虚构锻造
长篇写作引擎
火花
定义你故事的灵魂。选择体裁、目标篇幅，并种下故事前提的种子。
决定角色阵容规模和故事复杂度。
为代笔作者注入一个持续生效的文风指纹。
正在锻造故事前提…
角色锻炉
定义你的主要角色阵容，并让 AI 根据故事长度生成中层与次要配角。
正在锻造核心与配角角色……
核心阵容
配角阵容（中层与次要）
世界建造者
构筑冰山：表层、浅层、深处与深渊。
正在探入深处……
建筑师
分形大纲：从骨架到完整章节节拍。
正在起草骨架……
锻炉
代写循环。左侧为大纲，右侧为手稿。
Generating Beat...
装订坊
在出版前设计书籍封面，打磨完美书名，并撰写封底文案。
No Cover Art
Painting masterpiece...
如何使用 Fiction Forge
欢迎使用终极 AI 长篇写作引擎。按照 6 个阶段一步步打造你的杰作。
Phase 1: The Spark
定义你的体裁、目标字数和故事前提。引擎会根据你的初始提示，动态向你采访世界观细节。
Phase 2: Character Forge
角色阵容规模会自动调整。我们生成心理侧写、MBTI 性格类型，并可通过 Flux 模型绘制人物视觉肖像。
Phase 3: Worldbuilder
基于四层“冰山理论”，我们为情节建立表层、浅层、深层与深渊层的潜文本结构，为故事增添层次。
Phase 4: Architect
分形大纲会映射一个七点骨架，计算章节数量，并以数学方式将故事划分为微节拍。
Phase 5: The Forge
无限循环的代笔引擎。系统依托你的大纲跟踪器和滚动记忆缓存无缝写作。让它尽情运转！
Phase 6: The Bindery
生成标题、文案和 AI 封面艺术。最后点击“导出”，将整部手稿动态编译为可印刷的 PDF 格式。
v5.2 - 反馈记忆缓存
AUTHOR CORRECTION CACHE: Added a persistent feedback loop in Phase 5. Authors can input global rules (e.g., "Stop making characters smirk") which are stored in the state manager and injected as absolute commands in every subsequent ghostwriter prompt.
v5.1 - 全局操控与审计更新
GLOBAL STEERING AXES: Added live, hot-swappable dials in Phase 5 to dynamically control Tone, Action, Stakes, and Pacing. The active states are injected into the context of every single beat generated.
PROSE AUDITOR & RED PEN PASS: Implemented a $00 cost JavaScript regex auditor that intercepts generated prose and detects adverb clusters, repetitive sentence starters, and monotonous pacing. Flawed beats trigger a silent "Red Pen" rewrite using the fastest logic model to automatically repair the prose before inserting it into the manuscript.
CREATIVE DNA INTERVIEW: Replaced blank input fields in Phase 1 with an active binary/trinary questioning system that forces authors to build tension and friction into their premise.
v5.0 - 严格策展与矩阵大改版
STRICT CURATION ARCHITECTURE: Replaced dynamic/fuzzy credit-bucket sorting with a strict, hardcoded agent matrix. Prevents "clever" fallback overrides that obscured which AI models were actually firing.
ZERO CREDIT ENFORCEMENT: Draft Tier is now mathematically locked entirely to $00 cost agents. If no free model is available, the engine safely halts instead of silently charging credits.
MASTERWORK HONESTY: Removed logic that secretly downgraded outlining tasks on the Masterwork tier. If you pick Masterwork, you get Claude Sonnet for every single task, no exceptions.
v4.7 - 氛围切换与智能代理路由
VIBE SWITCH ARCHITECTURE: Added a user-facing toggle that masks complex model routing. "Standard" uses Gemini 3.0 Flash for clever plotting. "Edgy" routes to Grok 4.1 Fast for dark, anti-trope narratives.
THE WORKER BEE HANDOFF: Restricted Gemma 4 31B purely to Phase 5 prose generation. Since it no longer handles JSON logic, generation lag is cured while preserving masterwork-tier prose quality.
PARALLEL EXECUTION: Optimized The Bindery and Character Forge with concurrent execution, drastically reducing wait times.
v3.0 - 残酷真相节奏大改版
MATH CALIBRATION: Massively reduced word-count padding targets (e.g., from 750 down to ~450 for Premium) per beat to accurately map to human action boundaries, saving tokens and improving pacing.
POSITIVE FORMULAS: Added explicit positive constraints for "The 80/20 Rule", "In Media Res", and "Declarative Action" to eliminate internal monologue, inventory checks, and over-explaining gestures.
EXTERNALIZED EMOTION: Replaced tone planning with "Externalized Emotion" in the <plan> block, forcing the AI to lock in physical manifestations of feelings before generating prose.
v2.1 - 手稿完成界面更新
PERSISTENT UI STATE: The Forge tab now detects when `completedBeatsCount` equals `flatBeats.length` and permanently morphs the writing button into a disabled green "Manuscript Complete" state.
v2.0 - 文风执行与代理更新
POSITIVE FORMULAS: Swapped negative bans (e.g., "don't use filter words") for positive formulas ("Direct Camera Rule: Remove the observer"). This prevents Mid-Tier models from hyper-fixating on banned concepts.
AGENCY-FORCED OUTLINING: Rewrote the Phase 4 Architect instructions so every single beat is formatted around a protagonist's tactical choice, ensuring the writer never receives a passive scene prompt.
WEAPONIZED DIALOGUE: Banned characters from explaining the world to each other. They must use dialogue to manipulate, attack, or hide information.
v1.9 - 红笔优化
DEEP POV LOCKDOWN: Forced the model to respect physical limitations of the POV character (e.g., if deaf, no sound descriptions).
FILTER WORD BAN: Banned lazy writing crutches ("saw", "felt", "heard", "noticed") to enforce showing over telling.
ACTION CHOREOGRAPHY: Banned summarizing conflict. AI must describe physics, momentum, and visceral impact.
v1.8 - 微锚点与即时规划（JIT）
MICRO-ANCHOR (CONTEXT DIET): Severely truncated the prose-feeding memory window from ~600 words to just ~60 words. This strictly confines the LLM to the immediate physical present, entirely destroying the "echo chamber" loop of atmospheric bloat and forcing relentless, fast-paced action.
DYNAMIC CAST CONSTRAINTS: The engine now pre-parses upcoming beats to identify active characters in the scene. It weaponizes their stored Phase 2 profiles (Flaws, MBTI, Motivations) directly into the prompt as hard constraints, injecting natural friction into character dialogue and interactions.
JUST-IN-TIME (JIT) PLANNING: Added a forced Chain-of-Thought <plan> block execution prior to generating prose. The ghostwriter must explicitly define the POV, unspoken conflict, and physical action anchors before writing, structurally eliminating purple prose and internal monologue loops.
v1.7 - 智能缓存与专业装订
SMART CACHE (ROLLING MEMORY): Architect engine now maps dynamic inventory, active subplots, and physical injury status on a chapter-by-chapter level, passing this memory directly to the Worker Bee prose prompt to eliminate AI amnesia.
PRO BINDERY PDF: Rewrote the export compiler to generate a professional Title Page, Copyright & Disclaimer page, and a dynamic Table of Contents before the main manuscript.
MANUSCRIPT TYPOGRAPHY: Standardized print layout, automatically dropping the text-indent on the first paragraph of every chapter while enforcing perfect justification for the remainder.
v1.6 - 上下文瘦身与工蜂
CONTEXT DIET: Severely truncated the rolling memory context window to 150-600 words depending on tier. Prevents AI from falling into the "recap trap" and dramatically reduces input token costs.
DENSE BEAT ARCHITECTURE: Increased standard beats to 4 per chapter but reduced the ceiling per beat (e.g., 600 words). Creates vastly superior pacing, eliminates purple prose padding, and saves output tokens.
WORKER BEE ROUTING: Premium logic strictly routes to Claude 4.6 Sonnet / Gemini 3.1 Pro, but outsources raw prose execution to Gemini 2.5 Flash / Grok 4.1 Fast, effectively dropping the Masterwork cost by up to 85%.
v1.5 - 合著者（中阶）大改版
MATH CALIBRATION: Scaled Mid Tier pacing down from 1,000 words/beat to 800 words/beat (2,500 words per chapter). This aligns perfectly with the natural output limitations of rapid models like Haiku and Flash, eliminating prose padding and rambling.
CLAUDE HAIKU PROMOTION: DeepSeek V3 was still handling structural logic for the Mid Tier, risking the same timeouts. Claude 3.5 Haiku has been promoted to the primary structural brain for Mid Tier due to text reasoning speed.
v1.4 - Gemini Flash 推广版
MATH CALIBRATION: Adjusted Premium tier chapter calculations to estimate 3,400 words per chapter. This corrects a pacing bug where the engine would over-generate text.