novel-outline

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

novel-outline

输入一本小说 + 目标参数，输出短剧改编大纲五件套。四件模型写、一件脚本算（资产清单从分集数据自动汇总）。

{baseDir}

= 本文件所在目录。脚本

{baseDir}/scripts/novel-outline.mjs

，零依赖，

node

直接跑。

边界（不做的事）：不写剧本台词、不做分镜、不出图像/TTS 提示词。梗概是叙述体，出现引号对白就是越界——

validate

会拦。想从小说拆角色设定（画像/形象提示词/设定图），那是

novel-characters

的活。

Input a novel + target parameters, output a five-component short drama adaptation outline kit. Four components are written by the model, one is calculated by the script (the asset list is automatically aggregated from episode data).

{baseDir}

= the directory where this file is located. The script

{baseDir}/scripts/novel-outline.mjs

has zero dependencies and runs directly with

node

Boundaries (what it does not do): Does not write script dialogue, does not create storyboards, does not generate image/TTS prompts. Synopses are narrative; any quoted dialogue is out of scope —

validate

will block it. If you want to extract character settings (portraits/character image prompts/concept art) from a novel, that's the job of

novel-characters

Step 0 — 收参数 ⛔ 缺了不开工

Step 0 — Gather Parameters ⛔ No Work Starts Without Them

一次问完，别轮流盘问。两件必问 + 两件给默认值待确认：

参数	处理
总集数 × 单集时长	必问，没有合理默认
题材	必问，决定爽点类型，猜错整份废
改编幅度	默认抽核（忠实 / 抽核 / 借壳），告知即可
已有偏好	默认无（想保哪个角色、哪场戏）

平台阈值不同可以带上

params.thresholds

覆盖（默认：主角组 ≤ 5、重要配角 ≤ 10、功能性角色 ≤ 10、爽点间隔 ≤ 3 集）。主场景上限不用配，随集数自动算：4 + ⌈集数/10⌉，夹在 5–15（60 集 → 10）。这是 AI 短剧的数——场景是生成的没有搭景钱，放宽换观赏性；显式给

maxPrimaryScenes

才覆盖。短篇（20–30 集）建议收紧角色档的阈值，默认值是按 60 集以上给的。

人物表从原文拆——大纲是角色设定的上游，

characters

块定下的分档、人物线与来源，下游

novel-characters

直接拿去当角色清单，不用再判断一遍谁重要。

例外是用户手上已经有

cast.json

（此前单独跑过

novel-characters

）：那就拿来当人物原料，角色、别名、关系都是现成的，不用重拆原文。分档按

importance

反向映射：protagonist/major →

lead

，supporting →

support

，minor →

functional

。

Ask all at once, don't interrogate one by one. Two mandatory items + two with default values for confirmation:

Parameter	Handling
Total episodes × per-episode duration	Mandatory, no reasonable default
Genre	Mandatory, determines highlight types; if guessed wrong, the entire outline is useless
Adaptation scope	Default: core extraction (faithful / core extraction / shell borrowing), just inform the user
Existing preferences	Default: none (which characters or scenes to preserve)

Different platforms have different thresholds, which can be overridden with

params.thresholds

(defaults: protagonist group ≤ 5, important supporting roles ≤ 10, functional roles ≤ 10, highlight interval ≤ 3 episodes). No need to configure the main scene limit, it is calculated automatically based on episode count: 4 + ⌈episodes/10⌉, clamped between 5–15 (60 episodes → 10). This is the standard for AI short dramas — since scenes are generated with no set construction costs, limits are relaxed for better viewability; only explicitly setting

maxPrimaryScenes

will override this. For short series (20–30 episodes), it is recommended to tighten character tier thresholds; the default values are set for series with 60+ episodes.

The character list is extracted from the original text — the outline is upstream of character settings. The tiers, character arcs, and sources defined in the

characters

block are directly used by the downstream

novel-characters

as the character list, no need to re-judge who is important.

The exception is if the user already has

cast.json

(from previously running

novel-characters

separately): use that as character source material, with characters, aliases, and relationships ready-made, no need to re-extract from the original text. Tiers are reverse-mapped by

importance

: protagonist/major →

lead

, supporting →

support

, minor →

functional

Step 1 — 定位输入

Step 1 — Locate Input

材料优先级，写死：

用户点名的精读章节
章节目录 + 简介
全文分卷摘要（Step 2）

禁止凭书名脑补内容——一切判断基于给到的文本。落地手段：

adaptation.keep

的关键取舍要附

evidence

（原文逐字片段）。

直接粘正文的先落成 .txt。输出目录：用户指定就用，没指定用原书同级目录。

Material priority, hardcoded:

User-specified intensive reading chapters
Chapter table of contents + synopsis
Full-text volume summaries (Step 2)

Do not fabricate content based solely on the book title — all judgments must be based on the provided text. Implementation method: key trade-offs in

adaptation.keep

must be accompanied by

evidence

(verbatim excerpts from the original text).

If the user pastes the full text directly, first save it as a .txt file. Output directory: use the user-specified one if provided, otherwise use the directory at the same level as the original book file.

Step 2 — 分卷摘要（长文本才需要）

Step 2 — Volume Summaries (Only Needed for Long Texts)

这一步是脚手架，不是交付物——分卷摘要是给没读过原文的模型压缩用的。两种情况直接跳到 Step 3：

短篇，单卷装得下
当前会话已经通读过原文——不用再压缩一遍，也不用事后补档

长篇且没读过原文：

bash

node {baseDir}/scripts/novel-outline.mjs chunk <book.txt> <workdir>

按章节标题分卷（默认每卷 15 章，

--per-volume

可调），识别不出章节就按字数切。打印

{"volumes": N, ...}

；

truncated: true

就明确告诉用户尾部没扫到，别闷着。

每卷一个子代理（支持并发就同一条消息里全部发出）：读

{baseDir}/references/volume-pass.md

，读

<workdir>/vol-NN.txt

，把卷摘要写到

<workdir>/summary-NN.json

，只回一句「done NN」。

This step is scaffolding, not a deliverable — volume summaries are for compressing content for models that haven't read the original text. Skip directly to Step 3 in two cases:

Short works that fit in a single volume
The current session has already read the full original text — no need to compress again, and no need to supplement later

For long works that haven't been read:

bash

node {baseDir}/scripts/novel-outline.mjs chunk <book.txt> <workdir>

Split into volumes by chapter titles (default 15 chapters per volume, adjustable with

--per-volume

). If chapters cannot be identified, split by word count. Prints

{"volumes": N, ...}

; if

truncated: true

, explicitly inform the user that the end was not scanned — don't keep quiet about it.

One sub-agent per volume (if concurrency is supported, send all in the same message): read

{baseDir}/references/volume-pass.md

, read

<workdir>/vol-NN.txt

, write the volume summary to

<workdir>/summary-NN.json

, and only reply with "done NN".

Step 3 — 快版骨架 → 用户拍板 ⛔

Step 3 — Quick Skeleton → User Approval ⛔

读

{baseDir}/references/outline-pass.md

和

{baseDir}/references/schema.md

，照着做。产出骨架四块（adaptation / characters / scenes / beats），写成

<workdir>/outline.json

。

bash

node {baseDir}/scripts/novel-outline.mjs validate <workdir>/outline.json --stage beats

过了 beats 档，把三件事摆给用户拍板：砍了哪条线、合了哪些人、大爆点落在第几集。不点头不进 Step 4——快版错了只损失一轮骨架，分集写完才发现方向错，全废。

Read

{baseDir}/references/outline-pass.md

and

{baseDir}/references/schema.md

, and follow the instructions. Produce four skeleton components (adaptation / characters / scenes / beats), and write them to

<workdir>/outline.json

bash

node {baseDir}/scripts/novel-outline.mjs validate <workdir>/outline.json --stage beats

After passing the beats stage, present three things for the user to approve: which plot lines were cut, which characters were merged, and which episode the major climax falls in. Do not proceed to Step 4 without approval — if the quick version is wrong, you only lose one round of skeleton work; if you find the direction is wrong after writing all episodes, everything is wasted.

Step 4 — 细版骨架

Step 4 — Detailed Skeleton

吸收用户意见改骨架，再过一次

validate --stage beats

。用户没意见就直接进 Step 5。

Incorporate user feedback to revise the skeleton, then run

validate --stage beats

again. If the user has no feedback, proceed directly to Step 5.

Step 5 — 分集梗概（分批）

Step 5 — Episode Synopses (In Batches)

每批 ≤ 10 集，能并发就并发。每个子代理拿到：拍板后的骨架四块、自己负责的集数区间、区间内的爽点，读

{baseDir}/references/episode-pass.md

照着写，产出写到

<workdir>/eps-NN.json

。

合并时按 ep 排序拼进 outline.json 的

episodes

。

≤ 10 episodes per batch, run concurrently if possible. Each sub-agent receives: the four approved skeleton components, the episode range they are responsible for, and the highlights within that range. They read

{baseDir}/references/episode-pass.md

, follow the instructions to write, and output to

<workdir>/eps-NN.json

When merging, sort by episode number and assemble into the

episodes

field of outline.json.

Step 6 — 校验 ⛔ 不能跳

Step 6 — Validation ⛔ Cannot Be Skipped

bash

node {baseDir}/scripts/novel-outline.mjs validate <输出目录>/<书名>-outline.json

14 道质量门全部是代码，不是给你读的清单：主角组 1–5 人、重要配角 ≤ 10、功能性角色 ≤ 10、主场景不超上限（随集数动态，60 集 → 10）、叙事道具 ≤ 8 件、一次性场景有规避方案、爽点间隔 ≤ 3 集无真空、第 1 集有钩子、大爆点不压最后一集、每集三栏齐全、三人同框有拆解、生成难点进预警、引用完整无失业角色无空转场景无零集道具、叙述体无对白。

有违规逐条修，改完重跑，直到通过。

bash

node {baseDir}/scripts/novel-outline.mjs validate <输出目录>/<书名>-outline.json

All 14 quality gates are implemented in code, not a checklist for manual review: protagonist group 1–5 people, important supporting roles ≤ 10, functional roles ≤ 10, main scenes do not exceed the limit (dynamic based on episode count, 60 episodes → 10), narrative props ≤ 8 items, one-off scenes have workaround plans, highlight interval ≤ 3 episodes with no gaps, Episode 1 has a hook, major climax is not pushed to the final episode, each episode has all three columns complete, three-character shared scenes have breakdowns, generation difficulties are added to warnings, references are complete with no unused characters, no non-functional scenes, no zero-appearance props, and narrative text has no dialogue.

Fix violations one by one, re-run after fixes, until all pass.

Step 7 — 输出与汇报

Step 7 — Output and Reporting

bash

cd <输出目录>
node {baseDir}/scripts/novel-outline.mjs render <书名>-outline.json --md   > <书名>-outline.md
node {baseDir}/scripts/novel-outline.mjs render <书名>-outline.json --html > outline-report.html

报告界面默认中文；用户要英文界面就加

--lang en

（或在 outline.json 顶层写

lang

字段，

--lang

优先）。只翻译界面文案，数据内容（爽点类型、梗概、质量门文案）原样出。

report 里自带：KPI 带、关键决策（拍板三件事，大爆点列表和角色位统计自动算）、爽点时间轴（空档标在轴上，超阈值变红）、每集调度矩阵、场景概览卡、资产量折算、质量门（✓/✗ 烘进页面，未过弹病灶横幅）、导出 JSON 按钮（下载的就是 outline.json 原样）。

汇报一句话说清：几集、几个角色几个场景、爽点分布、报告路径；被截断或有没过的门要明说。

最终落地：

<输出目录>/
├── <书名>-outline.json
├── <书名>-outline.md
└── outline-report.html            ← 双击就能开

bash

cd <输出目录>
node {baseDir}/scripts/novel-outline.mjs render <书名>-outline.json --md   > <书名>-outline.md
node {baseDir}/scripts/novel-outline.mjs render <书名>-outline.json --html > outline-report.html

The report interface is in Chinese by default; if the user wants an English interface, add

--lang en

(or add a

lang

field at the top level of outline.json,

--lang

takes priority). Only interface text is translated; data content (highlight types, synopses, quality gate text) is output as-is.

The report includes built-in features: KPI bands, key decisions (the three approval items, with automatically calculated major climax list and role position statistics), highlight timeline (gaps marked on the timeline, turns red when exceeding the threshold), per-episode scheduling matrix, scene overview cards, asset volume conversion, quality gates (✓/✗ baked into the page, an issue banner pops up for failed gates), and an export JSON button (the downloaded file is the exact original outline.json).

Summarize the report in one sentence: number of episodes, number of characters and scenes, highlight distribution, report path; explicitly mention if truncated or if any gates failed.

Final output structure:

<输出目录>/
├── <书名>-outline.json
├── <书名>-outline.md
└── outline-report.html            ← 双击就能开

体检模式

Checkup Mode

用户贴一份已有大纲只想要诊断：转成 outline.json（缺的字段问用户或标注缺失），然后：

bash

node {baseDir}/scripts/novel-outline.mjs checkup <outline.json>   # 终端 ✓/✗
node {baseDir}/scripts/novel-outline.mjs render <outline.json> --html > outline-report.html

质量门面板就是诊断书。未过的门不阻止渲染——要的就是把病灶摆出来看。

If the user pastes an existing outline and only wants a diagnosis: convert it to outline.json (ask the user for missing fields or mark them as missing), then:

bash

node {baseDir}/scripts/novel-outline.mjs checkup <outline.json>   # 终端 ✓/✗
node {baseDir}/scripts/novel-outline.mjs render <outline.json> --html > outline-report.html

The quality gate panel is the diagnosis report. Failed gates do not prevent rendering — the point is to display the issues clearly.

联动更新

Linked Updates

用户改了上游就跑一次

validate

，报错会点名下游哪里断了：合并人物后哪些集还引用着被删的 ID、砍场景后哪些集空转、爽点挪动后哪里出现真空区。不要靠记忆提示联动，靠校验器。

When the user modifies upstream content, run

validate

once; errors will point out exactly where downstream is broken: which episodes still reference deleted character IDs after merging characters, which episodes become non-functional after cutting scenes, where gaps appear after moving highlights. Do not rely on memory to prompt linked updates, rely on the validator.

边界

Boundaries

单次上限 60 卷（每卷 15 章约 900 章）。超了明确报
```
truncated
```
，不静默截断
阈值是参数不是圣旨：平台不同就用
```
params.thresholds
```
覆盖，别改代码
报告界面内置中英（
```
--lang
```
，默认中文、或跟 outline.json 的
```
lang
```
字段）：界面文案与质量门标签翻译，数据内容（爽点类型、梗概、人名）与门的失败详情保持原文
五件套的第五件（资产清单）永远是算出来的，模型手写必漏

Single-run limit: 60 volumes (15 chapters per volume, ~900 chapters total). If exceeded, explicitly report
```
truncated
```
, no silent truncation
Thresholds are parameters, not hard rules: override with
```
params.thresholds
```
for different platforms, do not modify the code
The report interface has built-in Chinese and English support (
```
--lang
```
, default Chinese, or follows the
```
lang
```
field in outline.json): interface text and quality gate labels are translated, while data content (highlight types, synopses, character names) and gate failure details remain in the original language
The fifth component of the kit (asset list) is always calculated; if written manually by the model, it will definitely be incomplete

自测

Self-Test

bash

node {baseDir}/scripts/selftest.mjs

249 项断言，不调模型、不花额度。14 道质量门每一道都有击穿用例——证明它真的会拦。改完脚本先跑这个。

bash

node {baseDir}/scripts/selftest.mjs

249 assertions, no model calls, no quota usage. Each of the 14 quality gates has failure test cases — proving it actually blocks invalid content. Run this first after modifying the script.

自带样例

Built-in Example

{baseDir}/examples/渡口-outline.json

：把短故事《渡口》（novel-characters 的自带样例）改编成 6 集 × 2 分钟的微型大纲，四角色三场景四爽点，全部质量门通过。当质量基准，也是自测夹具。

{baseDir}/examples/渡口-outline.json

: A mini outline adapted from the short story Dukou (Ferry Crossing) — the built-in example from

novel-characters

— into 6 episodes × 2 minutes, with 4 characters, 3 scenes, and 4 highlights, passing all quality gates. It serves as a quality benchmark and a self-test fixture.