produce-long-form-novel
Produce, maintain, analyze, and export long-form Chinese novels through staged, editable artifacts, with progressive confirmation of high-impact creative settings for novice authors. Use when Codex needs to turn an idea into a novel plan, guide choices such as audience channel or genre, analyze public ranking metadata for hot genres, deconstruct an authorized reference novel or diagnose a manuscript, write fanfic from an authorized analysis workspace, design or revise volumes, generate chapter plans or prose, continue an existing novel, audit and repair chapters, export ready chapters as a TXT file, or decide the next production step from an existing Markdown/YAML workspace. Do not use for generic app development, database operations, unauthorized bulk copying, or unrelated short-form copywriting.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/explosivecoderflome--ani-book-skill)HTML
<a href="/listings/explosivecoderflome--ani-book-skill"><img src="/api/listings/explosivecoderflome--ani-book-skill/badge.svg" alt="findsafeskills safety rating" /></a>Install
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 283 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: produce-long-form-novel
- description: Produce, maintain, analyze, and export long-form Chinese novels through staged, editable artifacts, with progressive confirmation of high-impact creative settings for novice authors. Use when Codex needs to turn an idea into a novel plan, guide choices such as audience channel or genre, analyze public ranking metadata for hot genres, deconstruct an authorized reference novel or diagnose a manuscript, write fanfic from an authorized analysis workspace, design or revise volumes, generate chapter plans or prose, continue an existing novel, audit and repair chapters, export ready chapters as a TXT file, or decide the next production step from an existing Markdown/YAML workspace. Do not use for generic app development, database operations, unauthorized bulk copying, or unrelated short-form copywriting.
- ---
- # Produce Long-Form Novel
- ## Mission
- Help a novice author move from a vague idea toward a complete long-form novel. Produce concrete, editable artifacts; preserve author decisions; and recommend one clear next step.
- ## Codex-Native Execution Boundary
- **Codex is the only engine for creative understanding, planning, generation, review, and judgment.** This Skill is the process contract; Python scripts perform only deterministic state transitions, validation, indexing, conflict detection, and export. This repository is not an `AI-Novel-Writing-Assistant` runtime or submodule: do not add provider SDKs, Web APIs, database authorities, queues, or a custom agent runtime. `provider` and `model` fields are Token diagnostics only when the Codex host exposes them, never a generation dependency.
- ## Check Skill Currency Before Running
- Treat the checked-out repository as the editable Skill source and the locally installed Skill as its runtime mirror. Resolve the installed mirror from `$env:CODEX_HOME\skills\produce-long-form-novel`; when `CODEX_HOME` is unset, use `<UserProfile>\.codex\skills\produce-long-form-novel`. Before the first production action in a task, compare the repository root with that mirror:
- ```powershell
- python scripts/sync_skill_mirror.py check <repository-root> <installed-skill-directory>
- ```
- - Report `missing`, `changed`, and `extra` files before relying on the mirror.
- - After every functional edit to `SKILL.md`, `references/`, `scripts/`, `assets/`, `requirements.txt`, or `agents/`, run the repository's required compile, tests, and Skill quick validator. When they pass, automatically synchronize the repository to the resolved installed mirror and rerun `check`; do not ask for repeated authorization.
- - Never synchronize failed or unvalidated functional changes. Keep synchronization additive and preserve mirror-only files. Do not hardcode a user-specific absolute path.
- - This comparison and synchronization are deterministic maintenance operations; do not record them as creative model usage.
- When a user explicitly says they have no story idea, do **not** begin with the three opening settings or a full premise. Read [opening-seeds.md](references/opening-seeds.md), run its private divergence and quality-gate process, then output exactly five non-authoritative, one-sentence opening seeds. Ask the user to choose one, edit one, or provide an original idea.
- For a vague request without an explicit “no idea” statement, start with three high-impact choices: audience channel, publishing shape, and genre or primary reader reward. Do not ask about protagonist, style, or length in that first round or generate a full hook package. These choices remain provisional until the user confirms them or delegates the decision.
- When the user supplies an initial idea or selects one opening seed and still needs direction exploration, generate exactly two non-authoritative new-book brief previews before formal planning. If the user has already selected a concrete direction, supplied a complete brief, or explicitly delegated direct production, reuse those decisions and proceed to the next authorized artifact; do not reopen direction selection. Each preview must include a working title, one-sentence premise, target reader/reward, protagonist path, core conflict, early hook, progression loop, and high-level ending direction. The two previews must differ materially in their selling point, conflict, protagonist path, progression loop, tone, or ending direction. Do not write either preview to `novel-brief.md`, mark it `ready`, or use it for world, volume, or chapter production until the user selects, combines, or delegates a direction.
- Use Markdown for creative artifacts and a small `novel-state.yaml` file for progress. In a legacy workspace, keep continuity ledgers as readable Markdown. In a migrated workspace, YAML is the sole continuity authority and Markdown ledgers are generated read-only views; SQLite is a disposable local index. Do not require JSON unless the user explicitly requests machine integration or export.
- ## Choose Persistence Mode
- Choose one mode before producing files:
- - **Preview mode**: Return a bounded artifact in the conversation. Use for exploration, first-draft ideas, or any request without a confirmed save location. Do not create files or infer a workspace from chat history.
- - **Workspace mode**: Write Markdown artifacts and a small YAML state index into one workspace. Use when the user asks to save, create a workspace, analyze a long source, retain ranking snapshots, continue across tasks, protect edits, generate chapters, or otherwise confirms durable production. Novel production uses `novel-state.yaml`; reference-book analysis uses `analysis-state.yaml`; ranking-trend analysis uses `trend-state.yaml`.
- - **Promote preview**: When the user accepts two or more planning artifacts and wants to continue, recommend workspace mode once. After the user confirms, save the accepted artifacts first, create the state index, and resume from the recorded next action.
- Do not create a workspace merely because a conversation has multiple turns. If workspace mode is confirmed but no location is supplied, use `novels/<normalized-title>/` for novel production, `analyses/<normalized-title>/` for reference-book analysis, or `trends/<normalized-scope>/` for ranking-trend analysis under the current working directory, then state the chosen path before writing.
- ## Route the Request
- Choose one primary route before acting:
- 1. **Create a novel**: first route a vague idea through progressive confirmation; only after confirmation or AI delegation, turn it into positioning, a hook package, a story engine, and the first planning artifacts.
- 2. **Plan or revise volumes**: create or update volume strategy, skeletons, beat sheets, or chapter lists.
- 3. **Plan or draft a chapter**: create the chapter contract, draft prose, assess it, and update stable state.
- 4. **Audit or repair**: diagnose a supplied plan or chapter, apply the smallest useful repair, and preserve successful content.
- 5. **Continue an existing novel**: inspect current artifacts, recover the next valid production step, and continue without rewriting protected work.
- 6. **Analyze a reference work**: deconstruct an authorized local text, accessible online source, or the user's manuscript into evidence-backed notes, analysis sections, and reusable pattern cards without reproducing the source.
- 7. **Write fanfic from an analysis**: use an authorized analysis workspace as canon facts, keep original names within the read scope, and produce a fanfic novel workspace with a handoff before chapter drafting.
- 8. **Analyze hot genre trends**: inspect public official chart metadata or user-provided chart captures, aggregate surface signals, and produce a chart-level report plus 3–5 opportunity cards without reading novel prose.
- 9. **Establish or adjust a writing style**: turn author preferences, the author's own writing, or transferable methods from authorized analysis into a small editable style contract; bind it to a novel only after confirmation or delegation, then use it for prose, review, and repair.
- Read [workflow-routing.md](references/workflow-routing.md) when the request spans routes, the next step is unclear, or an existing workspace may be incomplete.
- Read [novel-brief.md](references/novel-brief.md) when creating or revising a novel brief. Keep its reader, length, narrative and writing-preference settings authoritative for later planning and chapter production.
- Read [fanfic.md](references/fanfic.md) when the user asks for fanfic, a continuation that keeps original names, or a new story that reuses an analyzed cast. Original novels still must not copy source names, recognizable appearance sets, or signature scenes from analysis cards.
- Read [style-engine.md](references/style-engine.md) when the user asks to extract, save, reuse, or change a writing style, or to apply anti-AI expression rules. Keep style optional during onboarding: a confirmed book-level `writing-style.md` can be created from author preferences, the author's own writing, or transferable methods from authorized analysis. Do not copy a third-party author's recognizable voice. Use the same effective style rules for drafting, humanization, review, and targeted repair; style cannot override story facts or chapter obligations.
- Read [remote-skill-sources.md](references/remote-skill-sources.md) when the user asks to obtain or update a writing-style or anti-AI Skill from GitHub. Treat `production/skill-sources.yaml` as an explicit, private source registry. Validate it with `scripts/skill_sources.py`, fetch candidates only on the user's request through the preinstalled `skill-installer`, pin an immutable commit before enabling, and require review plus author confirmation before merging rules into `writing-style.md`. Never fetch during ordinary chapter production, execute remote scripts, or let a remote `SKILL.md` override this Skill's contracts.
- ## Guide Creative Decisions Progressively
- Apply the progressive-confirmation contract in [novel-brief.md](references/novel-brief.md) across novel positioning, story engine, world and cast, volume planning, and the first chapter.
- - Before asking, extract explicit choices from the current request and read existing confirmation records and protected artifacts. Do not ask again about settings marked user-confirmed, delegated to AI, or not applicable.
- - For an explicit no-idea request such as “I have no idea; help me start a novel,” use the five-opening-seed contract in Mission. For a supplied idea or selected seed that still needs exploration, use the two-brief-preview contract before asking for opening settings; reuse an explicitly selected direction or direct-production delegation. Do not silently select a genre, channel, platform, or length and then present an authoritative plan.
- - For a vague request that does not explicitly say there is no idea, use the three-choice first round in Mission. At most, add one short illustrative premise marked non-authoritative.
- - Ask only the 2–3 unresolved choices with the greatest downstream impact. Give exactly 2–4 mutually exclusive content options per choice, put one recommendation first, and explain its effect briefly. Keep custom input, accept-all, AI delegation, and skip as one shared instruction after the choices; do not inflate each option list with these controls.
- - Treat “you decide,” “do not ask me about these,” or “generate directly” as revocable project-level delegation for the affected settings. Persist it in the novel brief when workspace mode is active.
- - When the request already specifies the chosen direction, summarize it and proceed using existing choices or delegation. Offer two meaningfully different brief previews only when the user is exploring alternatives; ask only about material unresolved conflicts. After a direction is selected, ask only the unresolved opening settings required by the current milestone.
- - For a bounded idea, title, name, or other preview request, ask only what that artifact needs; do not launch the full onboarding sequence.
- - Allow unconfirmed AI recommendations in bounded conversation previews only after presenting the current confirmation choices. Do not save them as authoritative, mark their artifact `ready`, or build formal downstream assets from them until the user confirms or delegates the decision.
- - When a confirmed high-impact setting changes, protect existing prose and immediately use the explicit `stale` status for each affected unwritten artifact; do not merely say it is “affected,” defer impact reporting, or silently rewrite it. A confirmed audience-channel or reading-promise change makes the existing unwritten volume strategy, volume skeleton, beat sheet, and chapter plans `stale` unless a dependency check proves a specific artifact unaffected.
- - Treat male-oriented, female-oriented, and broad-audience channels as reader-promise and packaging signals, never as rigid gender stereotypes.
- - Keep this confirmation sequence on the novel-creation route. Reference analysis, fanfic, and hot-genre trends retain their own scope and source-confirmation rules.
- ## Start from Facts
- When a workspace exists:
- 1. Locate and read `novel-state.yaml`, `analysis-state.yaml`, or `trend-state.yaml` first according to the route.
- 2. Read only the artifacts required by the current route.
- 3. Treat user-edited files and existing prose as protected unless the user explicitly authorizes replacement.
- 4. Compare the current artifact with its upstream dependencies before extending it.
- 5. Mark affected downstream plans as stale instead of silently rewriting them.
- When no workspace exists, remain in preview mode unless the user requests or confirms durable production. When promoting a preview, save only the artifacts the user has accepted; do not reconstruct unaccepted conversation drafts as authoritative files.
- Read [artifact-contracts.md](references/artifact-contracts.md) before creating a workspace, changing artifact status, or deciding which downstream files become stale.
- Read [token-usage.md](references/token-usage.md) before recording or reporting model usage. In workspace mode, account for every actual model generation call as `exact`, `estimated`, or `unavailable`; never invent a number when the runtime does not expose usage.
- Read [generation-contracts.md](references/generation-contracts.md) before running or changing the full production chain. Read [auto-director-and-recovery.md](references/auto-director-and-recovery.md) when the user delegates decisions, requests a chapter range, resumes an interrupted run, reaches a milestone approval, or encounters a blocking condition.
- Read [cross-book-asset-graph.md](references/cross-book-asset-graph.md) before publishing, importing, syncing, or resolving reusable assets or shared-IP canon. Only schema-v3 workspaces may link cross-book assets; graph results are candidates and the referenced YAML/Markdown remains authoritative.
- When a novel workspace has `production/skill-sources.yaml`, check its source status before applying any remote writing guidance. `candidate` and `update_available` entries are review material only; only an `enabled` entry with a confirmed fixed commit may contribute rules, and those rules still enter the local book snapshot before drafting.
- 当章节使用跨书资产时,Codex 先从计划和活跃链接提出 `context-packages/chapter-XXX.assets.yaml`,由脚本校验其固定快照、用途和约束;再通过 `novelctl context --asset-library ... --asset-selection ...` 生成有限上下文。选中资产的同步冲突会阻断上下文、正文和审校;共享更新只提示,默认继续使用本书锁定快照。只有正文验收且连续性提交后,才可在 `production/asset-candidates/` 形成待发布资产候选。
- 一个私有跨书资产库默认是一个共享 IP 宇宙。对 `universe` / `event` 候选,Codex 必须先基于已验收来源明确提出 `content.canon.sequence`、参与资产、影响和可选先后关系,再运行 `asset_graph.py canon-check` 与 `impact ... --workspace <明确工作区>`;影响报告只作风险审查,不能改写任何书或自动升级快照。只有作者批准、按资产委托或未撤销的宇宙级 Codex 委托之一成立时,才能用既有 `publish` 提交正史;`reusable` 资产不接受宇宙级委托。时间线与图谱均为 YAML/JSONL 的可重建派生物,不能替代回读权威 YAML/Markdown。
- In a schema-v3 workspace, use `scripts/novelctl.py` as the deterministic control surface for state transitions, validation, recovery, context, checkpoints, usage, and export. Codex remains the only creative generation engine; do not add or invoke a model-provider SDK from this skill.
- ## Run the Production Loop
- Follow this artifact chain, stopping at the milestone the user requested:
- `idea -> novel brief -> story bible -> world and cast -> volume strategy -> volume skeleton -> beat sheet -> chapter plan -> context package -> chapter draft -> humanization revision -> review/repair -> continuity update`
- For every step:
- 1. State necessary assumptions briefly. Apply confirmed or delegated defaults directly; present unconfirmed high-impact defaults as choices instead of silently treating them as authoritative.
- 2. Consume the authoritative upstream artifact rather than reconstructing it from chat history.
- 3. Produce one coherent artifact or one bounded batch.
- 4. Check its acceptance conditions.
- 5. Record the generation call in `production/token-usage.jsonl` before marking the artifact usable. Keep exact, estimated, and unavailable measurements distinct; do not count deterministic file or index operations.
- 6. Update `novel-state.yaml` only after the artifact is usable.
- 7. Report what was produced, what changed, the step's available Token usage, what remains uncertain, and the recommended next action.
- Book-level direction, the story engine, volume strategy, structural replanning, and any protected overwrite are milestone approvals by default. After the user approves a chapter range, continue serially within that range and stop at its end. A project-level AI delegation may satisfy these approvals until the user revokes it.
- Do not attempt to generate an entire long novel in one response. Advance through resumable milestones and keep each deliverable editable.
- ## Analyze Reference Works
- Read [book-analysis.md](references/book-analysis.md) before analyzing a reference novel, an online novel, a long uploaded manuscript, writing techniques, commercial hooks, character systems, plot structure, or a user's draft as a whole.
- Read [book-analysis-retrieval.md](references/book-analysis-retrieval.md) before indexing or querying a long analysis workspace, building BookGraph nodes or edges, tracing story relationships, using full-text retrieval, or attaching optional embeddings.
- - Persist analysis only under `analyses/<normalized-title>/` with `analysis-state.yaml` and the artifact layout defined in `book-analysis.md`. Do not use `novels/`, `novel-state.yaml`, or an invented competing layout for analysis unless the user explicitly asks to attach the results to an existing novel workspace.
- - Accept user-provided text, local text files, existing workspaces, and lawfully accessible online pages. Do not bypass paywalls, login controls, captchas, anti-bot restrictions, or site access rules; request a user-supplied file when access is unavailable.
- - On an authorized, accessible source, record the core reader promise and the visible payoff separately. A genre fantasy, psychological reward, or relationship triangle can be the core even when its delivery is a sequence of bodily, investigative, or power changes. Record rules, identity, and information gaps when they enable or block that promise; keep source facts and short quotes accurate.
- - Freeze the source scope and fingerprint before analysis. Copy a user-supplied local file or chapter directory into private `source/` as a snapshot, and never index `source/` in retrieval. Distinguish full coverage from sampling, and state omitted ranges and blind spots.
- - Before the first segment note, run `scripts/estimate_analysis_tokens.py` on the local source. Report the character count, planned segment count, and source-token range as a length estimate, separate from model usage. If it sets `confirm_required`, ask for full analysis or a bounded partial window; record `remaining_scope` for later incremental analysis.
- - After coverage is confirmed and before the first note batch, draft `analysis-lens.md` from 2–3 evidence-backed core candidates, then separate the selected core reader promise from its surface progression axes and gates. Include a counterfactual check for essential relationships or viewing positions. Self-check it, show the candidates, the selected promise, the progression axes, and primary questions, set it to `pending_confirm`, and wait for acceptance before notes. Existing notes without a lens remain intact; add the lens before the next batch.
- - If `character_sync.status` is `unconfirmed`, ask whether to maintain individual `characters/<CHAR-ID>.md` cards. When enabled, update relevant cards after each note batch, using only observed details and chapter evidence.
- - Build bounded segment notes first. Generate the overview before specialist sections and before handing an analysis into a novel so later judgments share a positioning anchor that includes the core promise, psychological engine, relationship structure, surface progression, and coverage limits.
- - After notes exist for a long source, run `scripts/analysis_retrieval.py build <analysis-workspace>`. Use graph traversal for explicit relationships, filtered lexical search for named facts, and optional vector recall only when compatible embeddings already exist.
- - Treat `retrieval/analysis-index.sqlite3` and embeddings as disposable derived caches. Keep Markdown/YAML, graph JSONL, source fingerprints, and user edits authoritative.
- - Separate source fact, supported inference, and open hypothesis. Bind important conclusions to chapter or segment evidence and preserve contrary evidence.
- - Produce reusable mechanism cards rather than imitation instructions. Describe the core promise, mechanism, prerequisites, reader effect, failure modes, and original transformation. Add stimulus object, viewing position, psychological progression, or return/attachment only when the source evidence makes those dimensions relevant; never reproduce substantial source text or promise stylistic cloning.
- - When using analysis resources in a new or continuing novel, read [analysis-to-novel.md](references/analysis-to-novel.md). Select mechanisms against the current book/chapter need, propose an original adaptation, and validate the pinned evidence before adding craft references to chapter context. Reference conclusions never become this novel's continuity facts.
- - Keep successful analysis sections when another section fails. Resume only missing or stale work, and mark downstream analysis stale when the source fingerprint or scope changes.
- - When the source is the user's own manuscript, use diagnosis mode and recommend bounded repairs without modifying the manuscript unless explicitly authorized.
- ## Write Fanfic from an Analysis
- Read [fanfic.md](references/fanfic.md) before continuing an analyzed novel under its original names or writing a new story with its cast. Use only facts covered by an authorized analysis workspace, confirm continuation versus a new story, and keep `fanfic-handoff.md` in progress until its source scope, branch point, storyline, and pace are ready. A continuation branches from the last fully noted chapter ending. Write fanfic prose only in its own `novels/` workspace and do not treat unread chapters as canon.
- ## Analyze Hot Genre Trends
- Read [hot-genre-trends.md](references/hot-genre-trends.md) before handling popular-ranking themes, hot genres, recent market directions, chart composition, or track opportunities.
- - Keep this route under `trends/<normalized-scope>/`; never mix its source material with `novels/` or `analyses/`.
- - If the platform or audience channel is missing, ask for the target channel first. Default to the top 20 entries per chart and no more than 3 charts per request.
- - Use only public official chart metadata or screenshots, tables, and text supplied by the user. Set `access_level` to `metadata_only`; do not bypass access controls or fetch novel prose.
- - Follow this order: confirm platform/channel/time window, capture the chart, save the snapshot, extract surface signals from title/tags/synopsis, aggregate facts, then create 3–5 opportunity cards.
- - Run `scripts/trend_snapshot.py validate` before reporting. Use `summarize` for one snapshot and `compare` only for two dates from the same platform, chart, and statistical window.
- - Call a single snapshot “current chart composition.” Require at least two comparable dates before claiming rise, decline, persistence, or new entry.
- - Do not infer whole-book pacing, character arcs, foreshadowing payoff, prose quality, or middle-to-late performance from metadata.
- - Move only an explicitly selected opportunity card into `novel-brief.md`. Never inject raw charts, synopses, reports, or unselected cards into novel-production context.
- - If the user requests analysis of a specific work, stop this route, explain the upgrade to reference-book analysis, and request an authorized source without automatically fetching prose.
- ## Build Story and Volume Plans
- Read [story-and-volume-planning.md](references/story-and-volume-planning.md) when creating or changing positioning, the hook package, story engine, world rules, cast, volume strategy, beat sheets, or chapter lists.
- Read [world-bible.md](references/world-bible.md) when creating, revising, or consuming the novel's world rules, faction constraints, or story stages. Treat its rule IDs and protected limits as chapter-level hard constraints, not decorative setting text.
- Read [character-asset-layout.md](references/character-asset-layout.md) when creating, splitting, migrating, reading, or updating character assets in a workspace.
- Preserve these priorities:
- - Lock the reading promise before expanding world detail.
- - Give the protagonist an active desire, repeatable action loop, escalating opposition, and visible rewards.
- - Separate volume strategy from volume skeleton.
- - Plan early volumes more firmly than distant volumes.
- - Protect user-fixed volume counts, milestones, and written chapters.
- - Generate chapter lists by the current beat or volume window when that reduces waiting and rework.
- - Use layered character presentation: a compact roster anchor for reserve roles, a complete identity and visual profile for active core roles, and a chapter-specific current presentation for participants. Do not leave an active character without a usable identity, appearance, dress/prop, voice or habitual action, and first-impression guidance.
- - In a workspace with recurring characters, keep a compact `characters/character-roster.md` index and one profile file per active core character. Read the index plus only the profiles relevant to the current chapter.
- ## Produce Chapters
- Read [chapter-production.md](references/chapter-production.md) before planning, drafting, continuing, reviewing, or repairing a chapter.
- Before writing prose, read [style-engine.md](references/style-engine.md) if `writing-style.md` exists in the novel workspace. Apply its positive guidance and the small anti-template baseline to the first draft; record chapter-specific expression overrides in the plan and context package. Keep the effective contract available for the final review and any repair.
- When the user asks to continue multiple chapters, finish a long chapter range, improve generation speed, work in parallel, or use subagents, keep the chapter-production route serial. The responsibility-specific subagent experiment is paused: do not create child agents or speculative next-chapter candidates. Produce, review, and commit one chapter before planning the next.
- Read [continuity-ledgers.md](references/continuity-ledgers.md) before bootstrapping continuity for an existing workspace, creating a context package, accepting a chapter, updating a ledger, recording quality debt, or recovering after an interruption.
- Read [structured-continuity-store.md](references/structured-continuity-store.md) when `continuity/data/` exists or when migrating continuity to YAML, rebuilding its SQLite index, creating a checkpoint, or assembling bounded long-form context.
- Read [chinese-novel-humanization.md](references/chinese-novel-humanization.md) before generating a complete chapter draft, producing a comparison rewrite, or reviewing prose for template-like machine patterns. Apply its protected-facts and no-artificial-noise rules.
- Use one shared chapter contract across drafting, acceptance, and repair. Include:
- - immediate chapter goal and resistance;
- - required events, facts, appearances, and payoff touches;
- - protected facts and forbidden crossings;
- - reader question, promised reward, key turn, net change, and ending pull;
- - previous-chapter handoff and current continuity constraints.
- - the relevant world-rule, faction, and stage IDs, plus any current state that limits this chapter.
- - a chapter-length contract: target length, acceptable range, its source in the book brief or an approved chapter override, and scene-level budget allocation.
- Generate the whole chapter as one coherent draft by default. Use scene beats for planning and targeted repair, not as a reason to stitch together many disconnected mini-drafts.
- For a multi-chapter request, repeat the complete single-agent chapter loop: plan and context package, one coherent draft, humanization, review, then continuity commit. Do not prepare later chapter candidates until the current chapter has become the stable source of truth.
- Before drafting a new chapter in a continuity-enabled workspace, create its concise context package. It must identify selected authority sources, required constraints, deliberately omitted material, and any missing hard constraint. Do not copy long prose into the package.
- After a complete draft, run one constrained Chinese-novel humanization pass by default. Preserve the chapter contract, effective writing style, and all protected facts; change only passages where a clustered expression problem harms reading, and leave sound passages intact. Keep the original draft when the user requests an experiment or comparison. Never promise a detector score or use a detector score as the sole acceptance criterion.
- After review, commit continuity only when the final prose is `accepted`, or is `continue_with_warning` and still safe to continue. In a migrated workspace, update and validate YAML first, then regenerate Markdown views and SQLite; in a legacy workspace, update the Markdown ledgers directly. Then update affected character assets, quality debt, recovery record, and state index. Do not commit planned events, local-patch candidates, rewrite-needed drafts, or replan-blocked drafts as facts.
- ## Manage Context and Continuity
- Read [context-and-continuity.md](references/context-and-continuity.md) when the request involves continuation, long-running consistency, character state, world rules, prior chapters, references, style, facts, resources, or payoffs.
- For a continuity-enabled workspace, read the recovery record and only the YAML-selected facts (or legacy ledgers), profiles, world entries, prior tail, and volume assets relevant to the current chapter. If a source fingerprint no longer matches its stable source, mark the dependent delta, ledger entry, and context package stale; never overwrite user-edited prose to make them match.
- Prefer the smallest context package that preserves coherence. Never load every chapter merely because it exists. Prioritize hard constraints, the current task, the previous handoff, relevant character facts, the active volume window, and unresolved promises.
- Keep durable facts in their authoritative modules (`characters/`, `world-bible.md`, and `continuity/`), while each `chapters/chapter-XXX/` directory holds that chapter's plan, context package, prose, and review. The chapter context package is a minimal assembly of references and chapter-specific obligations, never a duplicated or competing fact store.
- ## Export Ready Chapters
- When the user asks to merge, compile, or export generated chapters as a TXT file, run `scripts/novelctl.py export <workspace>` against the confirmed novel workspace. It exports only committed chapters whose accepted review still matches the prose and contract.
- - Export only chapter prose. Never include plans or reviews.
- - Default to `--source auto`: use the actual accepted prose path, including an accepted repair. The standalone exporter also defaults to stable chapters; use its `--include-drafts` only when the user explicitly requests unaccepted drafts.
- - Preserve numeric chapter order. Use `--start` and `--end` only when the user specifies a chapter range.
- - Run with `--dry-run` before a non-default range or source selection; otherwise export directly to the workspace `exports/` folder.
- - Report the exported chapter range, source selected for each chapter, skipped chapters, and final output path.
- ## Audit and Repair
- Read [quality-and-repair.md](references/quality-and-repair.md) before evaluating quality, rewriting prose, repairing continuity, or recommending replanning.
- Apply this escalation order:
- 1. Accept usable content.
- 2. Record non-blocking quality debt.
- 3. Apply a targeted patch when the problem is local.
- 4. Rewrite the chapter only when local repair cannot satisfy the chapter contract.
- 5. Recommend neighboring-plan or volume replanning only when the chapter responsibility itself is structurally impossible or misplaced.
- Do not discard a usable chapter because of a local style flaw. Do not create endless review-repair loops.
- ## Project Adapter Boundary
- This skill is a Codex-native writing workflow, not the runtime of `AI-Novel-Writing-Assistant-v2`. Read [ai-novel-writing-assistant-v2-adapter.md](references/ai-novel-writing-assistant-v2-adapter.md) only when the user asks to compare, align, import, export, distill production ideas, or develop against that project.
- Do not directly operate its database, task queue, HTTP routes, or production runtime unless the user separately authorizes an implementation task.
- ## Skill Maintenance
- For changes to analysis, adaptation, author corrections, or humanization, use the small [behavioral regression cases](references/behavioral-regressions.md) alongside deterministic tests. Check actual decisions and output evidence; field presence and successful scripts do not establish story quality. Honor the user's delegation preference when validating.
- When maintaining this skill, treat a user's explicit acceptance of a proposed workflow, artifact, directory, or constraint change as authorization to update the relevant skill instructions in the same task. Update `SKILL.md` and the directly affected reference contracts together, then run a focused consistency check for conflicting paths, duplicate authority, and stale terminology. Do not treat an unconfirmed recommendation as a specification change, and preserve unrelated user edits.
- ## Finish Clearly
- End each completed production action with:
- - artifact created or updated;
- - important assumptions or protected decisions;
- - quality or continuity risks;
- - state transition performed;
- - Token usage for the completed model-generation step, separated into exact, estimated, or unavailable;
- - one recommended next action.
README.md · README · 296 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- <div align="center">
- # Ani Book Skill
- **从灵感到定稿,让长篇创作持续推进。**
- 开书 · 拆书 · 原创机制迁移 · 同人创作 · 写法与去 AI 化 · 长篇连载
- [](https://github.com/ExplosiveCoderflome/ani-book-skill/actions/workflows/validate.yml)
- [](requirements.txt)
- [](LICENSE)
- [快速开始](#快速开始) · [完本路线](#从第一个想法到小说完本) · [拆书路线](#拆书如何进入创作) · [写法与去-ai-化](#写法与去-ai-化) · [本地阅读台](#本地阅读台) · [更新记录](docs/releases/release-notes.md) · [English](README.en.md)
- </div>
- 
- Ani Book Skill 是面向中文长篇小说的 **Codex 原生 Skill**。它把选题、拆书、规划、正文、修订和连续性维护组织成可恢复的工作流,产物保存在你能直接阅读、编辑和备份的本地 Markdown / YAML 文件中。
- **Codex 负责理解与创作,Skill 规定流程,本地工具核对状态、版本和依赖。** 你可以逐项决定关键设置,也可以把指定范围交给 Codex 推进。
- ## 功能一览
- | 你的需求 | 可以得到什么 |
- | --- | --- |
- | **从零开书** | 五条开书种子、双方向简报、目标读者、故事发动机与首卷规划;已有明确方向时可直接推进。 |
- | **寻找题材机会** | 基于公开榜单元数据的快照、题材构成与机会卡;有可比日期时再判断变化趋势。 |
- | **拆解参考作品** | 阅读范围与覆盖地图、分段笔记、核心阅读承诺、人物与关系、结构和写法机制,以及可核对的证据。 |
- | **把拆书用于原创** | 按目标小说需要选择机制,设计原创改造,接入当前章计划与上下文,并在审校时检查实际效果。 |
- | **创作同人** | 在已授权、已读范围内,从完整章末续写,或沿用人物与设定另起新故事。 |
- | **稳定写法、减少 AI 味** | 本书写法合同、人物声音与表达约束;从初稿起应用,二稿和审校按具体片段处理。 |
- | **持续连载与修稿** | 卷战略、节奏板、章节计划、完整正文、修订、审校、连续性提交、中断恢复与稳定章节导出。 |
- | **经营系列与共享 IP** | 跨书资产复用、独立分叉、更新提示、正史时间线与影响检查。 |
- ## 从第一个想法到小说完本
- **起步只需要一句想法。** 先确定读者想看什么,再逐步形成全书方向、当前卷和当前章;按已批准的范围逐章推进,最后检查全书收束并导出。
- ```mermaid
- flowchart TD
- idea["输入第一个想法"] --> choice["确定方向与关键设置"]
- choice --> brief["保存新书简报与小说工作区"]
- brief --> foundation["故事发动机、世界、角色与写法"]
- foundation --> volume["卷战略 → 当前卷骨架与节奏板"]
- volume --> plan["本章计划与上下文"]
- plan --> draft["完整初稿 → 受约束二稿"]
- draft --> review{"审校结果"}
- review -->|需要修文| repair["定点修复或重写"]
- repair --> review
- review -->|需要重规划| volume
- review -->|允许推进| continuity["提交连续性与恢复记录"]
- continuity --> scope{"当前进度"}
- scope -->|授权范围内还有下一章| plan
- scope -->|进入下一卷| volume
- scope -->|已到授权范围末尾| checkpoint["保存检查点,按后续指令继续"]
- checkpoint -->|更新授权范围| scope
- scope -->|已写到全书结尾| closure["全书收束检查与作者确认"]
- closure --> export["校验稳定章节 → 按顺序导出 TXT"]
- ```
- | 阶段 | 你要做什么 | Skill 推进什么、留下什么 |
- | --- | --- | --- |
- | **1. 给出起点** | 描述一个人物、矛盾、场景或想看的体验。 | 没有想法时给五条种子;需要探索时比较两份简报;已有选定方向或明确委托时直接复用。 |
- | **2. 确定开书方向** | 确认或委托决定当前必需的读者、题材、核心回报、发布形态和内容边界。 | 形成 `novel-brief.md`;进入正式生产后保存到 `novels/<书名>/`,由 `novel-state.yaml` 记录进度与下一步。 |
- | **3. 建立全书支撑** | 确认主线、人物选择与结局方向,提供想保留的写法偏好。 | 建立故事发动机、`story-bible.md`、世界与角色资料;需要时形成 `writing-style.md`。 |
- | **4. 规划当前卷** | 确认阶段目标、主要冲突和当前生产范围。 | 全书卷战略逐层落到当前卷骨架、节奏板和章节职责;近期具体,远期保留调整空间。 |
- | **5. 逐章生产** | 在约定节点阅读、纠正或验收,也可授权连续推进一个范围。 | 本章计划 → 上下文 → 整章正文 → 二稿 → 审校;修复后重新审校,结构问题回到规划。 |
- | **6. 持续推进** | 给出反馈;到达授权范围末尾后决定接下来的范围。 | 允许推进的章节先提交事实、伏笔、角色和关系变化,再进入下一章;卷末回看阶段兑现,滚动规划下一卷。 |
- | **7. 收束全书** | 确认结局效果和有意保留的开放问题。 | Codex 对照简报、故事圣经、卷规划与连续性检查主线兑现、人物弧线、伏笔回收及未解决问题,再修订需补足的部分。 |
- | **8. 验收与导出** | 确认最终正文和交付范围。 | 核对预定章节是否齐全、审校版本是否匹配、连续性是否提交,处理阻断问题后导出稳定正文 TXT。 |
- **完本检查关注故事是否完成承诺。** 主冲突与结局需要落地,人物和关系的变化需要有结果;伏笔与未完事项应已兑现,或被作者明确保留。全书收束由 Codex 结合现有工件检查、作者确认,导出工具负责核对稳定章节并合并正文。
- 可以这样开始:
- > 使用 `$produce-long-form-novel`。我的第一个想法是:……。请引导我从开书走到完本,保存小说工作区,逐步确认关键设置,按我批准的章节范围推进;写到结尾后检查全书收束并导出 TXT。
- 中断后可以说:
- > 继续 `novels/my-book/`,先读取状态和恢复记录,告诉我当前进度、待解决问题及下一步,再在已有授权内继续。
- [开书与设置](references/novel-brief.md) · [全书与卷规划](references/story-and-volume-planning.md) · [工件与工作区](references/artifact-contracts.md) · [章节生产](references/chapter-production.md)
- ## 拆书如何进入创作
- ### 从提供文本到拆书交付
- **先确定读什么、为什么拆,再逐段建立证据。** 可以只拆选定窗口,也可以按批次完成全文;每份结论都保留实际阅读范围,后续还能继续补拆或接入创作。
- ```mermaid
- flowchart TD
- source["提供有权处理的文本"] --> estimate["估算规模,确定目标、深度与阅读范围"]
- estimate --> snapshot["保存来源快照、指纹与分析状态"]
- snapshot --> lens["比较核心阅读承诺,确认拆书透镜"]
- lens --> notes["分批阅读 → 分段笔记与覆盖地图"]
- notes --> evidence["核对证据、反例与不确定性,建立检索"]
- evidence --> sections["总览 → 按目标生成专项分析"]
- sections --> cards["提炼机制卡与适用边界"]
- cards --> progress{"约定范围是否完成"}
- progress -->|尚有未读部分| notes
- progress -->|当前范围完成| delivery["交付结论、覆盖范围与缺口"]
- delivery -->|继续补拆| notes
- delivery -->|用于原创| original["原创改造 → 目标小说与章节合同"]
- delivery -->|用于同人| fanfic["核对已读边界 → 同人交接与独立工作区"]
- ```
- | 阶段 | 主要工作 | 留下的资料 |
- | --- | --- | --- |
- | **1. 来源与范围** | 提供授权文本,估算长度与 Token 区间,确定全文或部分范围,以及结构、人物、写法等拆解目标。 | `source-manifest.md`、`source/` 原文快照与来源指纹。 |
- | **2. 确认拆书透镜** | 比较有证据支持的核心候选,区分读者回报与表层事件;确认本次重点及是否同步角色卡。 | `analysis-state.yaml`、确认后的 `analysis-lens.md`。 |
- | **3. 分批阅读** | 按窗口记录发生了什么、人物如何选择、关系和承诺如何变化,避免用总评替代阅读。 | `notes/segment-NNN.md`、`coverage-map.md`,按需补充 `characters/`。 |
- | **4. 证据与分析** | 区分事实、推断和假设;保留来源锚点、反例与缺口,从笔记建立检索,形成总览及所需专项。 | `evidence-ledger.md`、`uncertainty-ledger.md`、`sections/overview.md` 与其他分析小节。 |
- | **5. 提炼可复用机制** | 解释为什么有效、需要什么条件、换一个故事怎样改造,以及哪些结论仍待验证。 | `pattern-cards.md`;接入原创章节时整理可校验的 `pattern-cards.yaml` 和证据引用。 |
- | **6. 交付、补拆或创作** | 查看覆盖与剩余范围;继续下一批,或选择原创机制迁移、同人新故事、同人续写。 | 可恢复的拆书工作区,以及按目标创建的改造方案或同人交接包。 |
- **部分范围完成会明确标注为部分拆书。** 只有实际读完约定的全文,才能声明全文覆盖。补拆先回读状态、透镜与已有笔记,再处理 `remaining_scope`;覆盖扩大或核心结论变化后,重新核对相关总览、机制卡和创作引用。
- 可以这样开始:
- > 使用 `$produce-long-form-novel`,拆解这份我有权使用的文本:`<文件路径>`。先估算规模,与我确定阅读范围和拆解重点,再比较并确认核心阅读承诺。按批次保留笔记与证据,最后交付总览、专项分析、可复用机制和覆盖缺口。
- 接着可以说:
- > 继续 `analyses/my-reference/` 的剩余范围;或根据已确认的拆书机制,设计原创开书方向,并说明保留哪些读者回报、怎样重建人物与事件。
- ### 提取支撑追读的核心机制
- 拆书先确定授权来源、阅读范围和分析目标,再比较可能的核心阅读承诺。分析会分别记录**读者持续想获得的体验**与**承载这种体验的表层事件**,结合证据、反例和反事实检查判断哪些要素真正不可替代。
- 你可以从中复用:
- - **故事发动机**:人物为什么行动,回报与代价怎样形成循环。
- - **人物与关系机制**:欲望、选择、信息差、关系变化及其后果。
- - **结构与节奏**:冲突推进、信息释放、阶段兑现和章末牵引。
- - **写法方法**:视角、场景组织、对白差异和细节选择。
- - **证据与适用边界**:结论来自哪里、覆盖了多少、需要什么前提、何时会失效。
- 按题材选择相关维度,避免把所有作品套进同一种力量升级或关系模型。分段笔记和机制卡可以检索,结论保留来源与覆盖范围;事实、推断与待验证假设分别记录。
- ### 同一份研究,三种创作方式
- | 方式 | 保留什么 | 怎样推进 |
- | --- | --- | --- |
- | **原创机制迁移** | 抽象原理、读者回报与运作条件 | 重建人物、世界和事件,把改造后的机制写入目标书的章节合同。 |
- | **同人新故事** | 已读范围内的原作人物与设定 | 明确起跳点和新故事边界,在独立小说工作区规划。 |
- | **同人续写** | 完整记录的章末状态与未完承诺 | 先完成交接包,再承接章末尚未发生的下一步。 |
- 交接与引用会锁定所采用的分析版本。**原文未变、拆书核心结论变了,旧交接也会失效。** 创作前重新核对来源,防止继续沿用旧判断。
- [拆书流程](references/book-analysis.md) · [检索与证据](references/book-analysis-retrieval.md) · [拆书接入原创](references/analysis-to-novel.md) · [同人流程](references/fanfic.md)
- ## 写法与去 AI 化
- 本项目把“写法引擎”落实为一份可编辑的 **本书写法合同**:说明希望呈现怎样的叙述距离、人物声音、对白、节奏与细节,以及哪些表达问题只提醒、哪些允许定点修订。
- 写法可以来自作者自己的样文、已验收章节、明确偏好,或授权拆书中的可迁移方法。单书快照保存在 `writing-style.md`,跨书复用卡保存在私有 `styles/`。首次开书无需先配置,写出章节后也可以逐步完善。
- `有效写法 → 初稿应用 → 一次受约束二稿 → 同合同审校 → 必要的定点修文`
- 去 AI 化关注成簇的模板表达、空泛解释、人物声音趋同和场景推进不足。修订需要指出具体证据并保留有效表达;“仅提醒”项由作者决定。作者纠正会按用词、表达或设定范围保存,供后续章节回读。
- ### 从 GitHub 获取补充方法
- 需要外部方法时,可以明确要求 Codex 获取写法或去 AI 化 Skill:
- `登记来源 → 获取候选 → 固定版本与核验内容 → 审阅规则 → 作者确认 → 本书写法快照`
- 候选保存在私有目录,可检查文件边界、内容摘要和大小。更新保留当前版本,经审阅再采用;章节生产使用本地快照。规则是否有用,仍依据正文效果判断,检测器分数不作为唯一验收标准。
- [写法合同](references/style-engine.md) · [中文小说反模板化修订](references/chinese-novel-humanization.md) · [GitHub Skill 来源](references/remote-skill-sources.md)
- ## 长篇如何持续推进
- 沿着[从第一个想法到小说完本](#从第一个想法到小说完本)的路线,以下检查贯穿每一轮章节生产:
- | 关键环节 | 当前支持 |
- | --- | --- |
- | **开书与决策** | 渐进确认高影响设置;复用已确认方向和授权范围,减少重复选择。 |
- | **章节上下文** | 结合本章计划、上一章实际尾段、简报、写法和相关连续性;缺少必需来源时明确报错。 |
- | **审校与定稿** | 验收绑定实际正文及章节合同版本;定稿自动保护,授权覆盖前保存旧内容。 |
- | **变更与修复** | 上游重做后旧审校失效;修复稿须重新验收,结构性重规划会阻断继续生产。 |
- | **连续性与恢复** | 维护事实、伏笔、资源、角色动态与关系约束;状态、恢复记录和检查点帮助定位下一步。 |
- | **多卷与导出** | 章节可明确所属卷;默认按数字章号导出已提交、验收版本仍匹配的稳定正文。 |
- 当前章完成连续性提交后再推进下一章。旧工作区支持预检、备份和显式迁移;YAML 与已验收 Markdown 保持各自的权威边界,SQLite 和阅读视图可重建。
- 系列创作还可从定稿中提炼私有资产:`fork` 为当前书保留独立版本,`sync` 追踪共享版本的更新或冲突。共享 IP 的正史发布由作者批准或有效委托控制,时间线与影响检查用于辅助决策。
- [章节生产](references/chapter-production.md) · [审校与修复](references/quality-and-repair.md) · [恢复流程](references/auto-director-and-recovery.md) · [跨书资产与正史](references/cross-book-asset-graph.md)
- ## 快速开始
- ### 1. 安装到 Codex
- 需要 **Codex、Git、Python 3.10+**。下面以 PowerShell 为例,在准备存放项目的位置运行:
- ```powershell
- git clone https://github.com/ExplosiveCoderflome/ani-book-skill.git
- cd ani-book-skill
- python -m pip install -r requirements.txt
- $skillBase = if ($env:CODEX_HOME) { $env:CODEX_HOME } else { Join-Path $env:USERPROFILE '.codex' }
- $skillTarget = Join-Path $skillBase 'skills/produce-long-form-novel'
- python scripts/sync_skill_mirror.py sync . $skillTarget
- python scripts/sync_skill_mirror.py check . $skillTarget
- ```
- `in_sync: true` 表示源码与安装副本一致。安装后在新的 Codex 对话中调用 `$produce-long-form-novel`。以后更新源码并通过项目验证后,可用同一组同步命令更新安装副本。
- ### 2. 直接描述你要做的事
- | 起点 | 给 Codex 的示例 |
- | --- | --- |
- | 没有想法 | “使用 `$produce-long-form-novel`,我还没有想法,先给我五条开书种子。” |
- | 已选方向 | “使用 `$produce-long-form-novel`,按下面这个已选方向创建小说,其余未决设置由你推荐:……” |
- | 拆书 | “拆解这份我有权使用的文本,先估算规模,确认范围,再比较它的核心阅读承诺。” |
- | 原创机制迁移 | “根据这份拆书设计原创方向,重建人物与世界,说明保留哪些机制以及怎样改造。” |
- | 同人 | “基于这份授权拆书,从最后一个完整记录的章末继续,先检查交接条件。” |
- | 写法与去 AI 化 | “根据我的样文提炼本书写法,检查人物声音和模板表达,先列出证据与处理建议。” |
- | 获取外部方法 | “从 GitHub 获取适合中文小说的去 AI 化 Skill,核验并审阅候选,给我可采用的规则。” |
- | 继续连载 | “继续 `novels/my-book/`,先检查状态;在已批准范围内逐章完成正文、审校和连续性。” |
- | 导出 | “检查 `novels/my-book/`,把验收版本匹配的稳定章节按顺序导出 TXT。” |
- 明确“没有想法”时先给五条种子;需要探索方向时比较两份简报。**已有具体选定方向、完整简报或直接生产委托时,复用已有决定继续。** 正式产物在确认保存或进入工作区后落盘,早期探索可以只在对话中进行。
- <details>
- <summary><strong>已有工作区:常用检查与导出命令</strong></summary>
- 在仓库根目录运行,将路径替换为你的实际工作区:
- ```powershell
- $book = 'novels/my-book'
- python scripts/novelctl.py status $book --format markdown
- python scripts/novelctl.py next $book
- python scripts/novelctl.py validate $book
- python scripts/novelctl.py export $book --dry-run
- python scripts/novelctl.py export $book
- ```
- 迁移旧工作区前,先运行 `python scripts/novelctl.py migrate $book --dry-run`。工具负责检查、维护与导出,创作步骤由 Codex 执行。完整说明见 [SKILL.md](SKILL.md) 与 [工件合同](references/artifact-contracts.md)。
- </details>
- ## 本地阅读台
- 可选的 **ANI 本地阅读台** 在浏览器里展示小说和拆书资料,便于浏览长篇工作区:
- - 书架、资料分类与完整文件树。
- - 中文全文搜索、页内目录、字号和明暗主题。
- - 正文上下章导航、来源文件与指纹查看。
- - 拆书覆盖、分析缺口、机制卡与证据状态。
- - 为目标小说与章节准备可复制的机制接入请求。
- 需要 **Node.js 22.18+ 和 npm**。在仓库根目录运行:
- ```powershell
- ./apps/workspace-reader/start-local.ps1
- ```
- 打开 [本地阅读台](http://127.0.0.1:8790)。其他位置的资料可通过 `-Roots 'C:/writing', 'D:/research'` 指定。阅读台只读明确配置的本地目录,创作、编辑与状态维护仍在 Codex 和原工作区进行。
- [启动方式、功能与限制](apps/workspace-reader/README.md)
- ## 文件、隐私与项目边界
- ```text
- SKILL.md Codex 创作流程入口
- references/ 开书、拆书、写法、章节、恢复与跨书规则
- scripts/ 状态、校验、检索、候选核验与导出工具
- apps/workspace-reader/ 可选的本地只读阅读台
- docs/ 更新历史与有限质量自检记录
- novels/ 私有小说工作区
- analyses/ 私有拆书、来源与证据
- styles/ 私有可复用写法卡
- trends/ 私有榜单快照与研究
- libraries/ 私有跨书资产与共享 IP 资料
- ```
- 真实稿件、来源文本、研究和私有资产目录默认被 Git 忽略;环境配置、密钥和备份文件也有对应忽略规则。作者可以直接编辑文件,工作流会保护用户编辑并检查下游是否失效。Token 用量区分精确值、估算值与运行时不可获取,不以字符数伪造用量。
- Codex 本身是唯一的创作理解、规划、生成、审校与判断引擎,Python 处理确定性工具任务;本地阅读台是独立可选组件。本项目吸收了 [AI-Novel-Writing-Assistant](https://github.com/ExplosiveCoderflome/AI-Novel-Writing-Assistant) 的长篇生产经验,采用可编辑文件与 Skill 合同组织工作,无需部署原项目的前后端服务。
- 榜单研究只用公开元数据;拆书使用有权处理的文本,保留来源、覆盖范围与证据。原创建议迁移抽象机制,同人创作遵守已读范围。去 AI 化修订保留事实、人物目的和作者选择,效果仍需结合成稿与反馈判断。
- [完整 Skill 规则](SKILL.md) · [贡献指南](CONTRIBUTING.md) · [安全政策](SECURITY.md) · [Apache License 2.0](LICENSE)
- ## 最新更新
- ### 2026-10-15
- - **新增**:GitHub 写法与去 AI 化 Skill 候选的内容、大小及版本核验。
- - **优化**:核心承诺贯穿拆书与创作,作者纠正持续生效;已选方向可以直接推进,题材专属机制按需采用;重设计首页横幅,补齐中英文完本路线、拆书路线与安装指引。
- - **修复**:定稿保护与审校版本绑定、必需上游和多卷检查、连续性恢复及稳定正文导出。
- 完整历史见[更新记录](docs/releases/release-notes.md)。
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.