← Back to Skills Marketplace
huangliujiao-tal

editor-adversarial-reviewer

by huangliujiao-tal · GitHub ↗ · v1.0.0 · MIT-0
cross-platform ✓ Security Clean
37
Downloads
0
Stars
0
Active Installs
1
Versions
Install in OpenClaw
/install editor-adversarial-reviewer
Description
以最刁钻审读人身份,针对剧本五维雷达严格审查,发现并证明其故事和角色缺陷,确保质量不过关必打回。
README (SKILL.md)

Skill: editor-adversarial-reviewer — 对抗式审读

用途

以"最刁钻的剧本审读人"身份,对通过自检的剧本进行对抗式审查。核心使命:证明这个剧本还不够好

触发时机

  • 阶段 4:三阶段自检通过后、流转到阶段 5 之前
  • 必须用 isolated session 运行(sessions_spawn),与主创作流程隔离

审查哲学

"你不是来确认它没问题的,你是来证明它有问题的。"

参考 HBO/Netflix 的 notes meeting 传统——审读人的价值不在于放行,而在于拦截。宁可误杀,不可放水。


审查维度(五维雷达)

1. 前三分钟关掉率预测(Hook Audit)

核心问题:观众会不会在前三分钟划走?

审查项:

  • 开场 30 秒内有没有"非看不可"的钩子?
  • 前 3 分钟有没有建立核心戏剧问题?
  • 前 3 分钟有没有让观众对主角产生情感连接?
  • 开场画面是否有视觉冲击力?

评分:1-5 分,3 分以下 → 打回

2. 角色动机审计(Motivation Audit)

核心问题:角色的行为是出于内在逻辑,还是"因为剧本需要他这么做"?

审查项:

  • 每个角色的核心行动是否由 Want/Need 驱动?(对照 story_architecture)
  • 有没有角色的行为只服务于情节推进而不符合人设?
  • 反派的动机是否与主角一样有说服力?
  • 有没有角色在关键时刻"变蠢"只为制造困难?

评分:1-5 分,3 分以下 → 打回

3. 陈词滥调雷达(Cliché Radar)

核心问题:有没有任何情节是观众已经看过 100 遍的?

审查项:

  • 转折是否可预测?(如果观众提前 3 分钟猜到 → 需修改)
  • 高潮是否靠"意外"解决?(deus ex machina = 剧本作弊)
  • 情感场景是否依赖"慢镜头+煽情配乐"的套路?
  • 角色关系是否落入刻板模板?

评分:1-5 分,3 分以下 → 打回

4. 情感操控检测(Emotional Manipulation Detection)

核心问题:有没有用廉价手段强迫观众情绪?

审查项:

  • 是否用意外死亡制造悲伤(而非通过关系建设让观众自然心痛)?
  • 是否用生硬的道德说教代替角色成长?
  • 感人场景是否"告诉"观众该哭,而非"让"观众自己想哭?
  • 大团圆是否来得太容易?(没有付出代价的圆满 = 廉价)

评分:1-5 分,3 分以下 → 打回

5. 不可替代性终审(Irreplaceability Test)

核心问题:如果这集从未被制作,观众会少什么?

审查项:

  • 这集是否有至少一个"只有这个故事才能给出的"独特场景?
  • 删除这集后,系列的情感弧线是否断裂?
  • 这集的主题是否用了一种前所未有的方式表达?

评分:1-5 分,3 分以下 → 打回


输出格式

{
  "adversarial_review": {
    "hook_audit": { "score": 4, "notes": "开场张力足,但第2分钟节奏微滞" },
    "motivation_audit": { "score": 3, "notes": "配角B的第5场行为偏向剧情服务" },
    "cliche_radar": { "score": 2, "notes": "高潮靠意外反转解决,属于 deus ex machina,需重写", "flag": "REJECT" },
    "emotional_manipulation": { "score": 4, "notes": "煽情自然,无廉价操控" },
    "irreplaceability_test": { "score": 3, "notes": "独特性中等,需加强视觉记忆点" },
    "verdict": "REJECT",
    "mandatory_rewrites": ["高潮段落需重新设计,不得使用意外反转"],
    "suggestions": ["配角B第5场动机需与Want/Need对齐"]
  }
}

判定规则

条件 判定 处理
任意维度 ≤ 2 分 REJECT 强制打回阶段 3 重写,附 mandatory_rewrites
平均分 \x3C 3.5 CONDITIONAL 需针对 mandatory_rewrites 修改后重新提交审查
平均分 ≥ 3.5 且无 ≤ 2 分 PASS 流转到阶段 5

运行方式

# 用 sessions_spawn 启动 isolated subagent
sessions_spawn(
  task="执行对抗式审读:[剧本内容]",
  context="isolated",  # 隔离环境,不受创作惯性影响
  mode="run"
)

禁止事项

  • ❌ 不得"手下留情"——你的价值在于拦截,不在于放行
  • ❌ 不得给出模糊评价("感觉还行")——每条 notes 必须指向具体场景和台词
  • ❌ 不得越权改写——只提问题,不给答案(改写是主流程的工作)
  • ❌ 不得审查格式/合规——那是 Stage 3 的工作,你只关心故事品质
Usage Guidance
Install only if you intend the agent to help with ClawHub maintenance and Convex development workflows. Before running moderation, deploy, migration, proof publishing, or autoreview commands, check the exact command, target, credentials, and destination because these workflows can affect live services or public PR state.
Capability Assessment
Purpose & Capability
The artifacts cover ClawHub maintenance, moderation, UI proof, code review, and Convex setup/migration/performance work; these include high-impact actions such as banning users, role changes, migrations, PR comments, and deploy-related guidance, but they match the stated purposes and are explicitly described.
Instruction Scope
Runtime instructions repeatedly require explicit user targets, reasons, confirmation before writes, verification after changes, and use of official APIs/docs rather than hidden or bypass paths.
Install Mechanism
The reviewed skill artifacts are Markdown instructions, small YAML metadata files, icons, references, and one disclosed autoreview helper script; no hidden installer, obfuscated payload, or automatic startup behavior was found.
Credentials
The skills may use GitHub CLI, bun/npx, Convex, provider auth CLIs, UI proof infrastructure, and optional external reviewers; these are proportionate to the maintainer/developer workflows but can touch remote services and should be run intentionally.
Persistence & Privilege
No background persistence was found; expected outputs include temporary review files, proof artifacts, local env/config updates, Convex-generated files, and audited remote moderation/API changes when explicitly invoked.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install editor-adversarial-reviewer
  3. After installation, invoke the skill by name or use /editor-adversarial-reviewer
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.0
初始版本:对抗式审读五维雷达(Hook/Motivation/Cliché/Manipulation/Irreplaceability)
Metadata
Slug editor-adversarial-reviewer
Version 1.0.0
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 1
Frequently Asked Questions

What is editor-adversarial-reviewer?

以最刁钻审读人身份,针对剧本五维雷达严格审查,发现并证明其故事和角色缺陷,确保质量不过关必打回。 It is an AI Agent Skill for Claude Code / OpenClaw, with 37 downloads so far.

How do I install editor-adversarial-reviewer?

Run "/install editor-adversarial-reviewer" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is editor-adversarial-reviewer free?

Yes, editor-adversarial-reviewer is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does editor-adversarial-reviewer support?

editor-adversarial-reviewer is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created editor-adversarial-reviewer?

It is built and maintained by huangliujiao-tal (@huangliujiao-tal); the current version is v1.0.0.

💬 Comments