黄金审稿 (Golden Manuscript Review)
Purpose
Provide a reusable "AI 网文主编" engine that takes a finished or in-progress manuscript and returns a diagnostic report + rewritten text. The engine does not invent new plot; it detects what repels readers (毒点), validates structural/emotional soundness, matches 爆款 benchmarks, and applies a master-polish rewrite protocol optimized for mobile vertical-screen reading.
This is the review/revise stage of the 黄金 writing toolkit, sitting after drafting and before publication:
拆书(deconstruct) → 人设(character) → 细纲(outline) → 文风(prose)
→ 章节(chapter) → 书名简介(title) → 审稿(review) ← this skill
When to use
- User says 审稿 / 改稿 / 润色 / 毒点检测 / 开篇优化 / 数据诊断 / 对标爆款 / 多平台转换.
- User pastes a draft, a chapter, or an opening and asks "这段哪里有问题" / "帮我改得更爽" / "适配番茄风".
- After running novel-golden-chapter, to polish the generated draft before publishing.
The Review Engine (四层逻辑)
Layer 1 — 三重动态检测引擎 (Three-Dynamic Detection)
Run three detectors in parallel over the manuscript:
- 主编思维模型 (Editor Mind-model): 20-year chief-editor intuition — judge 开篇切入、节奏、期待感、爽点密度 at a meta level.
- 万订作品数据库 (Bestseller DB): compare against patterns of 万订 / 高完读 works — is the structure on par?
- 毒点特征库 (Poison Feature Library): match text spans against the 37-dim 毒点雷达 (see
references/poison-dimensions.md).
Layer 2 — 三维重塑体系 (Three-Dimension Reconstruction)
Produce the actual fixes in three passes:
- 核心梗识别 (Core Hook ID): extract the 1–2 sentence central selling point. If absent or buried, flag it and surface it.
- 毒点排雷 (Poison Clearing): for every 毒点 hit in Layer 1, generate a targeted fix. Distinguish 明雷 (obvious, e.g. 送女/圣母) from 隐性病灶 (hidden, e.g. 爽点降级/情绪断层). Apply the 角色行为合理性推演: every character decision must close its 人设逻辑闭环 — if a fix would break 人设, choose the alternative that preserves it.
- 爆款元素灌注 (Hit Infusion): match 3–6 部 爆款对标书 from the platform hot list, then inject 微妙反差轴 (subtle contrast axis) and 信息差情节 (information-gap plot) where the manuscript is flat.
Layer 3 — 大神级精修协议 (Master Polish Protocol)
Rewrite the cleaned text:
- 牛马改文算法: simulate one of three 流派大神叙事风格 (see
references/review-protocol.mdfor the three style profiles) chosen to fit the manuscript's genre. - 手机阅读呼吸感: apply 动态段落切割 — target the 黄金 55% 单句成段率 for vertical-screen reading. Break long exposition; keep one beat per line at tension peaks.
- 反AI检测机制: keep 人性化叙事温度 — inject 口语节奏、具象细节、情绪留白; strip 模板化排比与空洞升华.
Layer 4 — 量化评估面板 (Quant Panels)
Emit four metrics so the author sees impact:
- 情绪曲线优化: track 200+ 读者情绪锚点 across the text; report the curve and flag 断层.
- 收益预判模型: simulate 章节完读率 / 首秀留存 from structural signals.
- 读者预期热力图: mark where 期待感 is high (hot) vs dead (cold).
- 章节级爽点浓度 / 人设冲突值: per-chapter numeric gauges.
Review Workflow (5 steps)
- 接入 (Intake): read the manuscript + any available 人设卡 (from novel-golden-character) + 细纲 (from novel-golden-outline) + 文风设定 (from novel-golden-prose). The more upstream artifacts present, the sharper the diagnosis.
- 扫描 (Scan): run Layer 1 — three detectors + 37-dim radar. Produce a 毒点清单 with severity (明雷 / 隐性 / 建议).
- 重塑 (Reconstruct): run Layer 2 — core-hook ID, poison fixes with 角色逻辑校验, hit infusion via benchmark match.
- 精修 (Polish): run Layer 3 — pick 大神风格, cut paragraphs to 55% rate, de-AI.
- 交付 (Deliver): emit the review report from
assets/review-template.md: 毒点清单表 + 修改建议 + 情绪曲线 + 量化面板 + 精修后全文。
Integration with the 黄金 toolkit
| Upstream artifact | How review consumes it |
|---|---|
| 人设 (novel-golden-character) | 12-field cards + 关系图谱 → 角色行为合理性推演 source of truth |
| 细纲 (novel-golden-outline) | 12-step loop + 三程度 → validate 期待感 / 章末钩 / 爽点升级 |
| 文风 (novel-golden-prose) | 9-dim framework → 大神精修 aligns to chosen 氛围/细腻/组合 mode |
| 章节 (novel-golden-chapter) | generated draft → primary input to review |
| 拆书 (novel-golden-deconstruct) | 28-dim + 情绪公式 → benchmark matching & 情绪曲线 baseline |
Applicable scenarios (触发映射)
- 新书开篇生死 2000 字优化 → focus Layer 2.1 + 3.
- 连载中期数据滑坡诊断 → Layer 1 毒点雷达 + Layer 4 量化面板.
- 完本作品爆款元素萃取 → Layer 2.3 爆款元素灌注.
- 多平台内容适应性转换 → Layer 3 大神风格切换 + 55% 段落率重切.
- 人设迭代风险预演 → Layer 2.2 角色行为合理性推演.
See references/poison-dimensions.md for the full 37-dim 毒点 radar, and references/review-protocol.md for the workflow detail, three 大神 style profiles, paragraph-cutting rules, and quant-panel formulas.
