ZenStory AI

Practical guide · 实用指南

Learn writing craft from fiction: evidence, scene mechanisms and a new-story exercise

怎样拆小说才能学会写：从原文证据、场景机制到自己的新故事

Source checked 2026-09-12 · 源码核对日期

A useful fiction analysis explains how a passage changes what a reader knows, expects or sees a character choose. Separate observable details from your interpretation, then state when a technique could work in a different story and when it would fail. A plot summary or a list of praise cannot do that work. Oh Story provides separate short- and long-fiction analysis workflows; analysis-only work is distinct from importing a manuscript for continuation. The original poster-proof example below demonstrates one narrow mechanism and an unsupported reading, then compares two new garden scenes with the same facts and outcome but different dramatic emphasis. It is not a model run, a full-book verdict or evidence of reader demand.

有用的拆文要说明：一段文字怎样改变读者知道的事、期待的结果，或人物能作的选择。先分开看得见的细节与自己的解释，再说这种手法换到别的故事里，什么条件下成立、什么情况下会失效。剧情复述和一串夸奖代替不了这一步。Oh Story 分别提供短篇与长篇拆文流程；只做分析和导入可续写工程是不同任务。下面用原创海报校样片段示范局部机制与缺乏依据的读法，再比较事实、结局相同而戏剧重心不同的两套菜园场景；这不是模型运行记录、全书结论或读者需求数据。

Oh StorySource on GitHub

Before you start · 开始之前

Steps · 操作步骤

  1. Choose analysis, not an accidental import For analysis-only work, use /story-short-analyze for a complete short piece or /story-long-analyze for a long work. /story-import serves a different goal: building a continuable writing project. The short workflow is a full analysis pipeline, not a command that promises to stop after your favorite paragraph. For the small exercise below, use its ordinary-chat request instead.
  2. Anchor each observation before naming a technique Record a short quotation or an accurate description with its paragraph or chapter location. Then describe the state before and after: who can act, who knows what, and which condition remains unresolved? “Tense dialogue” names an impression; the unsigned approval box and the promise of a new proof show why printing has not been authorized.
  3. Make the interpretation narrower than the evidence The poster passage supports a reading in which an omitted credit changes the approval process. It does not establish deliberate theft, a lifelong rivalry or healed trust. Name an alternative explanation when it matters. A plausible reader effect is an interpretation of the text, not a measured response from all readers or a claim about the author’s private intention.
  4. Extract a mechanism with conditions and failure cases One possible mechanism is that withholding a small, necessary cooperation makes an overlooked contribution consequential. It works here because approval is still needed and the disputed condition can still change. If the printing is already finished, or refusal has no practical effect, the same gesture will not do the same narrative work. A pause elsewhere may simply be hesitation; do not force every scene into this pattern.
  5. Change the conflict, not just the names Write a new exercise with a different desired outcome, relationship, cost and decision. Keep only the abstract mechanism you chose to practice. The garden example uses an openly stated availability limit and a smaller shared plan, not a hidden credit, a silent signature gesture and a replacement proof. This is a craft exercise, not a legal originality test or a shortcut to copying a passage’s wording. Compare the two garden options by the experience you want: acceptance of a stated limit or release of an unconfirmed workaround. More beats are not automatically better, and a surprise exception would change the agreed story facts.
  6. Can I study only the first 20 chapters? You can define a limited reading question, but distinguish it from a documented pipeline checkpoint. The long workflow normally builds the source’s chapter map, analyzes the first three chapters and pauses at 快速预览.md; an explicit full-analysis request continues after that preview. For a first-20-chapter study, label the authorized material as an excerpt and name what you want to learn from that range. A separate ordinary-chat reading brief can stay scoped to that material; it is not a demonstrated “complete first-20 analysis” mode, and twenty chapters are not guaranteed to fit a host’s context. Ask which chapters were actually read and keep any gaps visible. Do not call a novel excerpt a complete short story merely to use the short-analysis workflow.
  7. Can I use staged results before the whole book is analyzed? For questions about the passages actually covered, yes. That does not make the partial result a complete full-book analysis. For a supplied full novel, Stage 2 targets all indexed chapters, while full-book aggregation and reporting refer to the resulting chapter summaries and outputs. The large-book reference’s 10–20-chapter batches specifically describe Stage 3–5 processing of chapter summaries, not an automatic stop after reading twenty original chapters. Saved partial summaries can inform a question about the passages they actually cover. They do not establish a thousand-chapter novel’s final character arcs, cross-book clue payoffs or ending. Keep a range-limited reading note separate from a claimed complete analysis bundle; the cited contracts do not establish arbitrary-window completion and later seamless merging.
  8. What actually resumes after an interruption? Resume uses the saved _progress.md, not an agent’s promise to remember. The current contract requires schema_version: 2 and the chapter-boundary table. A paused_after_stage1 checkpoint resumes at Stage 2; other interrupted blocks resume from that block’s starting chapter and overwrite that block’s existing outputs. Preserve author notes separately before a repeat pass. For work that ran with failures, inspect the recorded gaps rather than treating a finished-looking report as complete evidence. A valid checkpoint describes progress on that source; it does not turn independently truncated inputs into one verified full-book reading.
  9. Carry selected lessons into a separate writing decision Choose which lesson serves your own premise and say what must not transfer: names, plot facts, distinctive props, reveal order or another work’s voice. A practice note is not a completed Oh Story analysis bundle; do not relabel it as one. The documented short-writing flow can read the full analysis bundle, including source text, so do not assume only your abstracted lesson will be used. Set the new story’s boundaries explicitly; the short-story planning guide helps turn that decision into an independent plan.
  1. 先选只拆文，不误走导入 只要分析时，完整短篇可用 /story-short-analyze，长篇可用 /story-long-analyze。/story-import 的目标不同，是建立可续写的写作工程。短篇拆文是完整分析管道，不是保证只读你最喜欢的一段就停的命令；下面的小练习应采用单独的普通聊天请求。
  2. 先定位观察，再给手法命名 记录短引或准确转述，并标明段落、章节位置。随后写清前后状态：谁能行动、谁知道什么、什么条件仍未解决？“对白紧张”只是印象；未签的确认格与重出校样的承诺，才说明为什么印制还没得到确认。
  3. 让解释不超出证据 海报片段支持“署名遗漏改变了确认流程”这一读法，却不能证明蓄意侵占、终身宿怨或信任已经修复。必要时列出另一种解释。可能的阅读效果是对文本的理解，不是对所有读者的实测，也不是作者真实意图的证据。
  4. 提取机制时，一起写条件与失效情形 一种可选机制是：暂停一项虽小却必需的合作，让被忽略的贡献真正影响下一步。这里成立，是因为确认仍然必要，争议条件也还能修改。如果海报已经印好，或拒绝配合根本没有后果，同一个动作就不承担相同作用。别处的停顿也可能只是犹豫，不能把每场戏都塞进这一模式。
  5. 换冲突，不只是换名字 新练习要有不同的目标、关系、代价与决定，只保留选中的抽象机制。菜园练习用公开说明的时间限制和缩小共同计划，替换署名缺失、无声挡住签字与重出校样。这是写作练习，不是法律上的原创判定，也不是沿用原段措辞的捷径。 比较两套菜园方案时，先选想写的体验：接纳明确限制，还是放下尚未成立的补救。节拍更多不自动更好，临时开一个意外例外则会改变已定事实。
  6. 能不能只研究前20章？ 可以限定精读问题，但要与文档中的管道停靠点分开。长篇流程通常先建立原文章节边界表，分析前三章，在 快速预览.md 停靠；一开始明确要求完整拆解时，则保留预览后继续。只想研究前20章，就把有权使用的材料标成节选，说清想从这个范围学什么。另行提出的普通聊天精读请求可以限定在这些材料内，但这不是已经证明存在的“前20章完整拆解模式”，也不保证20章一定装得进宿主上下文。要求说明实际读到了哪些章，缺失保持可见；不要为了改用短篇管道，就把长篇节选叫成完整短篇。
  7. 分析未完成，阶段性成果能先用吗？ 可以先研究实际覆盖的段落，但局部成果不等于已经完成的全书分析。对提供的完整长篇，Stage 2 的目标是覆盖已索引的全部章节，全书聚合和报告则引用相应摘要与前序产物。大书参考中的10–20章一批，具体指 Stage 3–5 对章节摘要的处理，不是读完20章原文便自动停止的设置。已保存的局部摘要可以帮助回答其实际覆盖段落中的问题，但不能确定千章长篇最终的人物弧、跨卷伏笔回收或结尾。范围有限的精读笔记应与声称完整的拆文资料分开；所引契约没有确立“任意区间完整分析、之后无缝合并”的保证。
  8. 中断后，真正接续的是什么？ 续跑读取保存的 _progress.md，不靠 Agent 口头说记得。当前契约要求 schema_version: 2 与章节边界表。paused_after_stage1 从 Stage 2 继续；其他分块中断则从该块起始章恢复，并覆盖该块已有产出。重跑前，把作者另加的笔记单独保留。若过程有失败，应查看记录中的缺口，不把外观完整的报告当成完整证据。有效断点说明同一份材料的处理进度，不会把各自截短的输入自动变成一份已确认覆盖全书的阅读。
  9. 把选中的收获变成另一项创作决定 决定哪条收获适合自己的梗，并说明不迁移哪些东西：人名、情节事实、独特道具、揭示顺序或另一部作品的声音。练习笔记不是完整的 Oh Story 拆文产物，不要改个文件名就冒充。文档中的短篇写作会读取包括原文在内的整套拆文资料，不能假设它只使用你提炼的抽象方法；新故事边界仍须明确。短篇规划指南可以帮助把这个决定变成独立设计。

Example · 示例

ORIGINAL PASSAGE AND READING EXERCISE — written for this guide; not a skill run or a quotation from another work.

The proof:
[P1] At the poster workshop, Lu Qing handed Zhou Mian the final proof. “Once you sign off, I’ll place the order.”
[P2] Zhou pulled a comparison copy from her notebook. Both their handwritten revisions covered it. At the bottom of the new proof, only Lu’s name appeared.
[P3] “Another typo?” Lu asked. Zhou laid her pen across the approval box without signing.
[P4] Lu looked from their revision marks to the credit line. After a pause, Lu changed the credit line to list both names side by side. “I’ll have a fresh proof made like this. You’ll see it before printing.”
[P5] Zhou picked up her pen, wrote “Awaiting new proof” on the back, and left the approval box empty.

Evidence → reading → limit:
• P1 makes Zhou’s approval a stated condition of ordering. P3 and P5 show that approval has not been given. The passage ends with a proposed next step, not an order placed or a poster printed.
• P2 juxtaposes both people’s revision marks with one printed name. Together with P3–P4, that supports reading the credit line as the immediate point of disagreement. It does not tell us who made the omission or whether it was deliberate. A layout oversight is one possible alternative to deliberate exclusion; the passage does not settle the cause.
• P4 offers a visible repair; P5 reserves the final decision until a new proof exists. This can make conditional cooperation legible to a reader. “Zhou forgives years of exploitation” invents both the history and the forgiveness.

Working lesson: a small cooperation can matter when another action genuinely depends on it. Preserve the unmet condition long enough for the response to have a cost or a limit. This is one reading of this scene, not a rule that every ending must withhold approval.

Different-situation exercise — a proposed new design, not facts inferred from the passage:
A community garden wants to plant a second bed. Wen has covered three consecutive Sundays and cannot cover the next one. The proposed expansion assumes Wen will cover that next Sunday too, without anyone having asked. During the meeting, Wen says this directly; nobody discovers a missing name or silently blocks a signature. The group must find another volunteer or leave the second bed unplanted. Design a short scene in which they choose the smaller plan. Give another member a credible reason to want expansion and let Wen care about the garden without promising unavailable time.
What transfers: a contribution previously treated as automatic becomes a real constraint on a shared goal.
What changes: the conflict is future workload rather than credit; the limit is stated aloud; the result is a reduced plan rather than a replacement proof. Do not reuse the original dialogue or its sequence of props and gestures.

TWO AUTHORED SCENE OPTIONS — alternatives for the garden exercise, not consecutive scenes or recorded model outputs.

Shared new design decisions for both options: this meeting is on Wednesday; extra seedlings must be confirmed with the nursery by Thursday for Sunday planting; the nursery cannot hold them without that confirmation. Care for the existing bed is already covered by another volunteer; the second bed needs additional Sunday coverage. Wen is unavailable that whole Sunday. Lin wants the extra bed so two waiting neighbors can join. Qiao might help, but cannot confirm Sunday availability until Friday. These are new story facts chosen for the exercise, not facts recovered from the poster passage. No order for the extra seedlings has been placed.

Option A — make the boundary the turning point:
1. Wen states the Sunday limit at the start. Lin names the disappointment: two neighbors have been waiting, and the expansion would give them space. Both concerns remain legitimate.
2. Lin accepts that there is no confirmed cover in time for the order but asks Wen to explain the smaller plan to the neighbors, thinking a direct explanation will be clearer. Wen agrees to join the conversation, not to describe an unpromised shift as a broken commitment.
3. The group chooses the existing bed only. Lin recognizes that the expansion counted on time nobody had secured and decides to lead the explanation: it was a gap in the group’s plan, not Wen cancelling a promised shift.
4. End with Lin telling the neighbors that the group planned more work than it could staff, so there will be no second bed this Sunday. Wen remains part of the garden without agreeing to cover it. The choice is settled; neither the disappointment nor all future staffing is magically fixed.
Intended emphasis: a clear limit can become part of a shared decision rather than a private refusal. Keep the scene short once that decision is made. The risk is making Lin concede too easily; give the loss of the extra places room in the exchange without turning Lin into a villain.

Option B — let a tempting workaround expose the same constraint:
1. The meeting begins with Qiao’s possible help. Lin suggests reserving the extra seedlings on that basis; Wen then states the Sunday limit. The limit is still explicit, not hidden until a final reveal.
2. Put the two dates in the conversation: the nursery needs Thursday confirmation, while Qiao can answer only on Friday. Qiao does not promise a shift that is still uncertain. In this design the nursery cannot hold the additional seedlings without that confirmation.
3. Wen asks who would actually cover Sunday if Qiao cannot. The group has no confirmed replacement. Rather than rely on Wen or announce a commitment nobody can keep, Lin withdraws the additional seedling request before placing the order.
4. Lin places the order for the existing bed only and gives the neighbors the same smaller-plan update. The tempting solution has been considered and rejected on its conditions, not defeated by an unrelated accident.
Intended emphasis: goodwill and a plausible workaround are not the same as an available contribution. The risk is letting the scene become a calendar explanation; keep each date tied to someone’s proposal, response and choice. If the workaround adds no meaningful conflict to your story, use the shorter option instead.

Choosing between them:
• Choose A if the main movement is whether the group will accept Wen’s stated limit and share responsibility for the disappointment.
• Choose B if the main movement is letting go of a comforting but unconfirmed solution. Its extra exchange earns space only if that hesitation matters for these characters.
• Both preserve the same agreed outcome and Wen’s availability. Do not quietly make Qiao available on Thursday, move the nursery deadline, or have Wen cover “just this once” to produce a tidier ending. A different outcome is possible, but it needs an explicit change to the brief and its consequences, not a covert exception.
• Neither option is a full story or a universal recipe. Keep the one that serves your own characters and desired experience; discard the borrowed mechanism if it pushes your story toward a conflict you do not want.

Ordinary-chat request for this small exercise — no analysis skill invocation:
Read only P1–P5 above. In chat, give three claims with paragraph evidence, one plausible alternative reading and two things the passage does not establish. Do not invent a full story or produce a complete analysis bundle. Then compare the two garden options above against the shared facts and the experience I want to create. Recommend one with a concrete tradeoff, or propose a clearly labeled third option without changing the agreed outcome or borrowing the poster scene’s wording and beat sequence. Mark any additional premise as a decision for me, not an established fact. Do not write files, import a book, start a full-book workflow or draft prose. Stop after the options.

A range-limited reading brief — ordinary chat, not a partial-pipeline command:
I am providing only chapters 1–20 of [authorized novel excerpt], not the complete novel. Study how the opening establishes the protagonist’s immediate goal and presents one unanswered question. State which supplied chapters you actually used. Separate observations, possible interpretations and what chapters 21 onward would be needed to establish. If the material is too large to read reliably here, say so and help narrow the question and material. Do not invent the ending, claim full-book coverage, run later analysis stages on my behalf or write files. A useful result is a bounded reading note, not a complete Oh Story analysis bundle.

For example, “this clue is still unresolved within chapters 1–20” may be supported; “the author forgot the clue in the entire novel” is not supported by that range. You can try a newly designed scene based on a local technique without pretending that you have already mapped the reference novel’s whole structure.

For an actual complete short-story analysis in an installed Claude Code workflow, use /story-short-analyze with [the authorized complete text path] and a distinct output location. That invokes the documented full pipeline, not the chat-only exercise. Request analysis only and no subsequent writing or import. Existing analysis under the same title has its own archive/resume/cancel decision; do not silently overwrite it.
原创片段与精读练习——为本指南撰写，不是技能运行记录，也不是引用他人作品。

《校样》：
[P1] 海报工作室里，卢青把最后一版样张递给周眠：“你确认后，我就下单。”
[P2] 周眠从标注本里抽出对照稿，上面留着两人各自的修改笔迹。新样张底部，只有卢青一个人的署名。
[P3] “正文还有错？”卢青问。周眠把笔横放在确认格上，没有签。
[P4] 卢青从两人的修改笔迹看向署名栏，停了一会，把署名改为两人并列：“按这版重出样，印之前再让你看。”
[P5] 周眠拿回笔，在背面写了“待看新样”，把确认格留空。

证据 → 读法 → 边界：
• P1 把周眠确认设为下单条件，P3 与 P5 显示她尚未确认。片段结束时只有下一步的提议，没有下单或印成海报。
• P2 把两人的修改笔迹与单人署名并置，结合 P3–P4，可以把署名看作眼前争议点。但不知道是谁遗漏，也不能判定是否故意。排版疏漏是不同于故意排除的一种可能解释；片段并未确定原因。
• P4 提出可见的修补办法，P5 将最终决定保留到新样张出现后。这能让读者看见有条件的合作。“周眠原谅了多年的剥削”则同时编造了前史与原谅。

暂取的收获：当另一项行动确实依赖合作时，小小的配合也可以改变局面。让尚未满足的条件保持有效，对方的回应才会有代价或边界。这只是本场景的一种读法，不是要求每篇都以暂不确认作结。

不同情境练习——以下是新的设计选项，不是从片段反推的事实：
社区菜园想增种第二块地。温宁已经连续代班三个周日，下周无法再代班。大家提出扩种时，仍把温宁下周代班算作已有的人手，尚未问过她。会上，温宁直接说清这件事；没有人发现漏掉的署名，也没有人默默拦住签字。大家必须另找志愿者，或暂时不种第二块地。设计一场最终选择缩小计划的短戏：让另一位成员有可信的扩种理由，也让温宁在关心菜园的同时，不承诺自己拿不出的时间。
迁移的是：原本被当成理所当然的投入，成为共同目标必须面对的实际限制。
改变的是：争议从署名转为未来分工；限制直接说出；结算是缩小计划，而非更换校样。不沿用原对白或那串道具、动作顺序。

两套作者自拟的场景方案——用于上面的菜园练习，二选一，不是连续两场，也不是模型运行输出。

两案共同确认的新设计：会议在周三；要在周日增种，必须于周四向苗圃确认追加苗木，苗圃无法在未确认的情况下保留这批苗。原有菜地的照料已由另一名志愿者承接，第二块地需要额外的周日人手。温宁整个周日都无法值守。林澄希望扩种，让两位等候的邻居也能参与。乔越可能帮忙，但要到周五才能确认周日是否有空。以上都是本练习新定的故事事实，不是从海报片段中还原的内容。追加苗木尚未下单。

方案 A——把接纳边界作为转折：
1. 温宁一开场便说清周日无法值守。林澄说出缩小计划的损失：两位邻居已经等了一阵，扩种才有他们的位置。双方在意的事都有理由。
2. 林澄接受下单前没有确定替班人的事实，却请温宁向邻居说明缩小计划，觉得本人解释更清楚。温宁愿意一起去谈，但不把从未答应的班次说成自己违约。
3. 大家选择只种原有菜地。林澄意识到扩种方案用了尚未落实的人手，决定由自己先说明：问题是共同计划的缺口，不是温宁取消了已承诺的班次。
4. 收在林澄告知邻居：大家规划的工作超过了能落实的人手，本周日不增种第二块地。温宁仍参与菜园，但没有答应值守。选择已定，失望与未来所有排班并没有一起消失。
想突出的是：明确的个人限制怎样进入共同决定，而不只停留在一次私人拒绝。决定之后不必拖长。风险在于林澄让步过快；给少了两个参与名额的损失留一点回应空间，但不把林澄写成反派。

方案 B——让看似可行的补救暴露同一限制：
1. 会议先讨论乔越可能帮忙。林澄据此提议预订追加苗木，温宁随后说清周日无法值守。限制仍然明说，不藏到最后才揭晓。
2. 把两个日期带进对话：苗圃周四要确认，乔越周五才能答复。乔越不把尚未确定的时间说成已承诺的班次。本设计中，苗圃也不能在未确认的情况下保留追加苗木。
3. 温宁问，如果乔越最终来不了，周日究竟由谁值守？大家没有确定的替代者。林澄不再押注温宁，也不宣布没人能承担的承诺，而是在下单前撤去追加苗木的请求。
4. 林澄只为原有菜地下单，并向邻居说明同样的缩小计划。看似可行的办法是在自身条件下被考虑、被放弃，不是被临时塞入的事故打断。
想突出的是：善意和听上去可行的办法，不等于真实可用的投入。风险在于写成日历讲解；每个日期都应牵动某人的提议、回应与选择。如果这次补救并未给自己的故事增加有意义的冲突，就用较短的方案。

怎样选：
• 核心变化若是大家是否接纳温宁说清的限制、共同承担解释失望的责任，可选 A。
• 核心变化若是放下一个让人安心却尚未成立的办法，可选 B。多出来的交锋，要对这些人物的犹豫有用才值得保留。
• 两案保留同一个已定结果和温宁的时间边界。不为收得漂亮就悄悄让乔越周四有空、挪后苗圃期限，或让温宁“就再顶这一次”。可以另选结果，但要明确修改创作要求并承担后果，不靠暗中开例外。
• 两案都不是完整故事或通用公式。保留服务自己人物与目标体验的那一案；如果借来的机制把故事推向了不想写的冲突，就放弃这个机制。

适用于本小练习的普通聊天请求——不调用拆文技能：
只读上面的 P1–P5。在聊天中给出三条带段落依据的判断、一种合理的替代读法，以及两件文本尚不能证明的事。不补全整篇故事，不生成完整拆文资料。再依据共同事实与我想写的体验，比较上面两案，带着具体取舍推荐其中一案；也可明确标注第三种设计，但不改变已定结果，不借用海报片段的措辞或情节节拍。进一步新增的前提标为需要我决定的内容，不当成已成立事实。不写文件、不导入书、不启动全书流程、不写正文。列出方案后停止。

范围有限的精读请求——普通聊天，不是启用局部管道的命令：
我只提供 [有权使用的小说节选] 第1–20章，不是完整小说。请研究开篇怎样建立主角的眼前目标，以及怎样呈现一个尚待解答的问题。说明实际用到了哪些提供的章节，分开观察、可能的读法，以及必须读到第21章以后才能确定的事项。若材料大到无法在这里可靠读完，请说明并帮我收窄问题与材料。不补结尾、不声称覆盖全书、不替我启动后续分析阶段、不写文件。我要一份范围明确的阅读笔记，不是完整的 Oh Story 拆文资料。

例如，“这条线索在第1–20章内尚未解答”可能有依据；“作者在整部小说里忘了这条线索”则不能只凭这个范围判断。可以把局部手法用于另行设计的场景，但不用假装自己已经掌握了对标小说的全书结构。

如果实际要在已安装的 Claude Code 环境中完整拆一篇短篇，可用 /story-short-analyze，提供 [有权使用的完整原文路径] 和独立输出位置。这会进入文档规定的完整管道，不是上面的聊天小练习。明确只做分析，不继续写作或导入。同名已有拆文还有归档、续跑或取消的选择，不应静默覆盖。

Expected files · 预期文件

Verify the result · 验证结果与边界

Sources and version notes · 来源与版本说明