关于 cw-debate
cw-debate 组织多个审稿 Agent 对草稿进行结构化讨论,适用于影响较大的写作决策。审稿者会相互质疑,并尝试达成共识、提出折中方案或确认僵局。
主要功能
- 01并行收集审稿者的初步分析
- 02围绕分歧生成质询
- 03整理回应与折中方案
- 04标记已解决的问题和僵局
- 05讨论最多进行四轮
适用场景
- 01处理相互冲突的草稿反馈
- 02检验审稿建议经受质疑的程度
- 03为重要写作决策探索折中方案
安装
复制下面的命令,在终端中运行。
npx skills add https:/ / github.com/ everyinc/ compound-writing --skill cw-debate安装后在 Claude Code 中输入/cw-debate确认已加载。
详细介绍
辩论评审
组织评审 Agent 之间进行结构化审议。评审者并非只是并行给出反馈,而是会交流彼此的观点、质疑对方的建议,并努力达成共识(或承认僵局)。
何时使用 /cw-debate,何时使用 /cw-panel
| 使用 | 适用场景 |
|---|---|
/cw-panel | 你希望综合多方观点。速度快。我们会指出分歧,由你决定。 |
/cw-debate | 你希望评审者真正辩论一番。更全面。分歧可能通过审议得到解决。 |
在以下情况下选择 /cw-debate:
- 作品影响重大,值得投入额外轮次
- 你想看看各种观点经受质疑后是否站得住脚
- 你怀疑有些分歧若让评审者彼此交流,或许能够化解
- 你想要提案和折中方案,而不是只听一句“由你决定”
在以下情况下选择 /cw-panel:
- 你想快速获得全面反馈
- 你愿意自行解决分歧
- 你需要考虑时间或 token 限制
审议流程
第 1 轮:初始立场
与 /cw-panel 相同——所有评审者独立分析草稿。
第 2 轮:质疑
主持人找出分歧,并向相关评审者提出质疑:
## Challenge: The fishing gear paragraph
**The passage:**
> "His tackle box was a museum of failures—rusted lures from the '70s..."
**hemingway said:**
> "47 words that could be 15. Cut it."
**sedaris said:**
> "This is the most alive moment in the piece. Keep every word."
---
**hemingway:** sedaris argues the specificity earns its place. How do you respond?
**sedaris:** hemingway argues this slows momentum. How do you respond?
第 3 轮:回应
评审者以其角色身份回应。他们可以:
- 让步——“仔细想想,这处细节确实值得保留。我收回之前的意见。”
- 坚持——“节奏上的问题依然存在。即使是好的细节,在这里也会造成负担。”
- 提议——“保留第一幅画面,删去较长的列表。这样双方都能接受。”
第 4 轮:达成结论
针对每项分歧:
- 已解决——评审者达成一致(通过让步或提出方案)
- 陷入僵局——评审者坚持各自立场,价值观存在根本差异
- 已有提案——已提出折中方案,由作者决定
如果提案引发新的讨论,主持人可以继续进行额外轮次,但总轮次最多为 4 轮。
实际操作流程
1. 加载上下文
读取 ../../references/context-contract.md、草稿,以及任何可用的简述、大纲、项目说明、反馈或状态元数据。确定:
- piece_type
- audience
- stage
- goals
根据草稿和请求推断缺失的上下文。无需提供 .status.yaml。
2. 提议评审小组
与 /cw-panel 相同——根据上下文提议 4-6 位评审者。
## Proposed Panel for Debate
**Draft:** [title]
**Context:** [piece_type] for [audience], [stage]
| Reviewer | Why |
|----------|-----|
| **[name]** | [Rationale] |
| **[name]** | [Rationale] |
**Note:** Debate works best with reviewers who have natural tensions.
Consider including pairs like:
- hemingway + sedaris (economy vs. specificity)
- reader + hitchcock (orientation vs. mystery)
- sorkin + vonnegut (speed vs. depth)
Proceed with this panel, or adjust?
等待用户确认。
3. 第 1 轮:运行评审者
如果当前平台支持隔离的工作进程,且用户明确提出的辩论请求授权使用它们,则并行运行所有评审者。在 Claude 上,有匹配的评审者 Agent 时使用它们;在 Codex 或其他运行时上,使用相应视角的评审技能或隔离的工作进程。收集各自独立的输出。
4. 第 2 轮:找出分歧并提出质疑
分析输出,找出冲突。针对每项重要分歧:
- 引用相关段落
- 呈现双方立场
- 生成质疑提示词
将质疑背景提供给受到质疑的评审者,并启动他们。
5. 第 3 轮:收集回应
每位评审者针对质疑作出回应:
- 让步、坚持或提议
- 说明理由
6. 第 4 轮:达成结论并进行额外轮次
确定结论状态。如果提案值得进一步讨论,则再进行一轮(总计最多 4 轮)。
7. 最终综合
返回辩论主持人的综合结果:
- 已解决的分歧(以及解决方式)
- 僵局(包括最终立场和影响)
- 提案(折中选项)
- 共识结论(未进入辩论的内容)
- 关键问题
辩论评审组的选择
当评审组中存在自然的张力时,辩论最有价值:
| 配对 | 张力类型 | 适用场景 |
|---|---|---|
| hemingway + sedaris | 简洁 vs. 具体 | 判断哪些内容值得占用篇幅 |
| reader + hitchcock | 清晰引导 vs. 悬念 | 区分有意保留信息与意外造成的困惑 |
| reader + hemingway | 背景 vs. 简洁 | 判断首次阅读的读者需要哪些铺垫 |
| reader + sorkin | 清晰引导 vs. 推进节奏 | 让读者了解情况,同时不拖慢文章节奏 |
| mom + hemingway | 通俗易懂 vs. 简洁 | 判断需要向非专业读者解释哪些内容 |
| sorkin + sedaris | 推进节奏 vs. 观察 | 描写性段落的节奏 |
| sorkin + vonnegut | 速度 vs. 深度 | 何时放慢节奏来刻画角色 |
| nemesis + sedaris | 严谨 vs. 风格 | 论说文的语气 |
**建议:**评审组中至少加入一组自然的张力配对。
示例会话
User: /cw-debate
Agent: I'll run a deliberation on your draft. Let me read it and propose a panel.
[Reads draft]
[Loads the brief and project context: personal essay, general audience, revision stage]
## Proposed Panel for Debate
**Draft:** father-and-the-sea
**Context:** personal essay for general audience, revision stage
| Reviewer | Why |
|----------|-----|
| **hemingway** | Check economy—personal essays can get overwritten |
| **sedaris** | Find the specific, human, funny moments |
| **sorkin** | Verify pacing through emotional beats |
| **reader** | Trace the first-time reading experience |
**Natural tensions:** hemingway/sedaris (economy vs. specificity),
sorkin/sedaris (momentum vs. observation)
Proceed with this panel, or adjust?
User: yes
Agent: Running Round 1—each reviewer analyzing independently.
[Spawns 4 reviewers in parallel]
[Collects outputs]
Round 1 complete. Analyzing for tensions.
## Tensions Identified
### Tension 1: The tackle box paragraph
**hemingway** wants to cut; **sedaris** wants to keep every word.
### Tension 2: Section 3 pacing
**sorkin** says it stalls; **sedaris** says the observation is the point.
### Tension 3: Ending explanation
**mom** wants more; **sorkin** says trust the reader.
Proceeding to Round 2 with 3 challenges.
[Spawns challenged reviewers with challenge context]
[Collects responses]
## Round 2 Responses
**Tension 1:**
- hemingway: HOLD — "Momentum matters more than any single detail."
- sedaris: PROPOSE — "Keep 'museum of failures' and the pug. Cut the Reagan line."
**Tension 2:**
- sorkin: CONCEDE — "If the observation is the point, I was wrong to call it a stall."
- sedaris: (no concession needed)
**Tension 3:**
- mom: HOLD — "I still don't know what he learned. That's a real gap."
- sorkin: PROPOSE — "One specific image of the lesson. Not explanation—scene."
Proceeding to final synthesis.
## Debate Synthesis
**Panel:** hemingway, sedaris, sorkin, mom
**Rounds:** 2
**Context:** personal essay for general audience, revision stage
---
### Resolved Tensions
**Section 3 pacing**
> [The observation passage]
**Resolution:** Sorkin conceded. The slow pace is intentional and earns its place.
**Recommendation:** Keep as-is. The observation is the point.
---
### Stalemates
*None—all tensions reached resolution or proposal.*
---
### Proposals
**The tackle box paragraph**
> "His tackle box was a museum of failures—rusted lures from the '70s,
> a reel that hadn't worked since Reagan, hooks that had never caught
> anything but his own thumb."
**Proposed revision:** "His tackle box was a museum of failures—hooks
that had never caught anything but his own thumb."
**Proposed by:** sedaris (compromise with hemingway)
**Would satisfy:** Keeps the voice and "museum" metaphor, cuts the catalog
**Trade-off:** Loses the temporal sweep ("'70s," "Reagan")
---
**The ending**
**Proposed revision:** Add one concrete image showing what you learned,
rather than explaining it.
**Proposed by:** sorkin (compromise with mom)
**Would satisfy:** Mom gets clarity; sorkin keeps momentum
**Trade-off:** Requires writing new material
---
### Consensus Findings
| Issue | Flagged By | Recommendation |
|-------|------------|----------------|
| Opening is slow | hemingway, sorkin | Start on the boat |
| "He taught me everything" unearned | all | Show, don't declare |
---
### The Deliberation Summary
**Resolved:** 1 tension resolved through concession
**Stalemates:** 0
**Proposals:** 2 compromise options for your consideration
---
### The Hard Question
> The debate kept circling one thing: you're telling us the relationship
> mattered without showing us what made it irreplaceable. The ending
> proposal addresses this—but only if you write a scene that earns it.
成本与范围
与评审组审阅相比,辩论明显更慢,也更消耗上下文,因为评审者会收到后续质询。请将其用于影响重大的分歧,限制轮数;对于常规审阅,优先使用 cw-panel。
错误处理
如果无法启动某位评审者:
- 记录这一缺口
- 继续使用可用的评审者
- 涉及该评审者的分歧将不会进入辩论
如果未发现任何分歧:
- 跳到综合总结
- 报告“没有显著分歧——评审者的意见基本一致”
如果评审者未回应质询:
- 视为默认 HOLD
- 在综合总结中注明未回应
如果辩论陷入循环:
- 立场重复后,由主持人宣布陷入僵局
- 无论如何,轮数上限为 4 轮
与 /cw-panel 的关系
/cw-panel 和 /cw-debate 共用基础设施:
- 相同的评审者 Agent
- 从当前项目和产物加载相同的上下文
- 相同的评审组提案逻辑
两者的区别:
/cw-panel→ 综合总结者(单轮审阅,指出分歧)/cw-debate→ 主持人(多轮辩论,围绕分歧展开讨论)
一种自然的工作流:先运行 /cw-panel 获取快速反馈,然后针对特定分歧运行 /cw-debate,展开辩论。
经验总结
[此技能的经验总结将在记录时添加]