novel-writing — Claude Code skill: plan, draft and revise fiction
A skill for fiction work: it plans chapters and arcs, carries a draft forward, and reviews finished prose. Handy when chapter versions and author voice matter.
npx skills add wgwtest/novel-writing --skill novel-writing -a claude-codeWhat novel-writing does
The skill splits fiction work into three stages. Planning covers what a scene, chapter, arc or whole story should accomplish before any prose exists. Drafting handles new text, a rewrite, or a continuation of a chapter already underway. Reviewing is for critique and revision advice on material that is already written.
Behind it sits a set of reference files the skill pulls in per task: references/planning.md, references/story-outline-and-causal-summary.md, references/dialogue-and-behavior.md, references/scene-and-structure.md, references/style-fidelity.md, references/realism-constraints.md among others. Long projects get a layered context strategy — the immediate job, hard constraints, full text of nearby chapters, structural cards for the distance, and a cold zone left out. Full text is read on demand rather than loading the whole novel.
Several rules are non-negotiable. The reader does not share the author's knowledge: if the prose never showed it, it counts as missing. A major character cannot walk on stage without something to anchor them — role, tie to the current event, a first impression. A character only knows what the scene permits: what they see, hear, reach, and can correctly infer. Poetry, screenplay formatting, pure copyediting without narrative judgement, and non-fiction fall outside the skill.
The skill does not invent its own version system.
What it helps with
- plan a chapter or arc before writing prose
- continue a chapter without flattening the author's voice
- check whether a major character is introduced clearly
- diagnose a scene where dialogue became turn-taking
- test a scene against real-world limits like visibility or procedure
Example prompts for the agent
- Plan chapter two: what it needs to do and how it ends
- Continue the scene in
chapter-04.md, keep the author's voice - Review the dialogue here — why does it read like a transcript
How to install novel-writing
1Skills CLI (Vercel)
npx skills add wgwtest/novel-writing --skill novel-writing -a claude-code2Manually: put the files in place
~/.claude/skills/novel-writing/SKILL.mdCopy the whole skill folder; for one project use .claude/skills/novel-writing/
1Skills CLI (Vercel)
npx skills add wgwtest/novel-writing --skill novel-writing -a codex2Manually: put the files in place
.agents/skills/novel-writing/SKILL.mdFor all projects use ~/.agents/skills/
1Skills CLI (Vercel)
npx skills add wgwtest/novel-writing --skill novel-writing -a cursor2Manually: put the files in place
.cursor/skills/novel-writing/SKILL.mdCursor also reads .agents/skills/ and .claude/skills/
1GitHub CLI
gh skill install wgwtest/novel-writing novel-writing2Manually: put the files in place
.github/skills/novel-writing/SKILL.md1Skills CLI (Vercel)
npx skills add wgwtest/novel-writing --skill novel-writing -a gemini-cli2Platform CLI
gemini skills install https://github.com/wgwtest/novel-writing --path novel-writingManually: put the files in place
.devin/skills/novel-writing/SKILL.mdThe legacy .windsurf/skills/ path still works
Skills CLI (Vercel)
npx skills add wgwtest/novel-writing --skill novel-writing -a opencodeClaude.ai
Zip the folder and upload it in Settings → Capabilities → Skills
Step by step, with commands for each platform: How to install skills in Claude Code: five ways
Questions about novel-writing
How do I install novel-writing in Claude Code?
Run npx skills add wgwtest/novel-writing --skill novel-writing -a claude-code. The skill lands at ~/.claude/skills/novel-writing/SKILL.md.
What does the novel-writing skill do?
It handles fiction across three stages: planning scenes and arcs, drafting or continuing prose, and reviewing finished chapters. Reference files are loaded to match the task at hand.
What is it not meant for?
Poetry-only requests, screenplay formatting, non-fiction, and copyediting where no narrative judgement is needed.
Source text
SKILL.md · MIT
--- name: novel-writing description: Use when setting up a fiction project, planning, drafting, revising, or reviewing fiction where chapter versions, author feedback, character knowledge, scene causality, dialogue, or style fidelity matter. --- # Novel Writing ## Overview Use this skill for fiction work that needs narrative craft, not just sentence-level cleanup. It covers planning, drafting or continuing prose, and reviewing finished fiction. When a task crosses chapter versions, it also coordinates the project's manuscript handoff and author-feedback learning. ## When to Use Use this skill when the user asks to: - plan a chapter, scene, arc, volume, whole-story outline, story synopsis, canon document, or reveal sequence - set up a fiction project that will keep formal chapters, working revisions, and author feedback across sessions - continue or rewrite fiction prose - check whether a chapter is structurally sound - review whether characters are introduced clearly - preserve an author's original voice while revising - compare an author's hand-edited chapter with an actual delivery and carry supported feedback into later work - deliver a revision within an existing project's manuscript and archive rules - test whether a scene obeys real-world limits such as visibility, procedure, age, or capability Do not use this skill for: - poetry-only requests - screenplay formatting work - pure copyediting where narrative judgment is not needed - non-fiction writing ## Stage Selection For project setup or work across manuscript versions, load the lifecycle reference as directed below. For prose tasks, decide which stage the task belongs to. ### 1. Planning Use when the user is deciding what a scene, chapter, arc, volume, or whole story should do before prose is written. Read: - `references/planning.md` for scene- or chapter-level planning - `references/story-outline-and-causal-summary.md` for an arc, volume, or whole-story outline, a story synopsis, or a standalone canon/theory document - `references/cognition-layers-and-language.md` when knowledge, motive, or disclosure drives decisions - `references/scene-causality-and-agency.md` for conflict, responsibility, action order, or chapter interfaces - `references/dialogue-and-behavior.md` when a planned meeting, interview, consultation, investigation, or technical scene could collapse into turn-taking exposition - `references/scene-and-structure.md` when chapter flow is the main concern - `references/realism-constraints.md` when the chapter depends on real-world limits ### 2. Drafting or Continuing Use when the user wants new fiction prose, a rewrite of prose, or a continuation of an existing chapter. Read: - `references/character-introductions.md` - `references/cognition-layers-and-language.md` when decisions or dialogue depend on unequal knowledge, judgment, or motive - `references/scene-causality-and-agency.md` for conflict, movement, authority, or continuity - `references/dialogue-and-behavior.md` for sustained dialogue, meetings, interviews, interrogations, domestic confrontations, or laboratory and research scenes - `references/scene-and-structure.md` - `references/style-fidelity.md` - `references/realism-constraints.md` when the scene depends on real-world constraints ### 3. Reviewing Use when the user wants critique, diagnosis, or revision advice on existing fiction. Read: - `references/revision-checklist.md` - `references/dialogue-and-behavior.md` when dialogue reads like a transcript, characters take orderly turns, or the scene contains motion but little human response - then load the specific reference files that match the problems you find ## Long-Fiction Context Strategy When the fiction project is long enough that loading the entire formal draft would be wasteful, use context LOD instead of full-text-first reading. ### Load order - `L0 Task`: the immediate writing or review job - `L1 Hard constraints`: outline, timeline, setting, issue records, local project rules - `L2 Near-field full text`: current chapter, adjacent chapters, and linked manuscript layers when fidelity matters - `L3 Far-field structure`: chapter cards, timeline slices, and outline slices - `L4 Cold zone`: unrelated future chapters and distant full text, excluded by default - If you are directly revising a chapter, read that chapter's full text. - If dialogue rhythm, psychological detail, ambiguity, flirtation, or body-detail writing matters, read the relevant full text instead of relying on summary material. - If structure cards conflict with the project's authoritative prose, flag the discrepancy and use the prose as evidence for this task. Update cards only when the project's rules and current authorization allow it; an unreviewed author edit is not automatically accepted canon. - If the available structured context is insufficient, expand by targeted full-text reads, not by loading the whole novel. - Future-chapter prose is excluded by default unless a confirmed continuity dependency requires it. ### Manuscript and author-feedback stewardship When creating a fiction project or working with chapter versions, a hand-edited or annotated author draft, a delivery, or cross-session recovery, read [Manuscript Lifecycle and Author Feedback](references/manuscript-lifecycle-and-feedback.md). This skill coordinates the setup or handoff among the current manuscript, working revisions, frozen evidence, and reusable author feedback; the project's own rules determine their paths and authority. Do not create a second version system or assume that a writing or review tool owns the project's formal manuscript. For parallel chapter ownership, follow the current project's explicit workflow. ## Hard Rules ### Reader Knowledge Is Not Author Knowledge Never assume the reader already knows: - who a character is - how people are related - what a place looks like - why a behavior matters If the author knows it but the prose has not shown it, treat it as missing. ### Important Characters Cannot Enter Naked When an important character first enters the story in an effective way, the prose must give the reader enough to anchor them: - identity or role - relationship to the current event or viewpoint - first impression If the character is a villain, love interest, rival, or core ally, the prose also needs at least one recognisable trait such as: - appearance - clothing - posture - voice - habitual gesture - other people's reaction ### Scene Information Must Obey Access Limits Characters may only know what the scene allows them to know. Always separate: - what they can see - what they can hear - what they can physically reach - what they can correctly infer Seeing something is not the same as understanding it. Getting close once is not the same as having time to inspect it. ### Cognition Must Change Choice and Language Separate information distribution from epistemic status before consequential decisions or dialogue. Track what is shared, specialized, or private; then distinguish observation, report, judgment, intention, and misrecognition. Use shared understanding to constrain the scene without redundant exposition, specialized understanding to change risk or available methods, and individual cognition to determine what a character chooses, hides, distorts, tests, or says. Classify all of it relative to the current character, time, and audience. ### Viewpoint Does Not Own Every Decision Separate the experiential center, problem owner, decision owner, domain actor, and execution owner. POV determines what pressure reaches the reader first; it does not automatically grant authority, expertise, or permanent control. Make consequential scenes legible as perception or misinterpretation, narrowed options, decision, action, consequence, and reassessment. Treat that chain as a diagnostic scaffold, not a mandatory formula or a ban on omniscient and multi-viewpoint fiction. ### Dialogue Must Happen Through Behavior Sustained dialogue must remain embodied and consequential. A scene is not active merely because characters handle props or instruments. Distinguish functional action, which reveals strategy or changes the next beat, from decorative action and procedural action. Do not repair transcript-like prose by attaching an emotion label or arbitrary gesture to every line. Action density is not a quality target, and beats must not be distributed evenly. Rapid dialogue may remain untagged when voice and pressure are clear. Add a behavioral beat where information lands, a lie is tested, power shifts, restraint fails, or a decision changes. Causal and response checks are author-side diagnostics, not a sentence-by-sentence narration template. Let context, dialogue, action, and pauses carry the exchange; direct statements of silence remain valid. Use the distinctions in [Dialogue and Behavior](references/dialogue-and-behavior.md#causal-coherence-is-a-diagnostic-not-narration) rather than routinely explaining whether a remark was answered. ### Every Scene Segment Must Earn Its Place Every chapter or functional segment should be able to answer: - What problem is this segment handling? - Who changes here? - What moves forward by the end? If the prose only accumulates content without changing the situation, treat it as structurally weak. ### Protect Style-Bearing Material Default to preserving the author's: - dialogue - inner thoughts - digressive texture - observational details - flirtation or ambiguity - rhythm-bearing repetition You may fix wording, punctuation, transcription noise, and paragraph breaks. Do not silently compress concrete material into summary. ### Review Output Must Be Specific When reviewing fiction, report concrete findings with: - location - problem type - why it fails - how to revise it Avoid vague comments such as "a bit stiff" or "could be improved." ### Project Rules Override This Skill If the fiction project already has local rules, style contracts, or manuscript-layer rules, follow those first. This skill supplies general fiction method, not project-specific law. ## Working Pattern 1. Identify whether the task is project setup, a prose stage, or work across manuscript versions or author feedback; load the lifecycle reference before any project-folder or manuscript-file edits in those cases. 2. Identify the minimum context tier needed for the task. 3. Load only the reference files needed for that stage. 4. Load project-local outline, timeline, setting, and issue files before pulling extra prose. 5. Pull near-field full prose when revising language, style, or chapter continuity. 6. Apply the hard rules before proposing prose changes. 7. If reviewing, produce structured findings before any summary. 8. When delivering plain-text manuscript files, run the bundled checker as `python <skill-directory>/scripts/check_manuscript_text.py <file-or-directory>`; review warnings instead of treating them as automatic prose defects. 9. If the task touches sensitive realism constraints and you are unsure, say what must be verified instead of bluffing. ## References - [Planning](references/planning.md): chapter tasks, optional planning aids, and entry/exit state - [Story Outline and Causal Summary](references/story-outline-and-causal-summary.md): story-first ordering, causal bridges, knowledge layers, and final-form outline checks - [Cognition Layers and Language](references/cognition-layers-and-language.md): information source, epistemic status, decisions, and disclosure - [Scene Causality and Agency](references/scene-causality-and-agency.md): POV pressure, responsibility, action topology, and continuity - [Dialogue and Behavior](references/dialogue-and-behavior.md): embodied exchange, functional action, transcript prevention, and dialogue-heavy professional scenes - [Character Introductions](references/character-introductions.md): first-entry rules for characters - [Scene and Structure](references/scene-and-structure.md): scene progression and chapter shape - [Style Fidelity](references/style-fidelity.md): preserving authorial texture during revision - [Realism Constraints](references/realism-constraints.md): institutions, bodies, crowds, spectacle, and access limits - [Revision Checklist](references/revision-checklist.md): structured review output and manuscript hygiene - [Manuscript Lifecycle and Author Feedback](references/manuscript-lifecycle-and-feedback.md): project-local version stewardship, author-edit comparison, and evidence-based case retrieval
© wgwtest · MIT · wgwtest/novel-writing