ZenStory AI
Agent skill pack · Agent skill 包
Oh Story
Turn a coding agent into a complete web-fiction writing workflow — chart scanning, deconstruction, drafting, de-AI editing, covers.
把编程 Agent 变成完整的网文写作流程——扫榜、拆文、写作、去AI味、封面图。
6,770 GitHub stars · 13 skills · MIT license · Format: Agent skill pack
Source on GitHub
English README
npx skills add zenstory-ai/oh-story-claudecode -y -g
What it is
Oh Story is an open-source skill pack (MIT) that gives coding agents — Claude Code, Codex CLI, Google Antigravity, OpenCode, ZCode, OpenClaw and Reasonix — a professional web-fiction author's method, for both long-form serials and short stories. It is not a prompt collection: the 13 skills ship with deterministic verifiers, blocking reference gates, layered context and state management, and a one-command deployer (/story-setup) for each harness.
它是什么
Oh Story 是一个 MIT 开源的 skill 包，让 Claude Code、Codex CLI、Google Antigravity、OpenCode、ZCode、OpenClaw、Reasonix 等编程 Agent 按职业网文作者的方法论工作，长篇连载和短篇都覆盖。它不是提示词合集：13 个 skill 内置确定性 verifier、写正文前会阻断的 Reference Gate、分层的上下文与状态管理，以及一条命令部署到各 Agent 环境的 /story-setup。
Who it is for · 适合谁
- Authors publishing on 起点 Qidian, 番茄 Fanqie, 晋江 JJWXC, 七猫 Qimao and 知乎盐言
- Studios running several serials at once
- Developers who want a mature, non-coding example of agent skills
- 在起点、番茄、晋江、七猫、知乎盐言等平台写作的作者
- 同时运营多本书的工作室
- 想看非编程场景下成熟 agent skills 的开发者
How it works · 流程
- 扫榜 · chart scanning Scrape the live bestseller boards of five platforms and look for patterns that repeat across books — genre, cast, angle — rather than trusting any single ranking.
- 拆文 · deconstruction Reverse-engineer a hit into outline rhythm, plot modules, character and style notes; the first three chapters (黄金三章) are analysed first as a make-or-break checkpoint.
- 商业化写作 · commercial drafting Write chapter by chapter from a per-chapter blueprint with hooks, payoff pacing (爽点) and an enforced word target; a tracking state file is the single source of truth for continuity.
- 去AI味 · de-AI editing A deterministic checker flags ~20 named Chinese AI-prose fingerprints and model-degeneration patterns; the skill rewrites what it flags.
- 封面 · covers Genre-aware cover generation from the project's own setting files.
- 扫榜 抓取五个平台的实时榜单，找跨样本重复的题材、人设、切入点，而不是信任单本排名。
- 拆文 把爆款拆成大纲节奏、剧情模块、角色与文风；先拆黄金三章作为检查点，再决定是否全拆。
- 商业化写作 按逐章细纲写正文：钩子、爽点节奏、机器校验的字数目标；追踪状态文件是连续性的唯一真相源。
- 去AI味 确定性检查器识别约 20 种中文 AI 文风指纹与模型退化模式，skill 据此改写。
- 封面 根据项目设定文件做题材感知的封面生成。
What makes it different · 有什么不同
- Deterministic verifiers, not LLM self-checks: chapter-blueprint contract (16 required fields), AI-pattern scan, degeneration scan, outline-copy detection — each with documented exit codes and explicitly forbidden from judging quality.
- A file-system project bible (设定 / 大纲 / 正文 / 对标 / 追踪) with a single committed tracking state that every Markdown view is re-rendered from.
- Seven cooperating agents (architect, writer, consistency checker, reviewer…) deployed per harness by
/story-setup; cross-session author memory. - Ships real output samples: a short-story deconstruction, a long-form continuation project, and a cover set.
- 确定性 verifier 而不是让模型自查：细纲契约（16 个必填字段）、AI 文风扫描、退化扫描、细纲抄写检测，每个都有明确退出码，且明确不评判「好不好」。
- 文件系统级的项目圣经（设定 / 大纲 / 正文 / 对标 / 追踪），所有 Markdown 视图由唯一的追踪状态文件派生。
- 7 个专业 agent 协作，
/story-setup按环境部署；跨会话的作者记忆。 - 仓库附真实产出：短篇拆文、长篇续写工程、封面示例。