A complete AI skill set for Chinese novel writing, built on the oh-my-openagent framework. 13 specialized skills covering every stage of fiction creation, from story ideation to manuscript completion, with core innovations in preventing worldbuilding info-dumps, reader-perspective testing, and AI-flavor elimination.
Dreampowers provides a disciplined, skill-driven workflow for writing Chinese fiction with AI. Instead of dumping all worldbuilding details into chapter one, it enforces gradual reveal through quantified rules and multi-stage review.
- 13 specialized skills covering the full fiction writing lifecycle
- Three-stage review for every chapter (plot, reveal check, prose)
- External review loop: reader + consistency review after three-stage pass, up to 3 iterations with TBD fallback for human review
- Six Iron Rules preventing info-dumps with zero-tolerance enforcement
- Foreshadowing tracking of setup/payoff across the entire manuscript
- Scene-type directing with sub-modes for action, emotional, and dialogue scenes
- Four continuous writing modes: per-chapter confirmation, batch N chapters, pause-at-act-break, full-auto
- Narrative timeline techniques: sequential/reverse/interpolated/supplementary with POV selection rules
- Description methodology: camera-language system integrated into scene directing
- External content import: bring in existing world set, character cards, outlines, and handwritten chapters
- Mid-story outline revision: 7-step freeze-and-revise workflow for course corrections
- Reader-perspective testing: cold-read simulation scoring page-turning desire, cognitive load, empathy, and pacing feel
- AI-flavor detection: quantified pattern matching to eliminate formulaic AI writing habits
- Writing style definition: 7-dimension questionnaire (narrative distance, sentence rhythm, sensory density, vocabulary register, metaphor strategy, emotional expression, dialogue ratio) with 54+ reference authors across 9 genres
- Opt-in mature content: intimate scene craft direction, installed separately via
--allflag - Three-layer isolation from coding skills (entry switching +
dp-prefix + directory isolation) - Git-based version management for manuscripts
| Skill | Purpose |
|---|---|
dp-using-dreampowers |
Entry skill: mode switching, skill routing, workflow overview |
dp-tool-research |
Story premise brainstorming, logline formula, platform comparison, internet-assisted worldbuilding research, author tuning |
dp-set-style |
Work-level writing style definition via 7-dimension questionnaire, style reference library (54+ authors across 9 genres), generates style.md voice profile |
dp-set-concept |
World/character set with iceberg annotations (50%+ underwater), concept isolation via file-level physical separation, story-level timeline construction, external content import |
dp-set-outline |
Act/chapter structure, 5 models, 6 iron rules with concept budget, foreshadowing tracking, theme weaving, narrative timeline techniques, POV rules, mid-story outline revision |
| Skill | Purpose |
|---|---|
dp-chapter-draft |
Chapter writing with pre-draft gate, 3-stage review, 4 continuous writing modes, timeline execution, handwritten chapter integration, era-accurate writing (period-correct objects and perspective), outline-as-skeleton discipline |
dp-chapter-summary |
Plain-text chapter summary (≤300 chars, no formatting) from output/chapter-NNN.md only, for cross-chapter continuity |
dp-chapter-direct |
Scene-type directing (action/emotional/dialogue sub-modes) + narrative pacing control, tension-relief law, camera-language description methodology |
dp-chapter-adult |
Opt-in: Intimate scene craft direction with intensity levels and emotional-depth rules |
| Skill | Purpose |
|---|---|
dp-character-style |
Character voice profiles, name-cover test, dialogue rules, subtext crafting |
dp-tool-version |
Git-based version management: structured commits, milestone tags, rollback, diff comparison |
| Skill | Purpose |
|---|---|
dp-review-reader |
Reader-perspective experience testing: page-turning desire, cognitive load, empathy verification, pacing feel |
dp-review-consistency |
Cross-chapter consistency verification (8 dimensions) + prose revision and AI-flavor detection/elimination + final completion checklist with full-book consistency scan |
Six Iron Rules enforced with zero tolerance:
- CURIOSITY_BEFORE_EXPLANATION: Mystery first, answers later
- SENSE_BEFORE_SYSTEM: Sensory experience before rules explanation
- EXPAND_DONT_ADD: Deepen existing concepts before introducing new ones
- CONCEPT_BUDGET: Ch.1 max 3 new concepts (can be 0), subsequent max 2/chapter (can be 0)
- NO_NARRATOR_LECTURES: All world info through character action/dialogue only
- JUST_IN_TIME: Reveal only what serves the current conflict
Individual Markdown files in docs/dreampowers/tracking/ (prefixed with thread-) tracking all foreshadowing threads with:
- Thread recovery pacing determined by outline planning, verified by consistency review
- Claremont Coefficient (CC = introduced - resolved), warning at CC > 2
- Overdue thread detection
Physical file-level isolation preventing AI from seeing future worldbuilding concepts or character states:
- One concept per file in
docs/dreampowers/set/concept/ - One character per file or directory in
docs/dreampowers/set/character/ - Chapter workspace folders (
docs/dreampowers/chapters/chapter-NNN/) as fully self-contained writing units - Chapter folders contain: symlinks to concepts/characters/foreshadowing threads +
spec.md(concept budget, gating criteria, dependency graph, reader assessment requirements) - Iron rules symlinked from
docs/dreampowers/tracking/iron-rules.mdinto each chapter folder - Previous chapter summaries symlinked (1-3 prior chapters as needed) for continuity
- AI reads only the chapter folder contents during writing, never the source directories directly
- Serial chapter writing: chapters must be written one at a time, in order
Every chapter passes through three sequential review stages:
- Plot Review: Conflict progression, character consistency, outline adherence
- Reveal Check: Zero-tolerance info-dump detection against the 6 Iron Rules
- Prose Review: Style consistency, dialogue voice, sensory detail, rhythm
After passing the three-stage review, every chapter enters an external review loop:
- Reader review (dp-review-reader): cold-reader experience testing across four dimensions
- Consistency review (dp-review-consistency): eight-dimension consistency check + prose revision with AI-flavor detection + writing style verification against style.md
- Fix and repeat: issues from both reports are merged and fixed, then the loop repeats (max 3 iterations)
- TBD fallback: if issues persist after 3 iterations, the output file is saved as
chapter-NNN-TBD.mdfor human review
Integrated across skills for consistent technique application:
- Timeline Methods (dp-set-outline, dp-chapter-draft): sequential, reverse, interpolated, supplementary with decision tree for selection and chapter-level execution guidance
- POV Rules (dp-set-outline): First/third-limited/third-omniscient comparison, 5 hard rules for POV switching (scene-boundary only, emotional anchor, no info-smuggling, consistent within scene, signal transitions)
- Description Methodology (dp-chapter-direct): Camera-language system integrated with the three scene sub-modes
- Four writing modes: per-chapter confirmation (default), batch N chapters, pause-at-act-break, full-auto (with quality gate)
- Mid-story outline revision: 7-step freeze-and-revise workflow
Simulates a "cold reader" who has no access to outlines, world set, or character cards, only the accumulated text of published chapters. Four test dimensions:
- Page-Turning Desire: Does the reader want to keep reading? (1-5 scale, flag at 2 or below)
- Cognitive Load: Is the reader confused or overwhelmed? (inverted scale, more than 3 unfamiliar elements in 500 words = overload)
- Empathy Verification: "Why should I care?" test for character motivation clarity
- Subjective Pacing Feel: "Want to skip this?" (dragging) vs "Wait, what happened?" (rushing)
Quantified pattern matching across 6 layers to eliminate AI writing habits:
- Lexical: Simile-word flooding, emotion labeling, degree-adverb stacking
- Syntactic: Triple-parallelism reflex, uniform sentence length, subject repetition
- Structural: Fixed three-paragraph pattern per scene, formulaic transitions, every ending is an epiphany
- Rhetorical: Synesthesia overuse, metaphor density, universal personification, false antiquity (nostalgia filter on period settings), era-anachronism detection (objects that don't belong to the story's time period)
- Formatting: Hyphen/em-dash intrusion (zero tolerance), emoji in prose (zero tolerance), Markdown asterisk marks (zero tolerance), HTML entity characters (zero tolerance)
- Register: Technical terminology density, definition-style sentences, passive/agentless constructions, enumeration patterns, over-explicit causal chains
- OpenCode with oh-my-openagent installed
# Clone or copy the skill_dreampowers directory
cd skill_dreampowers
# Install default skills (12 skills, excludes opt-in)
chmod +x install.sh
./install.sh
# Or install ALL skills including opt-in (e.g. dp-chapter-adult)
./install.sh --allThis creates symlinks from each dp-* skill directory to ~/.config/opencode/skills/. Opt-in skills (like dp-chapter-adult) are skipped by default; use --all or manually symlink them.
./uninstall.shRemoves all dp-* symlinks from ~/.config/opencode/skills/.
Switch to creative writing mode in an OpenCode session:
Use skill dp-using-dreampowers
This activates the Dreampowers routing table. The AI will now use dp-* skills for all tasks.
Starting a new novel from scratch, follow the full workflow:
我想写一个仙侠小说
The AI loads dp-tool-research and walks you through story premise brainstorming (genre, logline, platform fit). After you confirm a premise (stored as overview.md), it transitions automatically:
dp-tool-research --> dp-set-style --> dp-set-concept --> dp-set-outline --> dp-chapter-draft (loop)
Each stage produces artifacts and flows into the next. You don't need to manually invoke skills; the routing table handles transitions.
If you already have worldbuilding documents, character cards, outlines, or written chapters, use the import entry point:
我有一份已经写好的世界观设定，帮我导入
The AI loads dp-set-concept, identifies the content type, and processes it accordingly:
| What you provide | What happens |
|---|---|
| World set document | Iceberg annotation (50%+ marked "underwater"), stored in docs/dreampowers/set/world/ |
| Character cards | Voice profile extraction, stored in docs/dreampowers/set/character/ |
| Story outline | Structure validation, foreshadowing planning via dp-set-outline |
| Written chapters | Metadata injection, foreshadowing backfill, continuity check |
| Mixed document | Auto-split by type, each piece processed separately |
When the AI has written chapters 1-5 and you've handwritten chapters 6-8, you need the AI to continue from chapter 9:
AI 写到了第5章，第6到第8章是我自己写的，请接入后从第9章继续
[provide chapters 6-8]
The AI will:
- Create chapter workspace folders for your chapters
- Scan for foreshadowing threads and create/update tracking files
- Run continuity checking against existing chapters
- Generate chapter summaries via
dp-chapter-summary - Build a handoff context card summarizing where the story stands
- Enter the normal
dp-chapter-draftPre-Draft Gate for chapter 9
Your handwritten chapters are never modified, only analyzed.
Before the first chapter (or at any chapter break), the AI asks you to choose a writing pace:
| Mode | Behavior | Best for |
|---|---|---|
| Per-chapter confirmation (default) | Write, 3-stage review, you confirm, next chapter | Fine-grained control over every chapter |
| Batch N chapters | Write N chapters (each still reviewed), deliver all at once | Batch progress, review later |
| Pause at act break | Write until the current act ends, then pause | Managing progress by story arcs |
| Full-auto | Write all remaining chapters, deliver complete draft | Maximum trust in the workflow |
Mode switches take effect immediately. You can change anytime.
You can stop the AI at any point:
停
暂停
To resume:
继续写
从第12章开始重写
第二幕的走向不对，我想改大纲
This triggers the 7-step outline revision workflow in dp-set-outline.
Quality-layer skills can be invoked anytime:
帮我检查第3章到第7章的连续性
用读者视角测试一下第5章
检查最近三章的AI味
| Request | Skill invoked |
|---|---|
| Continuity check | dp-review-consistency |
| Reader-perspective test | dp-review-reader |
| AI-flavor detection + prose polishing | dp-review-consistency |
| Scene direction help | dp-chapter-direct |
| Character voice audit | dp-character-style |
After reading completed chapters, you may want to adjust the direction of upcoming chapters:
我看完前五章了，后面节奏太慢，想调整
作者调优
The AI loads dp-tool-research in author tuning mode, walks you through which unwritten chapters to adjust, collects your feedback (pacing, foreshadowing timing, character focus, tone shifts, etc.), and writes a tuning.md directive file into each target chapter's workspace folder. The tuning.md takes priority over outline defaults during drafting — only the Iron Rules outrank it.
Already-completed chapters are not affected; use the revision workflow for those.
提交当前进度
给第一幕打个里程碑标签
第3章改坏了，回滚到审查通过的版本
See dp-tool-version for the full commit convention.
Use skill using-superpowers
This deactivates all dp-* skills. Your manuscript and all Dreampowers artifacts remain in the project directory.
| You want to... | Say... |
|---|---|
| Start a new novel | 我想写一个[genre]小说 |
| Import existing materials | 导入我的世界观/角色卡/大纲/章节 |
| Continue writing | 继续写 / 从第N章继续 |
| Pause | 停 / 暂停 |
| Change writing mode | 改成连写N章 / 改成逐章确认 |
| Revise the outline | 我想改大纲 |
| Check continuity | 检查第X到第Y章的连续性 |
| Test reader experience | 用读者视角测试第N章 |
| Remove AI flavor | 检查AI味 |
| Tune upcoming chapters | 作者调优 / 我想调整后面的方向 |
| Manage versions | 提交 / 打标签 / 回滚 |
| Switch to coding | Use skill using-superpowers |
skill_dreampowers/
├── install.sh # Symlink installer
├── uninstall.sh # Symlink remover
├── README.md # English documentation
├── README_zh.md # Chinese documentation
└── skills/
├── dp-using-dreampowers/ # Entry skill
├── dp-tool-research/ # Story brainstorming + research
├── dp-set-style/ # Writing style definition + voice profile
├── dp-set-concept/ # World/character set + concept isolation + import
├── dp-set-outline/ # Outline + reveal pacing + theme weaving
├── dp-chapter-draft/ # Chapter writing + review
├── dp-chapter-summary/ # Chapter summary generation
├── dp-chapter-direct/ # Scene directing + pacing control
├── dp-character-style/ # Character voice profiles + subtext
├── dp-tool-version/ # Git version management
├── dp-review-reader/ # Reader-perspective testing
├── dp-review-consistency/ # Continuity + revision + final check
└── dp-chapter-adult/ # Opt-in: intimate scene craft
When used in a project, Dreampowers creates the following directory structure:
docs/dreampowers/
├── input/ # User-imported raw data (temp area)
├── set/
│ ├── world/ # World overview (logic, background)
│ ├── concept/ # Concept source files (one per file)
│ └── character/ # Character source files/directories
├── tracking/ # Cross-chapter persistent tracking
│ ├── overview.md # One-line story summary
│ ├── iron-rules.md # Iron rules (symlinked to chapter folders)
│ ├── style.md # Writing style profile (symlinked to chapter folders)
│ └── thread-NNN-*.md # Foreshadowing threads (thread- prefix)
├── timeline/ # Timeline + chapter summaries
│ ├── timeline.md # Initial timeline definition
│ └── summary-NNN.md # Per-chapter summaries
├── outlines/ # Outlines + full-book reviews
│ ├── outline-*.md
│ └── review-*.md
└── chapters/ # Chapter workspaces
└── chapter-NNN/
├── spec.md # Metadata (budget, gating, dependencies)
├── draft.md # Draft (work in progress)
├── review.md # Chapter-level review report
├── tuning.md # Author tuning directives (optional, high priority when present)
├── *.md -> set/concept/* # Concept symlinks
├── *.md -> set/character/* # Character symlinks
├── thread-*.md -> tracking/*# Foreshadowing symlinks
├── iron-rules.md -> tracking/iron-rules.md
├── style.md -> tracking/style.md
└── summary-*.md -> timeline/summary-*.md (prev 1-3 chapters)
output/ # Final versions
└── chapter-NNN.md
Chapter folders (docs/dreampowers/chapters/chapter-NNN/) are fully self-contained writing units. During writing, the AI reads only the chapter folder contents, never the source directories directly. Only invisible information is guaranteed unused.
Dreampowers and Superpowers (the coding skill set) are completely independent. Three-layer isolation ensures no interference:
- Entry switching:
dp-using-dreampowersandusing-superpowers - Namespace: All Dreampowers skills prefixed with
dp- - Directory: Installed via symlinks, separate from Superpowers source
GPL-3.0