# Peer Review

> Ask a group of LLM tools for their opinion on a given task. Use this when stuck on a problem, seeking a second opinion, validating an approach, or wanting diverse viewpoints on complex decisions.

- Skill: `pietz/peer-review` (Agent Skill)
- Install (CLI): `npx skillmds@latest add pietz/peer-review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/pietz/peer-review/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: pietz (https://skillmd.com/u/pietz)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/pietz/peer-review

---


# Peer Review

Query multiple LLM CLIs in parallel and collect their responses.

## Available CLIs

All three should be invoked via the Bash tool as background tasks, run in parallel:

- **Claude**: `unset CLAUDECODE && claude --model opus -p "$PROMPT"`
- **Codex**: `codex exec -m "gpt-5.5" --skip-git-repo-check "$PROMPT"`
- **Gemini**: `gemini -m "gemini-3.1-pro-preview" -p "$PROMPT"`

## Self-invocation rule

The orchestrating agent must NOT call its own CLI via Bash — it will fail or produce empty output. Use a **subagent** for your own model's contribution and Bash CLI commands for the other models.

## Workflow

1. Craft a clear, self-contained prompt. The CLIs have no history — fully re-brief them each time.
2. Identify which agent you are (the orchestrator) so you know which CLI to skip.
3. Check which of the remaining CLIs are installed (`which claude codex gemini`).
4. Launch the other CLIs as **parallel background Bash tasks** and your own model as a **subagent**, all in parallel.
5. Collect results from all models.
6. **Report transparently**: before synthesizing, tell the user which models responded, which failed or timed out, and any issues encountered.
7. Present each model's response in its own section with clear attribution.
8. Provide a unified synthesis highlighting where the models agree, disagree, and what unique insights each contributed.

## Guidelines

- Think about how much you want to reveal about your own opinion when crafting the prompt.
- If a CLI fails or times out, report it and continue with the results you have.
- Responses will arrive at different speeds. As each result comes in, consider giving the user a brief update or early summary rather than waiting silently for all models to finish. The final synthesis still needs all responses, but intermediate progress keeps the user informed.