gudastudio-skills: Delegate Claude Code Work to Codex and Gemini

Claude Code is your main editor, but sometimes you want a second model's eyes — Gemini's take on a UI, Codex's read on a gnarly bug. GudaStudio's skills collection wires both CLIs into Claude Code as callable skills, so delegating work is one Python-bridge invocation instead of a copy-paste relay between terminals.

Why This Skill Matters

The collection ships exactly two skills, and each has one job. collaborating-with-codex delegates coding tasks to OpenAI's Codex CLI for prototyping, debugging, and code review; collaborating-with-gemini does the same for Google's Gemini CLI. Both return a unified JSON result and support multi-turn sessions through a SESSION_ID, so a follow-up question continues where the first call left off.

Prerequisites from the README: Claude Code v2.0.56 or newer, Python 3.8+ (the project was developed and tested on 3.10), and the CLI for whichever collaborator you install the bridge for.

Installation

git clone --recurse-submodules https://github.com/GuDaStudio/skills
cd skills
./install.sh --user --all

--user installs to ~/.claude/skills/ for every project; --project scopes to the current repo; --target picks a custom path; --skill collaborating-with-codex installs one skill instead of all. On Windows, the same options exist as .\install.ps1 -User -All. Run ./install.sh --list first to see what is available, and after installing, ask Claude "list all available SKILLs please" to confirm both bridges are recognized.

Real Workflow: A Cross-Model Security Review

The bridge is a Python script with two required arguments, the working directory and the prompt:

python scripts/codex_bridge.py --cd "/project" --PROMPT "Review the security of login.py"

The response JSON carries a SESSION_ID. Pass it back for the follow-up:

python scripts/codex_bridge.py --cd "/project" --SESSION_ID "uuid-from-response" --PROMPT "Provide fixes for the issues found"

For a full pipeline, the README recommends a global protocol in ~/.claude/CLAUDE.md (it depends on the auggie-mcp context engine) that turns the pair into a five-phase workflow: retrieve full codebase context first, then hand the raw requirement to Codex and Gemini for independent analysis and cross-validate their plans. Claude shows you the merged plan and must stop for your approval before writing anything. Prototypes come back as route-specific outputs — Gemini handles frontend and UI with a context ceiling the README sets below 32k tokens, Codex handles backend logic — and both are required to answer with a Unified Diff Patch only. Claude then refactors the diff into production code, and both external models audit the result.

Tips

• The Codex bridge's --sandbox parameter defaults to read-only; the README's security guidance is to keep it that way and to treat every returned diff as a dirty prototype that Claude refactors before anything lands.
• The recommended protocol demands the string "OUTPUT: Unified Diff Patch ONLY" in every prompt to external models — external models get zero filesystem write permission by design.
• Skills themselves are free and open source; running Codex or Gemini may incur API costs on their own pricing.
• Gemini's prototype output needs a critical eye: the README warns that its backend-logic understanding is weak and should be limited to frontend work.

When Not to Use This

The README's own protocol says simple tasks may skip multi-model collaboration entirely — under its global config the agent must stop, report that reason, and wait for your permission. And if your code cannot leave your machine, a bridge to cloud CLIs is the wrong shape; a locally hosted model behind an OpenAI-compatible endpoint fits better.

See the leaderboard for more skills.