InkOS 多智能体小说创作 CLI 工具：AI 自主写作、审核与修订完整流程

inkos-multi-agent-novel-writing by aradotso/trending-skills

494 周安装量

10 GitHub Stars

GitHub

安装命令

npx skills add https://github.com/aradotso/trending-skills --skill inkos-multi-agent-novel-writing

AI/机器学习内容创作自动化自然语言处理

🇨🇳中文介绍

InkOS 多智能体小说创作

Skill by ara.so — Daily 2026 Skills collection

InkOS 是一个多智能体 CLI 系统，能够自主创作、审核和修订小说。智能体处理完整的流水线：写手 → 验证器 → 审核器 → 修订器，并在可配置的检查点设置人工审核关卡。

安装

npm install -g @actalk/inkos
# 或直接运行
npx @actalk/inkos --version

要求： Node.js ≥ 20.0.0

快速开始

# 创建一个新的小说项目
inkos book create --title "吞天魔帝" --genre xuanhuan

# 撰写下一章
inkos write next 吞天魔帝

# 审核特定章节
inkos audit 吞天魔帝 --chapter 3

# 运行完整守护进程（持续生产）
inkos daemon start

项目结构

执行 inkos book create 后，项目目录包含：

story/
  outline.md              # 故事大纲（架构师智能体输入）
  book_rules.md           # 每本书的自定义规则和审核维度
  chapter_summaries.md    # 自动生成的每章摘要
  subplot_board.md        # 支线进度跟踪（A/B/C 线）
  emotional_arcs.md       # 角色情感弧线跟踪
  character_matrix.md     # 角色互动矩阵 + 信息边界
  parent_canon.md         # 仅限衍生作品：导入的正传约束
  style_profile.json      # 风格指纹（如果使用了风格导入）
  style_guide.md          # LLM 生成的定性风格指南
  chapters/
    ch001.md
    ch002.md
    ...

广告位招租

在这里展示您的产品或服务

触达数万 AI 开发者，精准高效

联系我们

衍生作品（前传/续集/IF分支）

inkos book create --title "烈焰前传" --genre xuanhuan
inkos import canon 烈焰前传 --from 吞天魔帝       # 导入正传约束
inkos write next 烈焰前传                         # 写手自动读取正传约束

inkos detect 吞天魔帝 --chapter 3     # 检测章节中的 AIGC 标记
inkos detect 吞天魔帝 --all           # 检测所有章节
inkos detect --stats                  # 查看检测历史统计

守护进程（持续生产）

inkos daemon start                    # 启动调度器（默认：15 分钟/周期）
inkos daemon stop
inkos daemon status

全局配置 (`~/.inkos/config.json`)

{
  "llm": {
    "provider": "openai",
    "model": "gpt-4o",
    "apiKey": "sk-...",
    "temperature": 0.8
  },
  "daemon": {
    "intervalMinutes": 15,
    "dailyChapterLimit": 10,
    "parallelBooks": 2
  },
  "webhook": {
    "url": "https://your-server.com/hooks/inkos",
    "secret": "your-hmac-secret",
    "events": ["chapter-complete", "audit-failed", "pipeline-error"]
  },
  "aigcDetection": {
    "provider": "gptzero",
    "apiKey": "...",
    "endpoint": "https://api.gptzero.me/v2/predict/text"
  }
}

每本书规则 (`story/book_rules.md`)

# Book Rules: 吞天魔帝

## 禁忌 (Forbidden)
- 主角不得主动求饶
- 不得出现「命运」「天意」等宿命论表述

## 高疲劳词
- 震撼, 惊骇, 恐惧, 颤抖

## additionalAuditDimensions
- 数值系统一致性: 战力数值不得前后矛盾
- 角色成长节奏: 主角突破间隔不少于3章

## 写手特别指令
- 战斗场面优先感官描写，禁止数值报告

InkOS 按顺序运行五个专门的智能体：

ArchitectAgent  →  outline.md, book_rules.md
     ↓
WriterAgent     →  ch00N.md (reads: outline, summaries, arcs, matrix, style_guide, canon)
     ↓
ValidatorAgent  →  11 deterministic rules, zero LLM cost
     ↓  (error found → trigger spot-fix immediately)
AuditorAgent    →  33 LLM dimensions, temperature=0 for consistency
     ↓
ReviserAgent    →  spot-fix | rewrite | polish | anti-detect

写作后验证器规则（确定性，无需 LLM）

规则	条件
禁忌句式	`不是……而是……` 结构
破折号禁令	`——` 字符
转折词密度	仿佛/忽然/竟然 ≤1 每 3000 字符
高疲劳词	每本书列表，每章 ≤1
元叙事	编剧式旁白
报告术语	分析性框架术语出现在散文中
作者说教	显然/不言而喻等
集体反应	「全场震惊」陈词滥调
连续“了”	≥4 个连续句子包含“了”
段落长度	≥2 个段落超过 300 字符
书籍特定禁令	`book_rules.md` 禁止列表

审核维度（共 33 个，LLM 评估）

关键维度包括：

维度 1–23：核心叙事质量（情节、角色、节奏、伏笔）
维度 24–26：支线停滞、弧线扁平、节奏单调（所有 5 个流派）
维度 27：敏感内容
维度 28–31：衍生作品特定（正传冲突、未来信息泄露、世界规则一致性、伏笔隔离）
维度 32：读者预期管理
维度 33：大纲偏差检测

编程使用 (TypeScript)

import { BookManager } from '@actalk/inkos'
import { WriterAgent } from '@actalk/inkos/agents'
import { ValidatorAgent } from '@actalk/inkos/agents'

// 创建并配置一本书
const manager = new BookManager()
const book = await manager.createBook({
  title: '吞天魔帝',
  genre: 'xuanhuan',
  outlinePath: './my-outline.md'
})

// 为下一章运行写作流水线
const writer = new WriterAgent({ temperature: 0.8 })
const chapter = await writer.writeNext(book)

// 运行确定性验证（无 LLM 成本）
const validator = new ValidatorAgent()
const validationResult = await validator.validate(chapter, book)

if (validationResult.hasErrors) {
  // 自动触发定点修复
  const reviser = new ReviserAgent({ mode: 'spot-fix' })
  const fixed = await reviser.revise(chapter, validationResult.errors, book)
  console.log('已修复违规：', validationResult.errors.length)
}

Webhook 处理器 (Express)

import express from 'express'
import crypto from 'crypto'

const app = express()
app.use(express.raw({ type: 'application/json' }))

app.post('/hooks/inkos', (req, res) => {
  const sig = req.headers['x-inkos-signature'] as string
  const expected = crypto
    .createHmac('sha256', process.env.INKOS_WEBHOOK_SECRET!)
    .update(req.body)
    .digest('hex')

  if (sig !== `sha256=${expected}`) {
    return res.status(401).send('Invalid signature')
  }

  const event = JSON.parse(req.body.toString())

  switch (event.type) {
    case 'chapter-complete':
      console.log(`"${event.book}" 的第 ${event.chapter} 章已完成`)
      // 触发人工审核关卡
      notifyReviewer(event)
      break
    case 'audit-failed':
      console.log(`审核失败：${event.criticalCount} 个关键问题`)
      break
    case 'pipeline-error':
      console.error(`"${event.book}" 中的流水线错误：`, event.error)
      break
  }

  res.status(200).json({ received: true })
})

自定义流派定义

// genres/wuxia.ts
import type { GenreConfig } from '@actalk/inkos/types'

export const wuxia: GenreConfig = {
  id: 'wuxia',
  name: '武侠',
  chapterTypes: ['江湖相遇', '武功切磋', '恩怨纠葛', '门派争斗', '武林大会'],
  forbiddenPatterns: [
    '内力值', '战力', '等级提升',   // 无数值化力量体系
    '系统', '面板', '属性点'
  ],
  fatiguedWords: ['震惊', '无敌', '碾压', '秒杀'],
  languageRules: [
    {
      bad: '内力增加了100点',
      good: '一股暖流沿经脉漫开，指尖的颤抖渐渐平息'
    }
  ],
  auditDimensions: [
    '武功描写感官化',
    '江湖规则内部一致性',
    '恩怨情仇弧线完整性'
  ]
}

风格导入流水线

import { StyleAnalyzer } from '@actalk/inkos/style'

const analyzer = new StyleAnalyzer()

// 分析参考文本
const profile = await analyzer.analyze('./reference-novel.txt')
console.log(profile)
// {
//   avgSentenceLength: 18.3,
//   ttr: 0.42,           // 类符-形符比
//   rhetoricalDensity: 0.15,
//   paragraphLengthDist: { p25: 45, p50: 89, p75: 156 },
//   punctuationStyle: 'sparse'
// }

// 导入到书籍（生成 style_profile.json + style_guide.md）
await analyzer.importToBook('./reference-novel.txt', '吞天魔帝', {
  authorName: '某作者'
})

v0.4 强化的审核-修订循环可防止修订引入更多 AI 标记：

AuditorAgent (temp=0)
     ↓ 发现关键问题
ReviserAgent spot-fix
     ↓
AI 标记数量比较
     ↓ 标记增加？
  是 → 丢弃修订，保留原文
  否 → 接受修订
     ↓
重新审核 (temp=0)

审核一致性的关键设置：

审核器始终以 temperature: 0 运行 — 消除同一章节上 0–6 个关键问题的方差
默认修订模式为 spot-fix（仅修改问题句子）
提供 rewrite 模式，但实测会引入 6 倍多的 AI 标记
polish 模式有边界锁定：不增删段落，不改名，不添加新情节

衍生作品正传约束

当检测到 parent_canon.md 时，会自动激活 4 个额外的审核维度：

# parent_canon.md (由 `inkos import canon` 自动生成)

## 正传世界规则
- 力量体系: 炼体→炼气→炼丹→炼神→炼虚
- 地理: 九州大陆，东海以东无人居住
- 阵营: 正道五宗 vs 魔道三门

## 关键事件时间线
- 第1章: 主角获得吞天诀
- 第45章: 正道五宗盟约成立 [分歧点]

## 角色快照 (分歧点时状态)
- 林天: 炼气期第三层，未知父母身世
- 剑宗宗主: 在世，尚未叛变

## 伏笔状态 (正传专属，番外禁止回收)
- 古剑残片: 未解
- 神秘老人身份: 未解

验证器对每一章都触发 检查 book_rules.md 中的疲劳词列表 — 列出的词每章强制执行 ≤1 次。移除不常用的词。

审核结果在不同运行间差异巨大 确认审核器温度在配置中锁定为 0。如果使用代理 LLM API，确保其尊重 temperature: 0。

修订比原文引入了更多 AI 标记 这是预期行为 — InkOS v0.4 会自动检测并丢弃修订。如果反复发生，请明确从 rewrite 模式切换到 spot-fix 模式。

衍生作品审核错误地标记事件 验证 parent_canon.md 具有准确的分歧点时间戳。分歧点之前的事件受正传锁定；之后的事件衍生作品可以自由发挥。

守护进程在多次失败后停止某本书 检查 daemon status 中暂停的书籍。修复根本问题（通常是大纲模糊或 book_rules.md 矛盾）后，运行 inkos daemon resume 书名。

写手未应用风格指南 再次运行 inkos style import — 如果 style_guide.md 缺失或为空，写手会静默跳过风格注入。检查参考文本是否 ≥5000 字符以确保可靠的指纹提取。

🇺🇸English

InkOS Multi-Agent Novel Writing

Skill by ara.so — Daily 2026 Skills collection

InkOS is a multi-agent CLI system that autonomously writes, audits, and revises novels. Agents handle the full pipeline: Writer → Validator → Auditor → Reviser, with human review gates at configurable checkpoints.

Installation

npm install -g @actalk/inkos
# or run directly
npx @actalk/inkos --version

Requirements: Node.js ≥ 20.0.0

Quick Start

# Create a new novel project
inkos book create --title "吞天魔帝" --genre xuanhuan

# Write the next chapter
inkos write next 吞天魔帝

# Audit a specific chapter
inkos audit 吞天魔帝 --chapter 3

# Run the full daemon (continuous production)
inkos daemon start

Project Structure

After inkos book create, the project directory contains:

story/
  outline.md              # Story outline (architect agent input)
  book_rules.md           # Per-book custom rules and audit dimensions
  chapter_summaries.md    # Auto-generated per-chapter summaries
  subplot_board.md        # Subplot progress tracking (A/B/C lines)
  emotional_arcs.md       # Per-character emotional arc tracking
  character_matrix.md     # Character interaction matrix + info boundaries
  parent_canon.md         # Spinoff only: imported canon constraints
  style_profile.json      # Style fingerprint (if style import used)
  style_guide.md          # LLM-generated qualitative style guide
  chapters/
    ch001.md
    ch002.md
    ...

Core Commands

Book Management

inkos book create --title "Title" --genre xuanhuan  # genres: xuanhuan | xianxia | dushi | horror | general
inkos book list
inkos book status 吞天魔帝

Writing Pipeline

inkos write next 吞天魔帝           # Write next chapter (auto-loads all context)
inkos write chapter 吞天魔帝 5      # Write specific chapter
inkos audit 吞天魔帝 --chapter 3    # Audit chapter (33 dimensions)
inkos revise 吞天魔帝 --chapter 3   # Revise based on audit results
inkos revise 吞天魔帝 --chapter 3 --mode spot-fix   # Point fix only (default)
inkos revise 吞天魔帝 --chapter 3 --mode rewrite    # Full rewrite (use cautiously)
inkos revise 吞天魔帝 --chapter 3 --mode polish     # Polish (no structural changes)

Genre System

inkos genre list                        # List all built-in genres
inkos genre show xuanhuan               # View full rules for a genre
inkos genre copy xuanhuan               # Copy genre rules to project for customization
inkos genre create wuxia --name 武侠     # Create new genre from scratch

Style Matching

inkos style analyze reference.txt               # Analyze style fingerprint
inkos style import reference.txt 吞天魔帝        # Import style into book
inkos style import reference.txt 吞天魔帝 --name "某作者"

Spinoff (Prequel/Sequel/IF-branch)

inkos book create --title "烈焰前传" --genre xuanhuan
inkos import canon 烈焰前传 --from 吞天魔帝       # Import parent canon constraints
inkos write next 烈焰前传                         # Writer auto-reads canon constraints

AIGC Detection

inkos detect 吞天魔帝 --chapter 3     # Detect AIGC markers in chapter
inkos detect 吞天魔帝 --all           # Detect all chapters
inkos detect --stats                  # View detection history statistics

Daemon (Continuous Production)

inkos daemon start                    # Start scheduler (default: 15 min/cycle)
inkos daemon stop
inkos daemon status

Configuration

Global Config (`~/.inkos/config.json`)

{
  "llm": {
    "provider": "openai",
    "model": "gpt-4o",
    "apiKey": "sk-...",
    "temperature": 0.8
  },
  "daemon": {
    "intervalMinutes": 15,
    "dailyChapterLimit": 10,
    "parallelBooks": 2
  },
  "webhook": {
    "url": "https://your-server.com/hooks/inkos",
    "secret": "your-hmac-secret",
    "events": ["chapter-complete", "audit-failed", "pipeline-error"]
  },
  "aigcDetection": {
    "provider": "gptzero",
    "apiKey": "...",
    "endpoint": "https://api.gptzero.me/v2/predict/text"
  }
}

Per-Book Rules (`story/book_rules.md`)

# Book Rules: 吞天魔帝

## 禁忌 (Forbidden)
- 主角不得主动求饶
- 不得出现「命运」「天意」等宿命论表述

## 高疲劳词
- 震撼, 惊骇, 恐惧, 颤抖

## additionalAuditDimensions
- 数值系统一致性: 战力数值不得前后矛盾
- 角色成长节奏: 主角突破间隔不少于3章

## 写手特别指令
- 战斗场面优先感官描写，禁止数值报告

Agent Architecture

InkOS runs five specialized agents in sequence:

ArchitectAgent  →  outline.md, book_rules.md
     ↓
WriterAgent     →  ch00N.md (reads: outline, summaries, arcs, matrix, style_guide, canon)
     ↓
ValidatorAgent  →  11 deterministic rules, zero LLM cost
     ↓  (error found → trigger spot-fix immediately)
AuditorAgent    →  33 LLM dimensions, temperature=0 for consistency
     ↓
ReviserAgent    →  spot-fix | rewrite | polish | anti-detect

Post-Write Validator Rules (Deterministic, No LLM)

Rule	Condition
Forbidden patterns	`不是……而是……` constructs
Em-dash ban	`——` character
Transition word density	仿佛/忽然/竟然≤1 per 3000 chars
High-fatigue words	Per-book list, ≤1 per chapter
Meta-narrative	Screenwriter-style narration
Report terminology	Analytical framework terms in prose
Author moralizing	显然/不言而喻 etc.
Collective reaction	「全场震惊」clichés
Consecutive 了	≥4 consecutive sentences with 了
Paragraph length	≥2 paragraphs over 300 chars
Book-specific bans	forbidden list

Audit Dimensions (33 total, LLM-evaluated)

Key dimensions include:

Dims 1–23: Core narrative quality (plot, character, pacing, foreshadowing)
Dim 24–26: Subplot stagnation, arc flatness, rhythm monotony (all 5 genres)
Dim 27: Sensitive content
Dim 28–31: Spinoff-specific (canon conflicts, future info leakage, world rule consistency, foreshadowing isolation)
Dim 32: Reader expectation management
Dim 33: Outline deviation detection

Code Integration Examples

Programmatic Usage (TypeScript)

import { BookManager } from '@actalk/inkos'
import { WriterAgent } from '@actalk/inkos/agents'
import { ValidatorAgent } from '@actalk/inkos/agents'

// Create and configure a book
const manager = new BookManager()
const book = await manager.createBook({
  title: '吞天魔帝',
  genre: 'xuanhuan',
  outlinePath: './my-outline.md'
})

// Run the write pipeline for next chapter
const writer = new WriterAgent({ temperature: 0.8 })
const chapter = await writer.writeNext(book)

// Run deterministic validation (no LLM cost)
const validator = new ValidatorAgent()
const validationResult = await validator.validate(chapter, book)

if (validationResult.hasErrors) {
  // Auto spot-fix triggered
  const reviser = new ReviserAgent({ mode: 'spot-fix' })
  const fixed = await reviser.revise(chapter, validationResult.errors, book)
  console.log('Fixed violations:', validationResult.errors.length)
}

Webhook Handler (Express)

import express from 'express'
import crypto from 'crypto'

const app = express()
app.use(express.raw({ type: 'application/json' }))

app.post('/hooks/inkos', (req, res) => {
  const sig = req.headers['x-inkos-signature'] as string
  const expected = crypto
    .createHmac('sha256', process.env.INKOS_WEBHOOK_SECRET!)
    .update(req.body)
    .digest('hex')

  if (sig !== `sha256=${expected}`) {
    return res.status(401).send('Invalid signature')
  }

  const event = JSON.parse(req.body.toString())

  switch (event.type) {
    case 'chapter-complete':
      console.log(`Chapter ${event.chapter} of "${event.book}" complete`)
      // Trigger human review gate
      notifyReviewer(event)
      break
    case 'audit-failed':
      console.log(`Audit failed: ${event.criticalCount} critical issues`)
      break
    case 'pipeline-error':
      console.error(`Pipeline error in "${event.book}":`, event.error)
      break
  }

  res.status(200).json({ received: true })
})

Custom Genre Definition

// genres/wuxia.ts
import type { GenreConfig } from '@actalk/inkos/types'

export const wuxia: GenreConfig = {
  id: 'wuxia',
  name: '武侠',
  chapterTypes: ['江湖相遇', '武功切磋', '恩怨纠葛', '门派争斗', '武林大会'],
  forbiddenPatterns: [
    '内力值', '战力', '等级提升',   // No numerical power system
    '系统', '面板', '属性点'
  ],
  fatiguedWords: ['震惊', '无敌', '碾压', '秒杀'],
  languageRules: [
    {
      bad: '内力增加了100点',
      good: '一股暖流沿经脉漫开，指尖的颤抖渐渐平息'
    }
  ],
  auditDimensions: [
    '武功描写感官化',
    '江湖规则内部一致性',
    '恩怨情仇弧线完整性'
  ]
}

Style Import Pipeline

import { StyleAnalyzer } from '@actalk/inkos/style'

const analyzer = new StyleAnalyzer()

// Analyze reference text
const profile = await analyzer.analyze('./reference-novel.txt')
console.log(profile)
// {
//   avgSentenceLength: 18.3,
//   ttr: 0.42,           // Type-Token Ratio
//   rhetoricalDensity: 0.15,
//   paragraphLengthDist: { p25: 45, p50: 89, p75: 156 },
//   punctuationStyle: 'sparse'
// }

// Import into book (generates style_profile.json + style_guide.md)
await analyzer.importToBook('./reference-novel.txt', '吞天魔帝', {
  authorName: '某作者'
})

Audit Revision Loop

The v0.4 hardened audit-revision loop prevents revision from introducing more AI markers:

AuditorAgent (temp=0)
     ↓ critical issues found
ReviserAgent spot-fix
     ↓
AI marker count comparison
     ↓ markers increased?
  YES → discard revision, keep original
  NO  → accept revision
     ↓
Re-audit (temp=0)

Key settings for audit consistency:

Auditor always runs at temperature: 0 — eliminates 0–6 critical variance on same chapter
Default revision mode is spot-fix (only modify problem sentences)
rewrite mode is available but measured to introduce 6× more AI markers
polish mode is boundary-locked: no paragraph add/delete, no name changes, no new plot

Spinoff Canon Constraints

When parent_canon.md is detected, 4 additional audit dimensions activate automatically:

# parent_canon.md (auto-generated by `inkos import canon`)

## 正传世界规则
- 力量体系: 炼体→炼气→炼丹→炼神→炼虚
- 地理: 九州大陆，东海以东无人居住
- 阵营: 正道五宗 vs 魔道三门

## 关键事件时间线
- 第1章: 主角获得吞天诀
- 第45章: 正道五宗盟约成立 [分歧点]

## 角色快照 (分歧点时状态)
- 林天: 炼气期第三层，未知父母身世
- 剑宗宗主: 在世，尚未叛变

## 伏笔状态 (正传专属，番外禁止回收)
- 古剑残片: 未解
- 神秘老人身份: 未解

Troubleshooting

Validator fires on every chapter Check book_rules.md fatigue word list — words listed there are enforced ≤1 per chapter. Remove infrequently used terms.

Audit results wildly inconsistent between runs Confirm auditor temperature is locked to 0 in config. If using a proxy LLM API, ensure it respects temperature: 0.

Revision introduces more AI markers than original This is expected behavior — InkOS v0.4 automatically detects this and discards the revision. If it happens repeatedly, switch from rewrite to spot-fix mode explicitly.

Spinoff audit incorrectly flags events Verify parent_canon.md has accurate divergence point timestamps. Events before the divergence point are canon-locked; events after are fair game for the spinoff.

Daemon stops a book after repeated failures Check daemon status for the suspended book. After fixing the underlying issue (usually outline ambiguity or book_rules.md contradiction), run inkos daemon resume 书名.

Style guide not being applied by writer Run inkos style import again — if style_guide.md is missing or empty, the writer skips style injection silently. Check that the reference text is ≥5000 characters for reliable fingerprinting.

Weekly Installs

420

Repository

aradotso/trending-skills

GitHub Stars

First Seen

8 days ago

Security Audits

Gen Agent Trust HubPass SocketPass SnykPass

Installed on

cline414

gemini-cli414

cursor414

opencode414

codex414

github-copilot414

AI 代码实施计划编写技能 | 自动化开发任务分解与 TDD 流程规划工具

41,400 周安装