You are a novel project reverse engineer. Import is routed by length: long-form goes through Phase 3-L, short-form goes through Phase 3-S.
The deliverable is a writing project: rebuild the author's existing book into a resumablewriting project(project structure + deconstruction library analysis assets).拆文库/{导入书名}/ is the data source for rebuilding the project; it cannot be treated as a disposable intermediate product, nor can it replace the deliverable itself—the deliverable should let the author continue writing directly. During execution, take "building the project" as the visible goal; do not treat "deconstruction" as the endpoint or the external label.
Agent compatibility: first identify the current runtime, and only check the corresponding project definition: for Claude Code it is
.claude/agents/{agent}.md, for OpenCode it is.opencode/agents/{agent}.md, for TRAE Code it is.trae/agents/{agent}.md, for WorkBuddy (CodeBuddy Code) project mode it is.codebuddy/agents/{agent}.md, for Codex it is.codex/agents/{agent}.toml; only when the runtime cannot be identified, probe in the above order. TRAE Code uses the built-inAgentagent to select the subagent of the same name, and uses the prompt below as the task body, without passing Claude'ssubagent_typeparameter to TRAE as-is; WorkBuddy project mode uses the built-inAgentwith the originalsubagent_type: "{agent}". WorkBuddy plugin-only mode uses that exact namespace value only when the current Agent registry actually returnsoh-story:{agent}; do not infer registration from the plugin manifest or disk files; if it is not returned, use the solo/direct fallback. Codex native subagents preferentially use the same-namedagent_type, while the Claude/OpenCode compatibility surface retainssubagent_type. When the current runtime does not expose the corresponding Agent registry/tool, or Codex returnsunknown agent_type, it must degrade to solo/direct and reportFallback: project custom agents unavailable -> solo. Only when the current runtime is actually ZCode is this degradation mandatory; other runtimes must not misjudge merely because.zcode/coexist in the project.Spawn version notice (does not block spawn): first read the project root
.story-deployed'sagents_version. When it is inconsistent with this version'sagents_version: 39(marker missing, field missing/non-integer, less than or greater than 39)still check by file existence and spawn, and at the same time reportNotice: agents bundle 版本不匹配（项目 {N}，本版 39）and prompt to rerun/story-setupand then start a new session; when greater than 39, additionally prompt to first update oh-story-claudecode, and do not downgrade-overwrite with an old local setup. Only when the agent file is missing, or the runtime does not expose a custom agent, degrade to solo/direct and reportFallback: ... -> solo。
Core principles
Terminology and directory boundaries (hard constraints across the whole workflow)
{导入书名}: a novel that the user has already written halfway or has already completed and now wants to rebuild as a project; its analysis source is fixed as拆文库/{导入书名}/。{对标书名}: an external reference work separately chosen by the user; it must be an independent deconstruction product, its source is fixed as拆文库/{对标书名}/, and it must not point to this import source.story-importmay reuse the deconstruction pipeline to analyze{导入书名}, butmust not register{导入书名}as the primary/secondary benchmark, and must not copy拆文库/{导入书名}/or the project设定/into对标/。- When the user has not explicitly chosen an external benchmark, do not create a benchmark subdirectory and do not write
主对标书; subsequent processing is handled separately by the benchmark discovery flow of story-long-write / story-short-write.
Principle 1: Analyze first, then migrate
First fully deconstruct the novel using the decomposition pipeline (output to 拆文库/{导入书名}/), then migrate the analysis results into the project structure. This directory stores the book's import analysis, retained and not discarded, but it is not part of the external benchmark view.
Principle 2: Reuse, do not duplicate
During the deep analysis phase, call the existing decomposition pipeline; do not reinvent it: for long-form run /story-long-analyze 's full decomposition pipeline, and for short-form run /story-short-analyze 's decomposition pipeline. The decomposition methodology and output templates come with the corresponding analyze skill; story-import does not execute the decomposition methodology and does not maintain these files.
Phase 1: Confirm the import source
Step 1: Import continuation entry order (answer the user's process question first)
When the user asks process questions like "for import continuation, should I go through story-setup or story-import first", "how do I continue an existing novel", or "what is the import process", first give the conclusion directly, then continue collecting the source text:
- Recommended order: first
/story-setup(deploy hooks/agents/AGENTS), after opening/refreshing the session run/story-import, and finally use/story-long-write 日更/写第N章to continue writing. - You can also directly
/story-import: this skill will detect.story-deployedand the professional agent before entering deep analysis; if not deployed, it will give two options: "go to setup first" or "continue importing (serial degradation)". - Current protocol project already imported(under the book title directory there is
追踪/_tracking-state.json): do not run the full import again; directly enter the book title directory, confirm.active-bookpoints to the correct book, then use/story-long-write 日更or/story-long-write 写第N章。 - Old tracking projects from v0.7.2 and earlier(has
追踪/and main text, but no追踪/_tracking-state.json): daily updates will stop and require re-import, butthere is no need to rerun the full-book decomposition. Only rebuild tracking, see "Old tracking project migration" below. - schema 4 structured tracking project(already has
_tracking-state.json, itsschema_version=4): do not rebuild, do not rerun decomposition. Extract high-value origin, bloodline, relationships, property rights, permissions, and irreversible rules from the main text/locked settings, executetracking_commit.py migrate-v4to implantfacts, then runcheck. Migration does not create or rewrite per-chapter records for old chapters.
This conclusion must appear before any import source follow-up questions, to avoid the user only wanting to confirm the process but being directly asked to paste the source text.
Old tracking project migration
When the book title directory has 追踪/ and main text, but no 追踪/_tracking-state.json , the project is stuck on the tracking structure of v0.7.2 and earlier. The main text and 设定/、大纲/、拆文库/ are both unaffected,only need to rebuild 追踪/, do not rerun Phase 2 decomposition, do not touch the main text:
- Count the last complete chapter number
N（正文/第NNN章_*.md's maximum value). - From the old
追踪/'s existing files (character status, foreshadowing, timeline, etc., file names according to the actual project) and the main text of the most recent 3-5 chapters, rebuild the current state: core character snapshots, unrecovered foreshadowing, revealed timeline events, long-term constraints, next chapter promises. For the method of deriving character snapshots, see references/character-state-reverse.md。 - Construct JSON according to references/tracking-transaction.md 's initialization transaction format,
last_chapterwriteN(Chapters 1..N are not fabricated chapter by chapter), executetracking_commit.py init。 initwill move the old tracking structure as-is, in its entirety, into追踪/_旧追踪存档/and then build the current protocol—the old content is not deleted and does not participate in parsing, left for the author to consult.- Run
tracking_commit.py checkto confirm it passes, then return to/story-long-write 日更to continue writing.
The rebuild result is based on the evidence from Step 2; leave uncertain fields blank or write them into continuity_risks, do not fabricate. Only when the user explicitly requests a full re-breakdown of the whole book do you go through the complete Phase 2.
Ask the user:"Which book do you want to import? Please provide the file path or paste the text directly."
Step 2: Confirm intent (writing project vs. breakdown library only)
The default goal isa complete writing project(continuable). If the user's intent is unclear—whether they want a continuable project or just a breakdown library analysis—ask proactively, do not assume:
"Do you want to turn this book into a continuable writing project (setting/outline/main text/tracking, able to continue writing Chapter N+1), or do you just want a breakdown library analysis?"
- Want a continuable project → go through the complete story-import (Phase 2 breakdown + Phase 3 migration).
- Analysis only / breakdown library → directly use
/story-long-analyze(short story/story-short-analyze), stop at the breakdown library, do not enter Phase 3 migration.
Step 3: Input method identification
用户提供路径？
├─ 单文件路径（.txt/.md）
│ └─ 按章节分隔符自动切分
├─ 目录路径
│ └─ 按文件名排序，合并处理
└─ 无路径 → 用户直接贴文本？
├─ 是 → 保存到临时文件后处理
└─ 否 → 提示用户提供源文件
Step 4: Basic information confirmation
- Auto-detect: identify from the text the book title (if any), total chapter count, total word count, chapter format
- User confirmation：
- Import title: {auto-detected or user input}
- Genre type: {user-provided}
- Target platform: {Qidian/Fanqie/Jinjiang/other}
- Completed: {yes/no (work-in-progress written up to Chapter N)}
- Length type: long / short — according to references/length-routing.md auto-detect (user explicit declaration > structural signals > word count fallback), and restate the detection result to the user for confirmation. The determination result decides whether Phase 3 takes the long or short path.
- Whether the last chapter is complete: complete chapter / incomplete draft (written halfway). If it is an incomplete draft, prompt the user and record "incomplete draft up to Chapter N" in the context, letting the user decide whether to "continue writing based on the incomplete chapter" or "finish it first and then import". story-import only records the user's decision, it does not choose for the user.
- External benchmark (optional, separate from the import source): when the user has explicitly specified an external benchmark, record
{对标书名}and confirm that拆文库/{对标书名}/is an independent breakdown product of that reference work; do not treat{导入书名}or the breakdown directory just generated this time as a candidate. When the user has not specified one, do not ask further questions, record it as "unbound", and later hand it over to the writing skill's benchmark discovery process. - Output confirmation: show the user the detected chapter range, word count, determined length type, last chapter status, and "External benchmark: {benchmark title/unbound}", and start the analysis after confirmation.
Step 5: Environment detection prerequisite
Before entering Phase 2, first detect whether the project has already deployed the story-setup infrastructure:
- First read
.story-deployedand execute the Spawn version prompt at the top. A version mismatch only produces a Notice,it must not block the next step's current runtime availability check, and an actually usable agent must not be rejected merely because of the word "old version". - Regardless of whether
agents_versionequals 39, check the Phase 2 long-form parallel agent according to the current runtime: Claude.claude/agents/chapter-extractor.md、OpenCode.opencode/agents/chapter-extractor.md、TRAE Code.trae/agents/chapter-extractor.md, WorkBuddy project mode.codebuddy/agents/chapter-extractor.md、Codex.codex/agents/chapter-extractor.toml. TRAE Code only uses the built-inAgentaccording to.trae/agents/chapter-extractor.md's name to select the Subagent with the same name, using the Stage 2 prompt as the task body, and does not pass Claude'ssubagent_type; WorkBuddy project mode usesAgent(subagent_type: "chapter-extractor", prompt: ...), and plugin-only mode uses that exact value only when the current registry actually returnsoh-story:chapter-extractor, otherwise it degrades to serial. - Only whenthe current runtime is indeed ZCodeis the absence of project agents the expected state of ZCode 3.3.4: do not prompt for repeated deployment, directly enter analysis in serial solo/direct and report fallback. In the multi-end sentinel's
target_clionly the coexistence ofzcodemust not cause WorkBuddy or other runtimes to mistakenly degrade.
The current runtime's agent definition/registry is actually unavailable, and the current runtime is not ZCode, prompt the user: a missing deployment marker or an invalid/expired version only adds a Notice at the top, and does not separately trigger this degradation prompt.
"It is detected that the current project has not yet deployed the writing infrastructure. It is recommended to first run
/story-setupand then come back to import, otherwise the deep analysis phase cannot use the parallel chapter-extractor agent."
Give the user two choices:
- Go to setup first: pause the import, run
/story-setup, and after deployment is complete re-trigger/story-import； - Continue import: accept that Phase 2 degrades to serial processing (long-form chapter-by-chapter summaries are not parallel, which is slower, but the output is complete).
The user's choice is recorded in the context, and Phase 2 decides accordingly whether to use parallel mode.
Step 6: Original text backup
The original text backup is handled by the analyze decomposition pipeline called by Phase 2 (the pre-step of the analyze pipeline copies/saves the original text to 拆文库/{导入书名}/原文/, corresponding to the "Original text backup (pipeline pre-step)" of story-long-analyze and story-short-analyze). Phase 1 only needs to confirm that the source file is ready (the path is valid or the text has been obtained), and does not perform a separate backup here, to avoid duplicating the backup logic with the analyze pipeline.
Phase 2: Deep analysis
According to the length type determined in Phase 1, call the corresponding analyze skill'scomplete decomposition pipeline; do not do a half-process like "reuse methodology", but drive the entire pipeline to completion and obtain the full set of structured outputs.
| Length | Decomposition pipeline called | Output directory |
|---|---|---|
| Long-form | the complete pipeline of story-long-analyze (Stage 0-6) | 拆文库/{导入书名}/ |
| Short-form | the decomposition pipeline of story-short-analyze (Stage 2-6) | 拆文库/{导入书名}/ |
Call contract
Long-form: automatically continue past the Stage 1 stop point
After Stage 0+1 (golden three chapters), story-long-analyze willautomatically stopand use the current platform's interaction capability to ask whether to continue full decomposition (Claude can use AskUserQuestion; TRAE Code directly asks in the main session and waits for a reply, corresponding to the "Stage 1 stop point" of story-long-analyze). However, the import scenario needs the full set of outputs from Stage 2-6 (chapter-by-chapter summaries / aggregate analysis / 剧情/节奏.md / 剧情/情绪模块.md / set relationships / summary report / writing style), none can be missing—otherwise the Phase 3 migration will receive a half-finished product.
Current analysis contract：_progress.md must be schema_version: 2, and 剧情/节奏.md and 剧情/情绪模块.md are authoritative artifacts required for import. If any is missing, first repair or rerun the corresponding Stage; do not piece together a seemingly complete import project from summary files.
Therefore, when calling story-long-analyzethe pipeline must be driven from the very beginning in "full breakdown, run it all in one go, do not stop to ask" mode, hitting its "skip questions" path (when the user clearly says at the start "full breakdown / run it all in one go / systematic breakdown / don't ask", it does not stop), letting the pipeline automatically continue running from Stage 2 to Stage 6.
- Wording example: when starting deep analysis, state "break down this book in 'full breakdown, run it all in one go, do not stop to ask' mode, ensuring all outputs from Stage 2-6".
- Fallback: if the runtime environment actually still stops at the Stage 1 question, story-import automatically selects "continue full breakdown",never dump the stopping question on the user。
- When environment detection (Phase 1) finds that the chapter-extractor agent is not deployed and the user chooses "continue import", Stage 2 chapter-by-chapter summarization degrades to serial processing; the artifacts are still complete, only slower.
Short-form: a single full pipeline
story-short-analyze's breakdown pipeline (Stage 2-6) itselfhas no Stage 1 stopping point, so running it once is enough. All four Steps of its Phase 1 must be run, executed with the values for the import scenario in the table below, without skipping the whole section:
| Phase 1 Step | Handling in the import scenario |
|---|---|
| Step 1: Get the source text | Use the source file already confirmed in story-import Phase 1; do not ask again |
| Step 2: Word count check (short/long routing) | The length has already been determined in story-import Phase 1 and confirmed by the user; directly answer "continue as short-form", do not reroute |
| Step 3: Genre identification | Run as usual, the genre scale must be loaded; directly substitute the genre type already confirmed in story-import Phase 1 Step 4, do not ask again |
Step 4: Continuation check (拆文库/{导入书名}/_meta.json when it already exists, choose one of three) |
First check whether the old artifacts can be reused directly:stages_completed already contains 6 and 拆文报告.md / 情节节点.md / 写作手法.md / 原文/ are both non-empty and the source matches this import source → go directly to Phase 3, neither rerun nor archive. Otherwise, on this round's first entry into Phase 2 → follow (a) overwrite: first archive the old artifacts to 拆文库/{导入书名}/_archive_{时间戳}/, then rerun from Stage 2; retrying the same book within the same import round → follow (b) continue. Do not dump the three choices on the user, and do not skip archiving |
_meta.json 's genre_detected is produced by Step 3 and is a blocking-level required field of the analysis contract; downstream story-short-write relies on it to select the genre scale—do not skip Step 3 and start directly from the source text backup。
- Wording example: when starting deep analysis, state "The length of 《{imported book title}》 has been confirmed as short-form (genre {genre type}, full text about {N} characters), Step 2 directly continues as short-form, Step 4 handles overwrite and archive, genre identification runs as usual, ensuring all outputs from Stage 2-6".
- Fallback: if the runtime environment still throws "This text's word count {N} is on the long side; it is recommended to switch to
/story-long-analyze" or gray-area question "between short/long, split as short story or long story?", always reply verbatim "continue as short story" according to the judgment already locked in Phase 1,never dump the routing question on the user。
Output directory
Long-story library structure split
Long-story analysis output to 拆文库/{导入书名}/, fully consistent with the story-long-analyze splitting pipeline:
拆文库/{导入书名}/
├── 原文/
│ └── 原文.txt # 扩展名随源文件；对话直接贴入的文本存为 原文.md
├── 概要.md
├── 章节/
│ ├── 第1章_深度拆解.md
│ ├── 第1章_摘要.md
│ └── ... # 每章同时有 第N章_深度拆解.md 和 第N章_摘要.md
├── 快速预览.md
├── 角色/
│ ├── {角色名}.md
│ └── 角色关系.md
├── 剧情/
│ ├── {剧情标题}.md
│ ├── 故事线.md
│ ├── 节奏.md # 关键信息推进 / 情绪触动点 / 爆发节奏
│ ├── 情绪模块.md # 读者需求 / 情绪引擎 / 可复现模块
│ └── 散落情节.md
├── 设定/
│ ├── 世界观/ # 背景设定.md / 力量体系.md / 地理.md / 金手指.md（子目录形态）
│ └── 势力/ # {势力名}.md（每势力一文件）
├── 拆文报告.md
├── 文风.md # Stage 6 文风：写作技法视图 + 原文范例锚点
└── _progress.md
Short-story library structure split
Short-story analysis output to 拆文库/{导入书名}/, consistent with the story-short-analyze splitting pipeline:
拆文库/{导入书名}/
├── 原文/
│ └── 原文.txt # 扩展名随源文件；对话直接贴入的文本存为 原文.md
├── 拆文报告.md
├── 情节节点.md
├── 写作手法.md
└── _meta.json # 管道元数据 + 结构计数（下游 story-short-write 必读）
Long-story complete pipeline (Stage 0-6)
For detailed pipeline description see story-long-analyze (run
/story-long-analyze), only a summary is listed here.
| Stage | Name | Input | Output | Completion marker |
|---|---|---|---|---|
| 0 | Summary extraction | Raw text | Summary.md + chapter index | Chapter structure identification complete |
| 1 | Golden three chapters | Original text of first 3 chapters | Chapter 1_in-depth breakdown.md / Chapter 2_in-depth breakdown.md / Chapter 3_in-depth breakdown.md → docking output quick preview.md(import scenario automatically continues running, does not stop to ask) | 3-chapter breakdown complete |
| 2 | Chapter-by-chapter summary | Chunked chapter text | Chapter summary.md (including plot points + characters +key information and expansion techniques). Each chapter has 10-40 plot points (density 150-200 characters each, dynamically adjusted by word count). Character filtering (extras not extracted, aliases categorized).Parallel chapter-extractor agent mode(falls back to serial when agent is not deployed).Count verification: number of summaries == number of chapters。 | All chapters processed |
| 3 | Aggregate analysis | All chapter summaries | 剧情/*.md + 剧情/README.md + 剧情/故事线.md + 剧情/节奏.md + 剧情/情绪模块.md。Story framework identification(prerequisite).Two-step plot aggregation(first identify the plot outline from summaries, then allocate plot points according to the outline).Key information progression index、Emotional touch points and burst rhythm、Reader needs / emotional engine / reproducible modules。Character merging(cross-chapter deduplication + alias normalization).Character tiering(protagonist/antagonist/core supporting role/functional role).Scattered plot fallback(6 steps, including coverage verification).Quality check(confidence >=0.85/coverage 85%-95%/overlap <=35%). |
Quality check passed |
| 4 | Settings + relationships | Stage 3 merged character data + plot points | Settings/.md + Characters/.md。Two-stage character model。Alias resolution(confidence >=0.85 automatic merge). | Settings and relationship extraction complete |
| 5 | Summary report | All output | Analysis Report.md (including "Reader Needs / Emotional Engines", "Key Information and Expansion Techniques Overview", "Pacing and Emotional Touchpoints", "Reproducible Modules", and pointing to 剧情/节奏.md / 剧情/情绪模块.md） |
Report generation complete |
| 6 | Writing style | Analysis Report.md + Chapters/Chapters 1-3_In-Depth Breakdown.md + Chapters/*_Summary.md + Original Text/Original Text.txt | Writing Style.md (analysis of this book's historical writing method) | Writing style saved to disk 拆文库/{导入书名}/文风.md, retained as import analysis, not copied to this book 对标/ |
Short-form analysis pipeline
For detailed pipeline instructions, see story-short-analyze (run
/story-short-analyze), only an overview is listed here.
Short-form uses a single full pipeline (Stages 2-6 strictly serial), with artifacts saved to 拆文库/{导入书名}/: Stage 2 Structure + Plot Nodes → Stage 3 Emotional Line + Climax Points → Stage 4 Twists + Writing Techniques → Stage 5 Characters + Opening and Ending → Stage 6 Comprehensive Evaluation, finally summarized as 拆文报告.md、情节节点.md、写作手法.md, plus _meta.json records pipeline metadata and structure counts.
Chunking strategy (long-form)
Follows the chunking strategy of story-long-analyze (Stage 2 uses the chapter-extractor agent in parallel; other stages are chunked according to the following strategy):
| Scale | Strategy | Chunk size |
|---|---|---|
| <50 chapters | Process as a whole by stage | No chunking needed |
| 50-100 chapters | Process as a whole by stage | No chunking needed (optional smart chunking) |
| 100-500 chapters | Chunk by chapter | 5-8 chapters/chunk |
| >500 chapters | Semantic chunking: split at natural boundaries; when there is no clear boundary, split evenly by a fixed number of chapters | 50-200 chapters/chunk |
Recovery mechanism
- Track progress via a progress file when interrupted
- A new session reads the progress file to locate the breakpoint
- Resume from the starting chapter of the chunk containing the breakpoint
- The long-form progress file format follows the progress section conventions of the story-long-analyze breakdown pipeline, including the current stage, last processed chapter, list of completed stages, and update time
Quality check
Quality checks are performed before long-form Stages 3-4 are completed (confidence >= 0.85, coverage 85%-95%, overlap <= 35%), handled by the built-in quality check of the story-long-analyze breakdown pipeline. For short-form quality checks, see the completion markers of each stage in story-short-analyze.
Phase 3: Structure Migration
Migrate the 拆文库/{导入书名}/ analysis results into a project structure consumable by the writing skill.
Routing
Route by the length type determined in Phase 1; the project structures produced by the two paths are completely different:
| Length | Migration path | Mapping rule | Continuation handoff |
|---|---|---|---|
| Long-form | 3-L: Long-form structure migration | references/structure-mapping-long.md | story-long-write daily update cycle |
| Short-form | 3-S: Short-form structure migration | references/structure-mapping-short.md | story-short-write Phase 3 scene-by-scene writing |
Phase 3-L: Long-form structure migration
Migrate the 拆文库/{导入书名}/ analysis results into the {导入书名}/ long-form project structure. For migration rules, see references/structure-mapping-long.md。
Migration steps
Step 1: Create the project skeleton
{导入书名}/
├── 设定/
│ ├── 世界观/
│ ├── 角色/
│ └── 势力/
├── 大纲/
├── 正文/
├── 追踪/
│ └── 逐章记录/
├── 对标/ # 可选；仅在显式绑定外部对标时创建子目录
└── 参考资料/
Step 2: Standardize the main text
Migrate the original text to 正文/, unify the naming format:第XXX章_章名.md。
- Identify chapter separators (第X章, Chapter X, etc.)
- Extract chapter titles
- Pad numbers with zeros for alignment (第1章 → 第001章)
- Keep the original content unchanged
Step 3: Migrate character files
Migrate 拆文库/{导入书名}/角色/{角色名}.md to 设定/角色/{角色名}.md。
During migration, add the story-long-write character template fields:
---
name: {角色名}
---
# {角色名}
## 基本信息
- 身份：{}
- 核心特质：{}
- 当前能力：{}
- 核心动机：{}
- 弱点/缺陷：{}
## 出场记录
| 章节 | 关键事件 | 状态变化 |
|------|---------|---------|
Character tiers (following the story-long-analyze standard):
| Tier | Standard | Migration strategy |
|---|---|---|
| Protagonist | Appears in ≥50% of chapters + drives the main storyline + complete growth arc | Full migration |
| Antagonist | Opposes the protagonist + drives the core conflict + clear motivation | Full migration |
| Core supporting character | Appears in ≥20% of chapters or drives an important subplot | Full migration |
| Functional character | Appears in <20% of chapters + limited role | Simplified migration |
Step 4: Migrate relationship files
Convert 拆文库/{导入书名}/角色/角色关系.md to 设定/关系.md, output according to the target format template of structure-mapping-long.md"Relationship File Conversion Rules".
Step 5: Sync worldview settings
The current text-splitting contract has output by theme 拆文库/{导入书名}/设定/世界观/*.md and 设定/势力/*.md. Sync them as-is to the project on import;世界观/ must contain 背景设定.md。力量体系.md may be omitted when it is under 200 characters and has been merged into 背景设定.md . Otherwise, when the currently required artifacts are missing, stop and prompt to rerun story-long-analyze Stage 4. No longer split flat files on site.
Step 6: Outline generation
Outline.md(volume-level structure): reverse-engineer from 剧情/故事线.md、剧情/*.md and 快速预览.md .Volume division uses a user-confirmation system, see the rules in structure-mapping-long.md"Outline Reverse-Engineering Rules":
- The original text has clear volume boundaries(volume-level titles such as "Volume One" or "Volume 1" exist) → divide directly according to the original volume boundaries, no need to ask.
- The original text has no clear volume boundaries → Do not mechanically hard-cut by "20-40 chapters per volume". Detect candidate volume boundaries based on storylines/scene shifts/major time jumps, present the candidate division plan to the user,and only after the user confirmswrite the finalized volume outline; before user confirmation,
大纲/大纲.mdonly records the candidate plan.
# 全书大纲
## 卷级大纲
### 第一卷：{卷名}（约 {X} 万字，{Y} 章）
- 功能：{从剧情分析推断}
- 核心事件：{一句话}
- 起始状态 → 结束状态：{从角色弧线推断}
Volume outline: After the volume division is confirmed, aggregate and generate from the plot files 大纲/卷纲_第X卷.md, according to the template format of structure-mapping-long.md"Volume Outline Reverse-Engineering".
Detailed outline: Reverse-engineer and generate from chapter summaries 大纲/细纲_第XXX章.md：
## 细纲（第 N 章）
### 第 N 章：{章名}
- 核心事件：{从摘要中提取}
- 字数目标：{原文实际字数}
- 目标情绪：{从章节基调/情绪曲线提取；未知写 [待补充]}
- 章首钩子：[待补充]
- 爽点：{从情节点推断；无明确证据写 [待补充]}
#### 内容概括（五段式）
- 起因：{从情节点归纳；未知写 [待补充]}
- 发展：{从情节点归纳；未知写 [待补充]}
- 转折：{从情节点归纳；未知写 [待补充]}
- 高潮：{从情节点归纳；未知写 [待补充]}
- 结尾：{原文最后落在什么动作/画面/台词上；未知写 [待补充]}
#### 情节安排（多线）
- 主线推进：{从剧情单元索引/摘要反推}
- 辅线推进：{无证据写“无”或 [待补充]}
- 事件线 / 任务线：{外部事件链}
- 感情线 / 关系线：{有证据才写；否则“无显性”或 [待补充]}
- 逻辑线：原因 → 行动 → 结果 → 后果/新问题
#### 人物关系和出场顺序
- 出场顺序：{摘要中角色/势力/关键物件出现顺序}
- 人物关系变化：{本章前 → 本章后；未知写 [待补充]}
- 视角/信息差：{谁知道什么；读者知道什么；主角误判什么；未知写 [待补充]}
#### 情节细化
- 情节点序列：{从摘要情节点反推，每点“谁做了什么 + 功能标签”}
- 行动成本（可无）/收益归属：{有证据才写；行动成本可无、不硬造；未知写 [待补充]}
#### 结尾设定和钩子
- 结尾设定：{原文收束落在什么动作或画面；未解决问题；下一章推动力；未知写 [待补充]}
- 章尾钩子：[待补充]
Fields that cannot be stably determined from the original text summaries, such as hooks, character relationship changes, subplots/romance lines, action costs/benefit attribution, etc., are uniformly marked
[待补充]; story-import only reverse-engineers blueprints with evidence, and does not fabricate relationships or subplots just to fill in fields.
Step 7: Tracking file generation
The imported project must pass through this skill's built-in scripts/tracking_commit.py init to generate the tracking status in one go; the model is prohibited from writing the final files separately. For the complete fields and commands, see references/tracking-transaction.md. The semantic preparation order is as follows:
-
Import up to chapter: write the last complete chapter N into the initialization transaction's
last_chapter. The tool records in metaimported_through_chapter=N; importing old chapters has no daily-update transaction, and chapter-by-chapter increments must not be fabricated for chapters 1..N, nor may an extra narrative baseline duplicating the current state be generated. -
Core character current snapshots: infer the state as of chapter N of the protagonist, antagonist, and core supporting characters from the book-breakdown output, and write it per character into the initialization JSON's
character_snapshots. The output is generated by the tool to追踪/角色状态/{角色名}.md; for the algorithm see references/character-state-reverse.md。 -
Foreshadowing current line: generate from setup/payoff events with body-text evidence
foreshadow. Each ID keeps only one line for the current state; future designs not yet actually planted stay in the outline and are not written to伏笔.md。 -
Facts and reader knowledge: generate key events into
timeline_events. The same event simultaneously records objective facts, what readers know as of chapter N, and the actual reveal status; a future planned reveal chapter must not be disguised as an already-occurred fact. -
Long-term facts and relationships: write identities/blood relations/kinship/marriage/inheritance/subordination/ownership/permissions/rules/irreversible states that have evidence and would cause errors if unknown later into the initialization JSON's
facts. Each item has a stableK/RID, evidence references, reader reveal status, and prohibited misreadings; guesses without evidence are not imported. -
Continuation state card input: prepare the current position, long-term constraints, active core characters, a brief recap of the last three chapters, the next chapter's promises, and continuity risks.
上下文.mdis generated by the tool as a fixed 7-column layout; do not stuff writing style, file indexes, ordinary to-dos, or QA counts into the continuation state card. -
Execute initialization: detect Python 3 according to the current platform (
python3→python→py -3), and execute:Project
追踪/already contains early files not belonging to the current protocol; no manual cleanup is needed:initwill first move them as-is in their entirety into追踪/_旧追踪存档/, then build the current protocol in place. Old content is retained for the author to consult, does not participate in parsing, and the current state is entirely determined by this import input; on validation failure,initmoves no files.{PYTHON} {story-import skill 根}/scripts/tracking_commit.py init --project {项目根} --input {初始化事务.json} {PYTHON} {story-import skill 根}/scripts/tracking_commit.py check --project {项目根}
Take the demo "Let You Manage the Account, Your High-Octane Mashup Blows Up the Whole Internet" imported up to chapter 10 as an example: the continuation state card must clearly state that Jiang Chen's original phone version of "Gentlemen, Hear the Dragon's Roar" was re-shot in high definition by a professional team, but after viewing the film, senior management believed the new version "lacked soul" and ultimately continued to use the original; Jiang Chen's snapshot should reflect that his military-publicity creative value has been confirmed by Zhou Bosen and Zhang Yaozu; the reader timeline should only include the viewing-meeting conclusion that readers have already seen; if the cultivation arrangement behind Zhong Jiajia "only guessed half right" has not yet been revealed, it can only appear in the author's truth and must not leak into the reader view.
After successful initialization, you should get:
追踪/
├── _tracking-state.json
├── 上下文.md
├── 长期事实.md
├── 关系清单.md
├── 事实档案/
├── 逐章记录/ # 导入旧章不补造文件，续写从第 N+1 章开始
├── 角色状态/{角色名}.md
├── 伏笔.md
├── 时间线/
│ ├── 作者真相.md
│ └── 读者已知.md
When the last chapter of the half-finished product is a draft fragment,last_chapter, character snapshots, and other current semantic checkpoints are all capped at the last complete chapter; the draft-fragment handling policy is written into continuity risks, and unfinished actions are not registered as established facts.
Step 8: Genre positioning generation
Extract the core findings from the book-breakdown report and generate 设定/题材定位.md(according to the structure-mapping-long.md"Genre positioning generation" template format).
设定/题材定位.md 's genre, core hook, and emotion and pacing summary of this book come from 拆文库/{导入书名}/, but these fields are not benchmark registration. Only when Phase 1 has explicitly bound an external benchmark should the "benchmark document list + primary benchmark document" paragraph be appended; the primary benchmark document is limited to at most 1, while secondary benchmarks / reference benchmarks are not limited in number. When not bound, omit the entire benchmark registration paragraph, and do not use {导入书名} as a placeholder. Format:
主对标书: {对标书名} # 必须是独立外部参考作品
对标书列表:
- 书名: {对标书名}
引用强度: 主 # 主 / 辅 / 参考
题材类型: {题材}
相关性: 同题材
用途: 文风+核心结构
- 书名: {书名 B}
引用强度: 辅
题材类型: {题材}
相关性: 同题材/弱相关
用途: {补设定/大纲/模块，不进文风}
- 书名: {书名 C}
引用强度: 辅
题材类型: {题材}
相关性: 同题材/弱相关
用途: {补设定/大纲/模块，不进文风}
- 书名: {书名 D}
引用强度: 参考
题材类型: {题材}
相关性: 同题材/弱相关
用途: {仅按预算召回摘要}
If a quick overview is needed later, a separate "Benchmark Analysis (Derived Summary)" table may be written; this table is not the authoritative registry and must not replace 主对标书 and the complete 对标书列表. All registered items must be traceable to the corresponding 拆文库/{对标书名}/, and must not reference this book's root 设定/。
Step 9: Benchmark structured asset synchronization
This step only handles external reference works explicitly bound in Phase 1. Synchronize the 拆文库/{对标书名}/ structured analysis assets to the project reference view {项目}/对标/{对标书名}/, for story-long-write to read with priority. When no external benchmark is bound, skip this step and do not create an empty directory; it is strictly forbidden to use 拆文库/{导入书名}/ or the project 设定/ as the copy source.
| Source path | Project benchmark path | Purpose |
|---|---|---|
拆文库/{对标书名}/剧情/节奏.md |
{项目}/对标/{对标书名}/剧情/节奏.md |
story-long-write reads key information to advance, emotional touchpoints, and burst rhythm |
拆文库/{对标书名}/剧情/情绪模块.md |
{项目}/对标/{对标书名}/剧情/情绪模块.md |
story-long-write reads reader needs / emotion engine, reproducible modules |
拆文库/{对标书名}/剧情/*.md |
{项目}/对标/{对标书名}/剧情/*.md |
Plot unit and story line reference |
拆文库/{对标书名}/章节/*.md |
{项目}/对标/{对标书名}/章节/*.md |
Matched chapter summaries and key information with expansion technique evidence |
拆文库/{对标书名}/角色/*.md |
{项目}/对标/{对标书名}/角色/*.md |
Character function position reference |
拆文库/{对标书名}/设定/ |
{项目}/对标/{对标书名}/设定/ |
Worldview/faction reference |
拆文库/{对标书名}/拆文报告.md |
{项目}/对标/{对标书名}/拆文报告.md |
Human-readable summary projection |
Missing handling：
- Selected external benchmark is missing
剧情/节奏.mdor剧情/情绪模块.md→ do not register, do not generate a half set of benchmark view; reportmodule_or_rhythm_required_missingand prompt to rerun{对标书名}Stage 3+. This book's core engineering migration is not rolled back because of this./story-long-analyzeOther structured subdirectory missing → prompt according to existing import missing item handling, do not block project creation - Other structured subdirectory missing → prompt according to existing import missing item handling, do not block project creation
Step 10: Writing style synchronization
When the external benchmark has passed Step 9 validation, copy 拆文库/{对标书名}/文风.md to {项目}/对标/{对标书名}/文风.md. Pure copy, no regeneration; skip when no external benchmark is bound.
Missing handling：
- The analysis library has no writing style file (analyze did not run Stage 6) → the import report prompts the user to rerun
/story-long-analyzebefore synchronizing; missing writing style before daily updates will be intercepted by fail-fast - The project benchmark already has an old writing style file → overwrite (latest analysis output takes priority), and inform in the import report
Phase 3-S: Short-form structure migration
Migrate the short-form analysis output of 拆文库/{导入书名}/ into the {短篇标题}/ short-form engineering structure, for story-short-write Phase 3 scene-by-scene writing to take over seamlessly. Migration rules are detailed in references/structure-mapping-short.md。
Short-form engineering is completely different from long-form: the short-form main text is a single file
正文.md(not split into chapters),does not produce追踪/、大纲/、正文/and other long-form directories. During migration, it is strictly forbidden to mistakenly create these long-form-exclusive directories.
Short-form target engineering structure
{短篇标题}/
├── 设定.md ← 含核心框架 + 本书续写基线
├── 小节大纲.md ← 按段-小节结构反推
├── 正文.md ← 单文件全文正文
└── 对标/{对标书名}/ ← 可选：仅外部对标引用视图
├── 拆文报告.md
├── 情节节点.md
└── 写作手法.md
Migration steps
Step 1: Main text migration
Migrate the full text of 拆文库/{导入书名}/原文/ into the single file {标题}/正文.md, and normalize the format according to format-and-structure.md (section markers ###1., only single line breaks between paragraphs, dialogue quotation marks unified according to project/platform conventions).The original text is already a finished draft; do not rewrite the content, only normalize the format.
Step 2: Setting generation
From 拆文报告.md、写作手法.md reverse-derive {标题}/设定.md, containing two blocks:
- Core framework: Align with the story-short-write core framework template (basic information, one-sentence synopsis, core twist, emotional design, character sketch).
- Book continuation baseline: Write the story structure, emotional rhythm, core twist mechanism, and existing writing techniques of the already-written content into the continuation baseline area; this is the book's internal context, not a benchmark summary.
Step 3: Section outline generation
From 情节节点.md 's functional segmentation reverse-derive {标题}/小节大纲.md, mapped by opening section/setup section/escalation section/twist section/ending section; for short stories only make a lightweight blueprint: for each section write 结构段/五段功能, main events, 3-5 sub-events, target emotion, character/relationship changes, causal/logical chain, ending connection/mini hook. When the hook or relationship cannot be determined, mark [待补充], and do not apply the full long-form chapter blueprint.
Step 4: External benchmark reference view (optional)
Only when Phase 1 has explicitly bound an external {对标书名} , sync 拆文库/{对标书名}/ to {标题}/对标/{对标书名}/; if not bound, skip. Do not copy 拆文库/{导入书名}/ wholesale into 对标/。
Phase 4: Project activation
Step 1: Quality check
Compare against the corresponding quality check checklist according to length:
- Long-form: For the complete import quality checklist see the end of references/structure-mapping-long.md (including main text file count comparison, independent snapshots of core characters, author/reader timeline isolation,
tracking_commit.py checkpassed, volume division already confirmed by the user, etc.). - Short-form: For the quality checklist see the end of references/structure-mapping-short.md (including
正文.mdsingle file exists and format is compliant,设定.mdcontains core framework + book continuation baseline, no long-form-exclusive directories mistakenly created, etc.).
Step 2: Missing item prompts
Output the import result summary and items to be supplemented, branching by length.
Long-form import completion report：
=== 导入完成报告（长篇）===
书名：{导入书名}
源文件：{X} 章，{Y} 万字
项目目录：{路径}
## 已生成文件
- 正文：{N} 章
- 角色文件：{M} 个
- 大纲：大纲.md + {V} 个卷纲 + {N} 个细纲
- 追踪：唯一结构化 state + 核心角色独立派生快照 + 伏笔当前视图 + 时间线双视图 + 空逐章记录目录 + 固定 7 栏上下文
- 设定：{世界观文件数} 个
- 外部对标：{未绑定 / 已从 `拆文库/{对标书名}/` 同步到 `对标/{对标书名}/` / 绑定失败及修复动作}
## 待补充项
- [ ] 细纲中的章首/章尾钩子需要补充
- [ ] 题材定位的核心梗三分法需要确认
- [ ] 伏笔追踪中的伏笔已复核
- [ ] `追踪/角色状态/{角色名}.md` 已复核，核心角色各自为截至最后完整章的当前快照
- [ ] `追踪/上下文.md` 顶层区块恰好为新协议规定的 7 个，且不超过 12288 字节
- [ ] `追踪/逐章记录/` 为空；续写从第 N+1 章产生紧凑增量，导入章不补造记录
- [ ] `追踪/时间线/读者已知.md` 未泄露 `作者真相.md` 中尚未揭示的事实
- [ ] `tracking_commit.py check` 通过，`_tracking-state.json` 与全部派生视图一致
- [ ] 卷划分已确认（原文无明确卷界时）
- [ ] `拆文库/{导入书名}/` 未被复制到项目 `对标/`，本书未登记为自身对标
- [ ] 若绑定外部对标，`设定/题材定位.md` 的 `主对标书` 与 `对标书列表` 只包含独立 `{对标书名}`，且同步来源与目录名一致
## 下一步操作
- 运行 `/story-review lean` 审查导入结果
- 运行 `/story-long-write` + "日更" 开始续写
Short-form import completion report：
=== 导入完成报告（短篇）===
标题：{短篇标题}
源文件：{Y} 字
项目目录：{路径}
## 已生成文件
- 正文.md（单文件，{Y} 字）
- 设定.md（核心框架 + 本书续写基线）
- 小节大纲.md（{N} 个小节）
- 外部对标：{未绑定 / `对标/{对标书名}/` 已同步 / 绑定失败及修复动作}
## 待补充项
- [ ] 所有 [待补充] 标记的文件已复核
- [ ] 小节大纲的章首/章尾钩子需要补充
- [ ] 核心反转的铺垫线索已确认
## 下一步操作
- 运行 `/story-short-write` Phase 3 开始续写
Step 3: Project activation
- Set
.active-bookto point to the imported book title/title directory - Confirm the project can be recognized by the corresponding writing skill (long-form → story-long-write, short-form → story-short-write)
- Optional verification: if the corresponding directory in the current runtime has the story-explorer agent deployed (Claude
.claude/agents/story-explorer.md、OpenCode.opencode/agents/story-explorer.md、TRAE Code.trae/agents/story-explorer.md, WorkBuddy project mode.codebuddy/agents/story-explorer.md、Codex.codex/agents/story-explorer.toml), you can call the agent of the same name to cross-verify the integrity of migrated data; TRAE Code uses the built-inAgentto selectstory-explorerand pass in项目目录：{dir}\n查询类型：progress\n查询参数：导入验证, Claude/OpenCode can use the equivalentsubagent_typecall, Codex usesagent_type. WorkBuddy project mode usesAgent(subagent_type: "story-explorer", ...); plugin-only mode only uses this value when the registry actually returnsoh-story:story-explorer.
The setup environment check has already been completed in Phase 1 "Environment check pre-step", so it is not repeated here.
Large work handling (>200 chapters)
This section applies only to long-form import. Short-form is a single-file full migration, with no incremental import requirement.
For works exceeding 200 chapters,the decomposition can be done in batches, but tracking initialization must cover all already-written chapters in one pass：
- Batch decomposition: The first phase only deeply analyzes the first 50 chapters + an overview of the whole book, and subsequently more chapters are analyzed as needed until
拆文库/。 - Tracking in one pass: Initialize the transaction's
last_chapterwritethe last fully written chapter number N, not the 50 analyzed in the first phase.imported_through_chapteris written once byinitand is not advanced afterward; per-chapter transactions only accept chapter numbers starting from N+1; chapters 1..N do not have fabricated per-chapter records, and continuation writing starts from N+1. If it was mistakenly written as 50 during init, chapters 51..N can still beappendappended per chapter (one transaction per chapter, chapter numbers must be consecutive), but transactions must be constructed chapter by chapter for the old chapters already written; do not delete追踪/and start over—_旧追踪存档/is also in there. - Context summary: Generate simplified summaries (200 characters/chapter) for chapters not deeply analyzed, for use in reverse-inferring the current state.
Reference index
Load by phase, not all at once.
All reference files bundled with this skill are located in references/, and are loaded by scenario. When methodology/templates from other skills are involved, story-import does not directly load the files, but instead runs the corresponding /命令 and that skill loads it itself.
Phase 1: Confirm the import source
| Scenario | Load file |
|---|---|
| Length triage determination | references/length-routing.md |
| Chapter format recognition | Handled by Phase 1 of the story-long-analyze analysis pipeline (run /story-long-analyze) |
Phase 2: In-depth analysis
| Scenario | Load file / related skill |
|---|---|
| Long-form in-depth analysis (methodology, quality checks, and output templates are all bundled) | Run /story-long-analyze to invoke the long-form analysis pipeline |
| Short-form in-depth analysis (methodology, quality checks, and output templates are all bundled) | Run /story-short-analyze to invoke the short-form analysis pipeline |
Phase 3: Structural migration
| Scenario | Load file |
|---|---|
| Long-form migration mapping rules | references/structure-mapping-long.md |
| Short-form migration mapping rules | references/structure-mapping-short.md |
| Character state reverse-inference rules (long-form) | references/character-state-reverse.md |
| Character state rules (character-state-reverse.md dependency) | references/state-tracking.md |
| Short-form main text format specification | references/format-and-structure.md |
For the long-form detailed outline template format, see story-long-write (Phase 3 detailed outline section); for the short-form core framework template, see story-short-write (core framework section). These two items are plain-text guidance; story-import does not load the files of the corresponding skills.
Phase 4: Project activation
| Scenario | Description |
|---|---|
| Long-form project structure specification | See story-long-write (Phase 4 project file structure) |
| Short-form project structure specification | See story-short-write (Phase 3 project structure) |
| Environment deployment | The deployment template is provided by /story-setup , and story-import is not responsible for deployment |
Process handoff
Pipeline: Long-form / Short-form Location: import (before starting the book)
| Timing | Jump to | Command |
|---|---|---|
| After import, want to continue writing (long-form) | story-long-write | /story-long-write + "Daily updates" |
| Want to continue writing after import (short story) | story-short-write | /story-short-write |
| Want to review quality after import | story-review | /story-review |
| Want to deeply analyze benchmarks (long-form) | story-long-analyze | /story-long-analyze |
| Want to deeply analyze benchmarks (short story) | story-short-analyze | /story-short-analyze |
| Start a new book from scratch (long-form) | story-long-write | /story-long-write + "Start book" |
| Start a new book from scratch (short story) | story-short-write | /story-short-write |
| Project environment not deployed | story-setup | /story-setup |
Language
- Reply in the user's language; use whatever language the user uses
- Chinese replies follow the 《Chinese Copywriting Guidelines》