Fiction Writer

Structured creative writing using Obsidian files (source of truth) + Writing Engine UI (visual workspace) + Claude Code (full-context AI) + Voice API (hands-free dictation & TTS).

Reference files (read on demand)

SKILL_DIR = {home}/.claude/skills/creative-fiction-writer/. The harness loads only this file; read a sibling with the Read tool when you reach the step that needs it.

File	Priority	Read when
<SKILL_DIR>/fiction-memory.md	Required	Always, first — active projects, last-session state, accumulated craft notes
<SKILL_DIR>/templates.md	Required	Scaffolding a project or writing any brief / outline / progress / scene / character / world / _project.yml file
<SKILL_DIR>/writing-engine.md	Optional	Launching or explaining the Writing Engine UI, or the tool architecture
<SKILL_DIR>/voice-mode.md	Optional	The user wants voice writing — dictation, command mode, conversation, or TTS proof-listening

Always read <SKILL_DIR>/fiction-memory.md first — it holds active projects, last-session state, and accumulated craft notes.

Reference

• Memory: <SKILL_DIR>/fiction-memory.md
• Writing Engine: 10_Projects/writing-engine/ (web UI on port 7800)
• Projects: 10_Projects/<Title>/ (each novel/memoir/story is a project folder)

Three AI Brains, Three Strengths

Tool	Best For	Context
Claude Code (terminal)	Full-context drafting, continuity checks, structural revision	Reads entire project
Voice API (mic button)	Hands-free scene generation, dictation, TTS proof-listening	UI packages context into system_prompt
Writing Engine AI panel (Ollama/OpenAI)	Quick in-editor actions: expand, compress, grammar	Current scene + codex selection

Tool architecture + ASCII diagram: <SKILL_DIR>/writing-engine.md.

---

Trigger Table

User Says	Mode	Action
"写小说" / "new novel" / "start a novel"	new-project	Scaffold novel project
"写回忆录" / "new memoir"	new-project	Scaffold memoir project
"写短篇" / "new short story"	new-project	Scaffold short story project
"写大纲" / "outline" / "plan the story"	outline	Create/refine story outline
"写场景" / "draft scene" / "write next scene"	draft	Write scene with full context
"扩展" / "expand this"	revise:expand	Expand passage with detail
"压缩" / "compress"	revise:compress	Tighten prose
"对话润色" / "dialogue pass"	revise:dialogue	Improve dialogue voice
"文风统一" / "voice pass"	revise:voice	Consistent narrative voice
"连续性检查" / "continuity check"	continuity	Check consistency
"角色档案" / "new character"	character	Create/update character
"世界设定" / "worldbuilding"	world	Create/update world note
"组装章节" / "assemble chapter"	assemble	Compile scenes → chapter
"项目状态" / "story status"	status	Progress report
"打开写作界面" / "open writer" / "writing UI"	ui	Launch Writing Engine
"语音写作" / "voice mode" / "talk to write"	voice	Activate voice mode in UI
"读给我听" / "read it back" / "TTS"	voice:tts	TTS playback of current scene/chapter

Project scaffolds + all note templates: <SKILL_DIR>/templates.md.

---

Core Workflow

new-project → outline → [DRAFT LOOP] → assemble → revision passes

Draft Loop (per scene)

1. Claude reads context (see Context Loading Protocol)
2. Claude drafts scene → writes to manuscript/<path>.md
3. User reviews in Writing Engine UI or Obsidian
4. User annotates / requests changes via Claude Code
5. Claude revises
6. Every 3-5 scenes: continuity check
7. Update _progress.md

---

Context Loading Protocol

Before writing ANY scene, Claude MUST read (in order):

1. _brief.md — premise, themes, tone, audience
2. _outline.md — find the current scene's beat/purpose
3. _progress.md — what's done, open threads, continuity flags
4. Previous 2 scenes — for continuity and flow
5. POV character file — voice, speech patterns, arc position
6. Other characters in scene — relationships, tensions
7. Relevant world file — if new setting introduced

This is what makes Claude Code superior to the Writing Engine's generic AI panel: every scene is written with full project context.

Context Loading — Quick Reference

READ _brief.md          → tone, themes
READ _outline.md        → this scene's beat
READ _progress.md       → what's happened, open threads
READ prev 2 scenes      → continuity, voice
READ character files    → POV voice, relationships
READ world files        → setting details (if needed)

---

Mode Details

new-project

1. Ask user for: title, type (novel/memoir/short), premise, language
2. Create project folder in 10_Projects/<Title>/
3. Scaffold all folders and template files
4. Create _project.yml for Writing Engine compatibility
5. Write initial _brief.md from user's premise
6. Update fiction-memory.md with new active project

outline

1. Read _brief.md for premise and themes
2. Propose act/chapter structure with scene beats
3. Iterate with user until outline is solid
4. Write to _outline.md
5. Generate initial _progress.md scene table

draft

1. Execute full Context Loading Protocol (see above)
2. Identify next scene from _progress.md (or user specifies)
3. Write scene matching: outline beat, character voice, tone, previous flow
4. Save to manuscript/<path>/<scene>.md
5. Update _progress.md (status → draft, word count)

revise:expand

Read the scene, expand thin passages with:

• Sensory detail (sight, sound, smell, touch, taste)
• Internal thought (for POV character)
• Environmental texture
• Micro-actions and body language

revise:compress

Read the scene, tighten by:

• Removing redundant descriptions
• Cutting filler words and phrases
• Combining sentences where possible
• Preserving voice and key imagery

revise:dialogue

Read the scene + all character files for characters present:

• Make each character sound distinct
• Remove "said" synonyms overuse
• Add subtext (what's unsaid)
• Ensure dialogue advances plot or reveals character

revise:voice

Read _brief.md tone notes + sample of earlier scenes:

• Flag voice inconsistencies
• Normalize POV distance
• Harmonize tense usage
• Suggest specific fixes

continuity

Scan recent 3-5 scenes + character files + world files:

• Character consistency: behavior matches profile? contradictions?
• Timeline: chronological logic? impossible overlaps?
• Plot threads: which are open? any dropped?
• Setting details: physical descriptions consistent?
• Tone/voice: has narrative voice drifted?

Output: Update _progress.md with findings under "Continuity Flags"

character

1. Ask: name, role, want/need/flaw/arc
2. Create file in characters/ (or people/ for memoir)
3. Link to relevant scenes in outline
4. If updating existing: read current file, preserve existing info, add new

world

1. Ask: name, type, description, rules, significance
2. Create file in world/ (or places/ for memoir)
3. Link to scenes where it appears

assemble

1. Read all scenes for target chapter (sorted by scene number)
2. Check transitions between scenes
3. Add chapter-level framing if needed
4. Calculate chapter word count
5. Write assembled chapter to revisions/ch01-assembled.md

status

Report on demand:

• Total word count (by chapter/act)
• Scene completion status (draft/revised/final)
• Open plot threads
• Character appearance tracker
• Continuity flags
• Estimated completion %

ui

Launch the Writing Engine:

cd "10_Projects/writing-engine" && python server.py &

Then open http://localhost:7800 in browser.

Note: The Writing Engine reads from 10_Projects/ — it sees the same files Claude Code writes.

---

Writing Engine UI details: <SKILL_DIR>/writing-engine.md. Voice mode (dictate / command / conversation + TTS): <SKILL_DIR>/voice-mode.md.

---

Writing Principles

These guide Claude's drafting and revision:

1. Show, don't tell — action and sensory detail over abstract statements
2. Character voice distinction — each character sounds different
3. Scene purpose — every scene must advance plot OR reveal character (ideally both)
4. Tension on every page — conflict, mystery, or emotional stakes
5. Specific over general — "a cracked blue mug" not "a cup"
6. Subtext in dialogue — what's unsaid matters more
7. End scenes with forward momentum — reader wants to turn the page

---

Post-Action

After any mode execution:

1. Update _progress.md if scene status changed
2. Update <SKILL_DIR>/fiction-memory.md whenever a project changes (new or active project, phase/status shift) or a significant craft insight emerges — it is the durable memory across sessions
3. Report what was written/changed and next suggested action