cross-review

name

cross-review

description

Multi-perspective cross review for product, design, and technical proposals. Reviews proposals from user, system, and cost perspectives using isolated reviewer roles. Works out of the box with zero configuration (single-model with sub-agent isolation). Optionally configure a second model (GPT, Claude, Gemini, DeepSeek, GLM, Qwen, Kimi, MiniMax) for higher quality independent review. Use this skill whenever the user asks for cross review, second opinion, multi-perspective review, or wants their proposal challenged from different angles. Also triggers on: 交叉评审, 第二意见, 评审一下, quick review, full debate.

Cross Review

Run multi-perspective cross reviews on any proposal. The AI reviews your proposal from different expert perspectives (user experience, system design, cost/complexity) using isolated reviewer roles, then synthesizes the results into actionable output.

Two operating modes:

• Zero-config (default): Uses sub-agent isolation or role-play to simulate independent reviewers. Works immediately, no API keys needed.
• Multi-model (enhanced): Calls a second AI model for truly independent review. Higher quality, requires one external API key.

Prerequisites

Zero-config mode (default): No setup needed. The skill uses sub-agent isolation or role-play to simulate independent reviewers.

Multi-model mode (enhanced): A config file at ~/.cross-review-config.json with at least one external model's API key. See the Configuration section below.

First-Time Setup (Multi-Model Mode Only)

If the user wants to use a second model and no config file exists, guide them:

1. Ask which AI product they are currently using (claude / gpt / gemini / etc.)
2. Ask which reviewer model(s) they have API keys for
3. Collect the API key(s) and base URL if non-default
4. Run the setup script to create the config:

python3 {SKILL_DIR}/scripts/setup.py \
  --current-model claude \
  --add-model gpt --api-key "sk-..." \
  --add-model deepseek --api-key "sk-..." --base-url "https://api.deepseek.com/v1"

If python3 is not available, tell the user to install Python 3.8+ first.

If {SKILL_DIR} does not resolve on the current platform, use the installed skill path directly (e.g. ~/.claude/skills/cross-review/scripts/setup.py).

Trigger Phrases

• cross review, second opinion, multi-model review
• 交叉评审, 第二意见, 让另一个模型看看
• quick review, full debate, deep review

Entry Flow

Step 1: Detect Model Environment

python3 {SKILL_DIR}/scripts/reviewer.py --list-models

Parse the returned JSON:

• has_reviewer: true → multi-model mode, continue to Step 2
• has_reviewer: false → single-model mode, jump to modes/solo-review.md

If the command fails (config file missing, python3 not found, etc.), use single-model mode — jump to modes/solo-review.md. This is the default experience and works without any configuration.

Step 2: Select Review Mode (multi-model mode)

Check the trigger phrase or user input for an explicit mode:

• Contains "quick review" or "quick" → use Quick Review, skip asking
• Contains "full debate" or "full" or "deep" → use Full Debate, skip asking
• No explicit mode → check default_mode in config:
  • Has value (quick or full) → use that default, tell user they can switch
  • No value (first use) → ask the user:

This is your first time using cross-review. Pick a default mode:

• Quick Review: Primary model proposes → reviewer model critiques → synthesis. ~3 steps, good when you already have a draft and want fast validation.
• Full Debate: Both models propose independently → cross-critique → rebuttals → synthesis. ~7 steps, good for early-stage ideas that need thorough challenge.

I'll remember your choice. Say "quick review" or "full debate" anytime to switch.

After the user confirms, save their choice:

python3 -c "
import json, pathlib
p = pathlib.Path.home() / '.cross-review-config.json'
c = json.loads(p.read_text())
c['default_mode'] = 'quick'
p.write_text(json.dumps(c, ensure_ascii=False, indent=2))
"

Replace 'quick' with 'full' if the user chose Full Debate.

Step 3: Execute

• Quick Review → follow modes/quick-review.md
• Full Debate → follow modes/full-debate.md

Unified Output Schema

All modes MUST produce output in this exact structure:

**Final Proposal**
(Recommended proposal incorporating review feedback, complete and actionable)

**Resolved Disagreements**
(Points where reviewers initially disagreed but reached consensus. Write "None" if empty)

**Decisions Needed From You**
(Each in this format, or "None" if empty)

> **Disagreement N: [one-line summary of the decision]**
> - Context: [1-2 sentences on why this is a disagreement]
> - Option A: [approach] → Pros: ... / Cons: ...
> - Option B: [approach] → Pros: ... / Cons: ...
> - My recommendation: A/B, because [one-line reason]

**Suggestions to Ignore for Now**
(Each: [suggestion] → Reason to skip: [no current need / complexity not worth it / future optimization])

Configuration File Format

Create ~/.cross-review-config.json:

{
  "current_model": "claude",
  "default_mode": "quick",
  "reviewer_priority": ["gpt", "deepseek", "glm", "gemini", "qwen", "kimi", "minimax", "claude"],
  "models": {
    "gpt": {
      "api_key": "sk-...",
      "base_url": "https://api.openai.com/v1",
      "model": "gpt-5.4"
    }
  }
}

Field reference:

Field	Required	Description
current_model	Yes	The AI product you are using now (claude / gpt / gemini / etc.). The reviewer will skip this model.
default_mode	No	Default review mode (quick or full). Set automatically on first use.
reviewer_priority	No	Order of preference for selecting the reviewer model. Defaults to all configured models.
models	Yes	At least one model with an api_key.
models.*.api_key	Yes	API key for the model. Can also use env var CROSS_REVIEW_{MODEL}_KEY (e.g. CROSS_REVIEW_GPT_KEY).
models.*.base_url	No	API endpoint. Required for non-default providers (DeepSeek, GLM, Qwen, Kimi, MiniMax).
models.*.model	No	Model name to use. Defaults vary by provider.
models.*.max_tokens	No	Max response tokens. Default: 4096.
models.*.temperature	No	Sampling temperature. Default: 0.3.
models.*.wire_api	No	API protocol: chat (default, OpenAI-compatible), anthropic, gemini, responses (OpenAI Responses API).
models.*.ssl_verify	No	Set to false to skip SSL verification. Default: true.

Minimal config (one reviewer model is enough):

{
  "current_model": "claude",
  "models": {
    "gpt": {
      "api_key": "sk-...",
      "model": "gpt-5.4"
    }
  }
}

Full config with all supported models:

{
  "current_model": "claude",
  "default_mode": "quick",
  "reviewer_priority": ["gpt", "deepseek", "glm", "gemini", "qwen", "kimi", "minimax", "claude"],
  "models": {
    "gpt": {
      "api_key": "sk-...",
      "base_url": "https://api.openai.com/v1",
      "model": "gpt-5.4"
    },
    "claude": {
      "api_key": "sk-ant-...",
      "model": "claude-sonnet-4-20250514"
    },
    "gemini": {
      "api_key": "AIza...",
      "model": "gemini-2.5-pro"
    },
    "deepseek": {
      "api_key": "sk-...",
      "base_url": "https://api.deepseek.com/v1",
      "model": "deepseek-chat"
    },
    "glm": {
      "api_key": "...",
      "base_url": "https://open.bigmodel.cn/api/paas/v4",
      "model": "glm-5.1"
    },
    "qwen": {
      "api_key": "sk-...",
      "base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",
      "model": "qwen-max"
    },
    "kimi": {
      "api_key": "sk-...",
      "base_url": "https://api.moonshot.cn/v1",
      "model": "kimi-k2.6"
    },
    "minimax": {
      "api_key": "...",
      "base_url": "https://api.minimax.chat/v1",
      "model": "MiniMax-M2.7"
    }
  }
}