所有对用户的回复在发出前都必须通过对抗性审查。
何时使用
每一次对用户的回复，无一例外。这是一个强制质量关卡。
工作原理
先按常规起草回复（但不要发送），然后生成一个带有批判性审查者人格的子智能体。该子智能体检查草稿中的问题，你再根据其反馈修改。
步骤 1：起草回复
照常撰写回复。暂勿发送。
步骤 2：生成审查子智能体
使用 sessions_spawn，设置 mode="run"，并选用更便宜的模型（如 qwen3.5-plus、gpt-4o-mini）以降低成本。审查者无需完整工具集——它只阅读文本并输出批评意见。
注意： 这会把草稿文本发送到已配置的模型提供商 API。若使用云端模型，数据将传输到提供商服务器。若处理敏感内容，请查阅提供商的数据政策。要实现纯本地审查，请配置本地模型（如 ollama、LM Studio）。
使用以下提示模板生成：
`` You are an adversarial reviewer. Examine the following draft response critically. Draft: """ [YOUR DRAFT HERE] """ Check for: 1. Factual errors or unsubstantiated claims 2. Missing important caveats or edge cases 3. Logical contradictions 4. Tone issues (too apologetic, too confident, dismissive, etc.) 5. Missing follow-up suggestions that would be valuable 6. Overly verbose sections that could be trimmed 7. Any advice that could backfire or cause problems Respond with exactly one of: - PASS + one-line reason why it's fine - List of specific issues, each with: - Where the problem is - Why it's a problem - Suggested fix Be harsh. Be picky. Better to over-catch than to miss. You are the quality gate. `
步骤 3：采纳反馈
- 子智能体说 PASS → 原样发出草稿
- 子智能体提出有效问题 → 修改草稿，再发出改进版
- 子智能体明显错误 → 相信自己的判断，发出你的版本
步骤 4：发出
若审查导致实质性改动，可简要说明改进（如“审查发现 X，已修正 Y”）。仅微调则无需提及。专注于交付最佳结果。
隐私与安全
- 草稿文本通过已配置的模型 API 发送至第二个 AI 模型实例。若使用云端模型（如 qwen3.5-plus、gpt-4o-mini），数据将传输到提供商服务器。要实现纯本地审查，请使用本地模型提供商（如 ollama、LM Studio）。
- 仅共享草稿文本（不含完整对话历史）。
- 若草稿含敏感数据（PII、凭据等），智能体应自动跳过审查步骤。
- 审查交互不会在当前智能体会话之外留存。
例外情况
以下情况无需审查：
- HEARTBEAT_OK`
- 系统级确认（工具结果、NO_REPLY）
- 纯粹机械确认，不含任何主观内容
为何重要
LLM 输出可能包含细微错误、缺失上下文或语气问题，创作者视角容易忽略。第二双“对抗性眼睛”能在问题抵达用户前捕获它们。这相当于智能体界的代码审查。
注意：此审查步骤会增加每次回复的延迟与 token 消耗。
技术细节
无需特殊配置。要禁用审查，卸载此 skill 即可。
Every response to the user must pass adversarial review before delivery.
When to Use
EVERY response to the user, without exception. This is a mandatory quality gate.
How It Works
After drafting your response (but before sending it), spawn a sub-agent with a critical reviewer persona. The sub-agent examines your draft for problems, then you revise based on its feedback.
Step 1: Draft Your Response
Write your response as you normally would. Do not send it yet.
Step 2: Spawn Reviewer Sub-Agent
Use sessions_spawn with mode="run" and a cheaper model (e.g. qwen3.5-plus, gpt-4o-mini) to minimize cost. The reviewer does not need the full tool suite — it only reads text and outputs critique.
Note: This sends your draft text to the configured model provider's API. If you use cloud-based models, data will be sent to the provider's servers. Check your provider's data policy if handling sensitive content. For truly local review, configure a local model (e.g. ollama, LM Studio).
Spawn with this prompt template:
You are an adversarial reviewer. Examine the following draft response critically.Draft:
"""
[YOUR DRAFT HERE]
"""
Check for:
- Factual errors or unsubstantiated claims
- Missing important caveats or edge cases
- Logical contradictions
- Tone issues (too apologetic, too confident, dismissive, etc.)
- Missing follow-up suggestions that would be valuable
- Overly verbose sections that could be trimmed
- Any advice that could backfire or cause problems
Respond with exactly one of:
- PASS + one-line reason why it's fine
- List of specific issues, each with:
- Where the problem is
- Why it's a problem
- Suggested fixBe harsh. Be picky. Better to over-catch than to miss. You are the quality gate.
Step 3: Apply Feedback
- Sub-agent says PASS → deliver your draft as-is
- Sub-agent raises valid points → revise your draft, then deliver the improved version
- Sub-agent is clearly wrong → trust your own judgment, deliver your version
Step 4: Deliver
When the review leads to substantive changes, briefly note the improvement (e.g. "Review caught X, fixed Y"). For minor edits, no need to mention. Focus on delivering the best result.
Privacy & Safety
- The draft text is sent to a second AI model instance via the configured model API. If you use cloud-based models (e.g. qwen3.5-plus, gpt-4o-mini), this will send data to the provider's servers. For local-only review, use a local model provider (e.g. ollama, LM Studio).
- Only the draft text (not full conversation history) is shared with the reviewer.
- If the draft contains sensitive data (PII, credentials, etc.), the agent should skip the review step automatically.
- Review exchanges are not persisted beyond the current agent session.
Exceptions
These situations do NOT need review:
HEARTBEAT_OK- System-level acks (tool results, NO_REPLY)
- Purely mechanical confirmations with zero opinion content
Why This Matters
LLM outputs can contain subtle errors, missing context, or tone issues that are easy to miss from the creator's perspective. A second "pair of eyes" that is explicitly adversarial catches problems before they reach the user. This is the agent equivalent of code review.
Note: This review step adds latency and token usage per response.
Technical Details
No special configuration needed. To disable review, uninstall this skill.
by