegate Skill

Claude is the supervisor. Claude decides scope, supplies context, and performs final review. Gemini is the specialist for large-context synthesis, long-form drafting, bilingual or CJK writing, and second-opinion analysis.

Prerequisite check (do this first)

This skill emits Gemini CLI invocations. Before producing any task file, wrapper command, or handoff prompt, verify the binary is on $PATH:

gemini --version

If that command is not found, stop and tell the user:

This skill needs the Gemini CLI. Install it with:

npm install -g @google/gemini-cli
gemini --version

Then re-run your request.

Do not prepare a task prompt, write a wrapper command, or fabricate a result.json. Without the binary on PATH, every "successful" wrapper run is a hallucination.

When to Use

Do not use Gemini as a mirror copy of Codex. Its value is different.

Route to	Best for	Avoid
Gemini	Large-context summarization, English or zh-TW/CJK writing, bilingual synthesis, reviewer-style second opinion, release-note drafting	Bulk code generation, architecture decisions, security-sensitive coding
Codex	Mechanical implementation, refactors, test scaffolding, batch edits	Large-context reading and nuanced synthesis
Claude	Requirements, acceptance judgment, debugging root cause, final publication review	Long repetitive drafting

If the task is "read a lot, synthesize, compare, or rewrite carefully in English or Chinese," Gemini is a good candidate.

Required Output Contract

Every wrapper run must leave machine-readable status in:

<log-file>.result.json

Required fields:

{
  "status": "success|fallback|error|verify_failed",
  "delegate": "gemini",
  "model": "gemini/<model>",
  "log_file": "<path>",
  "summary": "",
  "risks": [],
  "files_changed": [],
  "tests_run": [],
  "timestamp_utc": "2026-04-24T00:00:00Z"
}

The wrapper contract is transport status only. Claude still owns factual review, terminology checks, and publication quality.

If the task was queued as part of a multi-agent run by agent-task-splitter (from the agent-collab-skills marketplace), its task file lives at .ai/gemini_task_<NNN>_<slug>.md and the round plan is at .coord/plan.yml. Read the plan for context (which other agents are running, what each task's success criteria are).

Good Delegation Targets

• Summarize a long English report into concise English or zh-TW
• Compare multiple docs and produce one synthesized brief
• Rewrite translated content into more natural Traditional Chinese
• Draft release notes, updates, or FAQs from source material in English or Chinese
• Provide a second-opinion review over a long design or doc set
• Align terminology across bilingual documents

Bad Delegation Targets

• Generate or refactor production code across many files
• Diagnose a flaky test or deep runtime bug
• Decide architecture or API boundaries
• Review auth, secret handling, or validation logic
• Publish translation output without Claude review

Supervisor Workflow

1. Write a task file

Create .ai/gemini_task_<name>.md:

# Task: <descriptive name>

## Context
- Repo: C:\path\to\repo
- Read these files first:
  - docs/spec.md
  - docs/changelog.md
- Output file(s):
  - docs/output_zh-TW.md

## Goal
<what Gemini should synthesize or draft>

## Language
- Output language: Traditional Chinese
- Tone: formal / concise / executive / technical
- Audience: <who will read it>

## Constraints
- Preserve dates and proper nouns exactly
- Keep terminology consistent with glossary.md
- Do not invent facts missing from the sources

## Acceptance
- Required verification files: <paths>
- Required sentinel string: <string if useful>
- Claude will perform a terminology and factual review before shipping

2. Launch Gemini synchronously

From Claude Code Bash:

bash .claude/skills/gemini-delegate/scripts/run_gemini.sh \
  --prompt "Read .ai/gemini_task_<name>.md and execute all instructions inside." \
  --log-file .ai/gemini_log_<name>.txt \
  --verify-file docs/output_zh-TW.md

PowerShell direct call is also supported:

& "C:\Users\wenyu\mispricing-engine\.claude\skills\gemini-delegate\scripts\run_gemini.ps1" `
    -Prompt "Read .ai/gemini_task_<name>.md and execute all instructions inside." `
    -LogFile "C:\Users\wenyu\mispricing-engine\.ai\gemini_log_<name>.txt" `
    -VerifyFile "C:\Users\wenyu\mispricing-engine\docs\output_zh-TW.md"

3. Read wrapper status first

cat .ai/gemini_log_<name>.txt.result.json

If status is verify_failed, the process exited but required files were missing or incomplete. Treat that as failure.

4. Claude publication review

Claude must review:

• factual accuracy against source files
• dates, names, and terminology consistency
• tone and audience fit
• any banned or sensitive phrasing required by the project

Gemini can draft or synthesize. Claude decides whether the output is publishable.

Non-Interactive Execution Rules

These wrappers already enforce the critical Gemini CLI rules:

• run from the target working directory instead of using a fake -C
• pass --approval-mode yolo
• pipe prompts through stdin instead of a giant positional argument
• verify expected files after exit when --verify-file is supplied

If you write your own wrapper, preserve all four behaviors.

Quality Boundaries

Gemini is useful for first drafts and synthesis, but it can still:

• drift terminology across files
• over-translate proper nouns
• miss project-specific banned phrases
• guess dates or context if the prompt is underspecified

Do not ship its output unreviewed.

Minimal Review Checklist

Before accepting delegate output:

• Did Gemini cite or preserve the correct source facts?
• Did it keep terminology consistent with project vocabulary?
• Did it stay within the requested language and tone?
• Did required output files actually exist on disk?
• Does the result belong to Gemini, or should this have stayed in Claude/Codex?

If the answer to the last question is wrong, fix your routing first.