Produce Long Form Novel

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:

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, 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, generate exactly two non-authoritative new-book brief previews before formal planning. 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.

Read workflow-routing.md when the request spans routes, the next step is unclear, or an existing workspace may be incomplete.

Read 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 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.

Guide Creative Decisions Progressively

Apply the progressive-confirmation contract in 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 any supplied idea or selected seed, use the two-brief-preview contract before asking for opening settings. 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 is already specific, summarize the understood idea, produce two meaningfully different brief previews, and ask only for direction selection or material 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 before creating a workspace, changing artifact status, or deciding which downstream files become stale.

Read 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 before running or changing the full production chain. Read 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 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.

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 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 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 visible reader payoff first. Record rules, identity, and information gaps when they enable or block that payoff; 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 this book's visible reader payoff and the gates that affect it. Self-check it, show its one-sentence payoff 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 so later judgments share one positioning anchor.
• 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 prerequisites, reader effect, failure modes, and a safe transformation direction; never reproduce substantial source text or promise stylistic cloning.
• When using analysis resources in a new or continuing novel, read 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.