AI Fiction Team (Market Research)

1. Team Introduction and Industry Background

1.1 Positioning of This Team Within the AI Harness System

AI Harness is the engineering carrier layer between foundation models and real business systems; it converts the model's uncertainty into engineering-level predictability (see the core definition in the parameter cards). AI fiction is the most tension-laden proving ground for this definition:

• The model's "uncertainty" is not a flaw in the fiction scenario but a product feature — what readers want is surprise;
• yet "unpredictability" must be constrained in engineering terms — after hundreds of thousands of words, a broken character or a forgotten planted detail means the work is ruined.

This contradiction determines the core point: AI fiction is not a model-capability problem but a Harness engineering problem. The 8 platform documents in this team are, in essence, 8 engineering answers to the same question.

Decomposed according to the six-layer Harness capability model, the pressure distribution in the AI fiction scenario is:

1.2 Industry Development Trajectory

The evolution of AI fiction tools is largely in sync with the three-generation Harness architecture evolution, but has its own timeline:

1.3 Core Thesis of This Team

AI fiction is the scenario with the greatest pressure on L4 (Memory & State) and L1 (Context Engineering) among the six Harness layers. The core challenge of long serialized works is "still not breaking characters or forgetting planted details after hundreds of thousands of words", which is essentially a long-horizon state management problem rather than a text generation problem. Whoever makes L4 solid — Bible, State, Timeline, symlink snapshots, Lorebook triggers — is the one who can support long-form works.

This thesis has three lines of supporting evidence:

1. Evidence on the problem side: the web-novel industry names this failure mode "chishu" (吃书, contradicting earlier settings), showing it is an inevitable problem of long serialized works rather than an occasional one. Yuewen's "Miaobi Tongjian" is therefore positioned not as "chapter generation" but as a "second brain" — dedicated to planted-detail mining and detail retrieval.
2. Evidence on the solution side: the whole industry has converged on three L4 paradigms — structured rolling summaries, keyword-triggered entry libraries, and versioned state snapshots (see Section 3.4). The three strongest platforms each occupy one; the paths differ but the goal is the same.
3. Evidence on the benefit side: Claude Book's State snapshots let "writing Chapter 15 precisely retrieve what happened in Chapters 1—14 without needing 100K tokens of context" — this is the reverse decompression of L1 that results from making L4 solid.

Corollary: to evaluate an AI fiction tool's long-form capability, do not look at its single-generation quality, but at its state-layer design.

2. Glossary

3. Platform Landscape & Cross-Comparison Matrix

3.1 Platform List & Grouping

This team covers 8 platforms / workflows, divided into four categories by creative form:

Note: "Xiezuo Shi" (写作狮), originally suggested for coverage in the task brief, returned no results at all after 3 rounds of searches and was removed from the platform list; "Mita Writing Cat" (秘塔写作猫) and "Biling AI" (笔灵 AI), upon verification, are primarily positioned for academic / workplace writing rather than web novels, so no dedicated page is created for them, and the positioning conflict is truthfully explained in 07-chinese-webnovel-ai.md.

3.2 Cross-Comparison of Basic Information

3.3 Harness Six-Layer Capability Maturity Matrix

Key points for reading this table:

1. The strong correlation between L4 and L1 is obvious: for platforms with a solid L4 (Claude workflow, Yuewen, NovelAI, Lingxie), L1 pressure is generally relieved by externalizing state; platforms with weak L4 (Fanqie AI, Biling AI) generally expose "continuity error" problems in long-form works.
2. L6 strength is strongly correlated with the business model: Fanqie, as a distribution platform, has the strongest L6 because it bears review and compliance responsibility; NovelAI, as a subscription tool, embodies L6 in privacy (encryption, no training) rather than content compliance.
3. L5 is a widespread weakness: most platforms only have weak evaluation such as "multi-candidate regeneration", lacking quantifiable quality gates. The only exceptions are the Claude workflow's Perplexity Gate and Sudowrite's AI Beta Reading.

3.4 Comparison of Three Long-Range State Management Paradigms

This is the most critical cross-comparison in this team, directly corresponding to the core thesis:

Engineering recommendation: the three paradigms are not mutually exclusive but complementary. Production-grade long-form workflows are best used in combination — B manages "entry-level settings" (character profiles, techniques, locations), C manages "timeline-level state" (who is where, what they know, what they hold), and A serves as the fallback compression layer.

4. Harness Six-Layer Design Points

图 4-1｜AI 小说场景：六层 Harness 模型压力分布

数据来源：基于本文分析绘制的示意图。

4.1 L1 Context Engineering Layer

Context assembly in the AI fiction scenario must answer four questions simultaneously; missing any one causes problems:

How the four assembly strategies are implemented in the industry:

• Full injection: Sudowrite reads the first 20,000 words of the manuscript; Chapter Continuity can link up to 25 documents.
• Conditional trigger: NovelAI Lorebook, AI Dungeon Story Cards — inject only when a keyword hits.
• State externalization: Claude Book uses state/current/ snapshots to replace the "full preceding text", compressing 100K tokens into a single state file.
• Pruning agent: Chinese practice introduces a dedicated subagent that gathers one chapter's context, producing a distilled result for the writing Agent, avoiding main-context bloat.

4.2 L2 Tool & Execution Layer

The tool layer in the AI fiction scenario has a counterintuitive rule: the more "boring" the tool, the more critical it is.

Conclusion: "tools as scripts" is the maturity hallmark of third-generation Harness in the fiction scenario — stripping deterministic tasks away from the model so that the model only does uncertain creative work.

4.3 L3 Orchestration & Control Layer

Three orchestration forms:

1. Stage pipeline (Fanqie AI, Lingxie Creation): strung together by creation stage — before starting the book (inspiration → outline) → during writing (continuation / blocking / expansion-rewrite) → before submission (review).
2. Multi-role agent pipeline (Claude workflow): orchestrator-worker architecture, PLANNER writes beats → WRITER generates → REVIEWERS verify in parallel → STATE-AGENT updates state.
3. Human-in-the-loop cadence (MidReal, AI Dungeon): hands control back to a human every 300 words, or drives the narrative loop through free player input.

The key engineering detail of the multi-role agent pipeline is the responsibility boundary: Claude Book explicitly requires the reviewer subagent to only verify, not write fiction; the writers are only coordinator and writer. This boundary prevents "review agents overstepping to rewrite and causing state inconsistency".

4.4 L4 Memory & State Layer

The core layer of this team. A complete state layer should contain six types of information:

The last item is the easiest to overlook, yet it is the biggest trap of Agent writing: if the knowledge graph stores all the information of the whole book, the model will "accidentally" reveal the secret of Chapter 30 as early as Chapter 3. The countermeasure among Chinese practitioners is — the knowledge graph retains only the information at the novel's current progress, auto-syncing after every plot advancement.

4.5 L5 Evaluation & Observability Layer

Evaluation in the AI fiction scenario must be layered; mixing layers leads to wrong decisions:

Claude-Code-Novel-Writer makes an explicit statement on this: automation/quality-check.sh only records mechanical signals such as word count, paragraph length, and dialogue ratio; it can reveal anomalies but does not judge literary quality; revision decisions should be based on chapter intent, voice, pacing, continuity, motivation, and reader effect. This "boundary statement" is itself a manifestation of L5 maturity.

Meanwhile, L5 also includes business observability. Yuewen is one of the few platforms disclosing verifiable business metrics: since the new version launched, Miaobi's daily active users have more than doubled, average daily token consumption has grown over 90%, author–AI interaction frequency has more than doubled, driving a 56% year-over-year increase in "Writer Assistant" daily active users (source: Tencent News, "Comic drama annual growth over 45%"); the weekly usage rate of the Miaobi large model by writers has exceeded 75% (source: Geek Park).

4.6 L6 Governance & Safety Layer

L6 in AI fiction has three main threads, and Chinese platforms and overseas platforms focus on different things:

For the Chinese-side compliance baseline and the strongest governance sample, see Section 6.

5. Harness Highlight Materials

5.1 Claude Book's Four Directories & Symlink Snapshots

The responsibilities and change strategy of the four directories:

The State versioning + symlink model:

state/
├── current/      # Symlink → 最新章节状态
├── chapter-01/   # 第 1 章后归档
├── chapter-02/
└── template/

Core benefit (in the original's wording): when writing Chapter 15, the model can precisely retrieve what happened in Chapters 1—14 without needing 100K tokens of context.

bible/ can be automatically extracted from source books by the book-analyzer skill to obtain quantitative style rules: average sentence length, dialogue-tag frequency, vocabulary constraints, and explicitly forbidden elements; bible-merger can merge analysis results across multiple source books.

Example of forbidden elements: under the Famous Five children's-book style, the bible explicitly writes "character death (even the villain is only arrested)" and "no bloody violence or detailed injuries". The original beats wrote "the keeper Mr. Le Goff fell down the stairs in the storm, and the body was only found two days later"; the style linter judged it too dark; it was rewritten to "Mr. Le Goff vanished in that year's storm... no one ever saw him again".

5.2 Perplexity Gate Against the AI Slope

Design note: turning "anti-mediocrity" into an automatic checkpoint rather than a post-writing subjective feeling is a successful attempt to convert a non-quantifiable literary problem into an engineering signal.

5.3 NovelAI Lorebook Keyword Triggering

This is the industrialized prototype of Paradigm B (keyword-triggered entry library); it saves more context than "injecting the full preceding text" and is more controllable than a "rolling summary".

5.4 Lingxie Creation SoloEnt.md Project Constitution

The project constitution centrally records seven types of information:

1. Worldview rules
2. Character personality and behavioral boundaries
3. Main-plot progress
4. Side-plot progress
5. Timeline
6. Planted details already set
7. Information that must not be disclosed early

Engineering implication: write the "project constitution" as the single authoritative context source into the workspace root file, so that any Agent, any model, and any moment of access can obtain the same set of constraints. This is a perfectly isomorphic idea to AGENTS.md in software engineering.

The accompanying three-column creation interface (left: characters / outline / chapters / materials; center: body text editing; right: AI Agent dialogue) solves the "context visibility" problem — the author can see the constitution and the body text at the same time, preventing the AI from improvising while it cannot see the constraints.

5.5 Claude-Code-Novel-Writer v4.1

Note: this is a GitHub community project, not any vendor's official standard. Within the current search scope, it is the only usable commercial-writing AGENTS.md / SKILL.md reference implementation.

Directory structure:

├── AGENTS.md                 # 规范化的项目指引（harness-neutral）
├── CLAUDE.md                 # Claude 侧 import AGENTS.md
├── .agents/
│   ├── roles/                # 规范化专能角色
│   └── skills/               # 规范化写作工作流
├── .claude/agents|skills|output-styles|settings.json
├── .codex/agents|config.toml|hooks.json
├── .pi/prompts/
├── manuscript/chapters/      # 权威源
├── planning/  characters/  worldbuilding/
└── automation/

The distinction between "authoritative source / derived" is the most valuable engineering idea in this project: the manuscript is the authoritative source; state tracking is derived and regenerable. This principle downgrades L4 from "memory that may be corrupted" to "an index that can be rebuilt from the source", greatly reducing the risk of long-form engineering.

6. Governance & Compliance Topic

6.1 Compliance Baseline

The "Measures for the Labeling of AI-Generated Composite Content" takes effect on September 1, 2025. Its Article 6, Paragraph 4 requires platforms to provide a labeling feature and remind users to declare proactively.

On September 23, 2025, Fanqie popped up a "Did you use AI?" checkbox in the author backend, prompting "If the work you upload was synthesized/generated using AI-related tools, you must check it truthfully" — a direct implementation of that provision.

6.2 Fanqie Novels: This Team's Strongest L6 Sample

Capability-side red line: AI continuation and AI expansion capabilities are only available to authors of non-guaranteed-minimum books (disabled for guaranteed-minimum books). This is the clearest "tiered authorization by contract type" design found in this team's research.

Official characterization: "It is ultimately up to each author to decide whether and how to use tool-generated content."

6.3 Diverging Platform Stances

7. Selection Recommendations

7.1 Selection by Creative Goal

7.2 Selection by Harness Capability

7.3 General Engineering Recommendations

1. Build the state layer first, then talk about generation. The first step in starting any long-form project is not writing Chapter 1, but establishing a Bible / project constitution.
2. Distinguish the authoritative source from derived indexes. The manuscript is the authoritative source; state tracking must be regenerable.
3. Separate mechanical signals from literary quality. Use scripts to check anomalies and humans to judge quality; do not treat word count and paragraph length as quality scores.
4. Set "what the Agent must not say". The knowledge graph stores only current-progress information, preventing future information from being disclosed early.
5. Chinese submission projects must reserve a declaration process. Since 2025-09-01, AI labeling is a legal obligation, not an option.

8. Document Navigation & Reading Paths

Recommended paths:

• Engineering-management view: README → 05-claude-fiction.md → 03-yuewen-miaobi.md → 04-fanqie-ai.md
• Content-creation view: README → 02-sudowrite.md → 01-novelai.md → 07-chinese-webnovel-ai.md
• Compliance-governance view: README Section 6 → 04-fanqie-ai.md → 03-yuewen-miaobi.md → the L6 parts of 05-claude-fiction.md

9. Summary

The 8 platform documents in this team jointly support one judgment: competition in the AI fiction track has shifted from "whether the model can write good sentences" to "whether Harness can hold up hundreds of thousands of words".

Three conclusions:

1. L4 determines the long-form ceiling. Bible, State, Timeline, symlink snapshots, Lorebook triggering — the five forms are essentially the same thing: externalizing state out of the model's context into an auditable, rollback-able, regenerable engineering object.
2. The way to solve L1 is to "reduce rather than expand". The industry's best solution is not to force everything into a bigger context window, but to use state externalization, keyword triggers, and dedicated pruning agents to compress what must enter the context to a minimum.
3. L6 is already a hard constraint in the Chinese scenario. After the "Measures for the Labeling of AI-Generated Composite Content" took effect on 2025-09-01, Fanqie's mandatory declaration and its disabling of expansion/continuation for guaranteed-minimum books have been upgraded from platform rules to a compliance baseline. Apart from Fanqie, public compliance-implementation documentation from other platforms is largely absent — this is this team's biggest information gap.

One point worth emphasizing repeatedly: for all the normative content in this team about AGENTS.md / SKILL.md, apart from the GitHub community project Claude-Code-Novel-Writer, no official standard documents from any commercial platform were found. Any normative form given in this document is a community practice or a draft proposal and does not constitute an official standard.

Information-Gap Declaration

During the research and writing process, this team identified the following unresolved items, all marked with [To be filled] or [To be verified], with no speculative filling:

1. "Xiezuo Shi" (写作狮): no results at all after 3 rounds of searches; removed from the platform list and not written into any platform document's main text.
2. Fanqie AI pricing: the official announcement did not disclose standalone pricing; third-party claims of "¥19/month tied to membership" and "completely free" contradict each other and cannot be relied upon, marked [To be filled].
3. Yuewen Miaobi / Writer Assistant AI pricing: no public price list was found; suspected to be provided free with Writer Assistant, marked [To be filled].
4. NovelAI context window: five figures exist — 2,048 / 6,144 / 8,192 / 28,672 / 128k tokens — not unified by the official, presented in parallel and marked [To be verified].
5. MidReal developer background: "jointly developed by MIT, NYU, Cambridge, Princeton" appears only on a single aggregator site, unconfirmed on official and App Store pages, marked [To be verified].
6. MidReal pricing naming: the App Store lists both "Premium/Ultra" and "Low/Medium/High Dose" in-app purchase naming sets concurrently, suspected to be old and new versions coexisting, presented in parallel.
7. Biling AI positioning conflict: "academic writing platform" and "web-novel creation platform" come from different sources, and the official-site entry is inconsistent; explained in 07-chinese-webnovel-ai.md.
8. AIGC labeling compliance implementations across platforms: apart from Fanqie (mandatory declaration), no public documentation was found on whether other platforms write AIGC metadata fields.
9. Commercial-platform AGENTS.md / SKILL.md: no commercial platform has made such files public. Any normative form must be labeled "community practice / draft proposal, not an official standard".
10. Sudowrite developer entity and founding year: all sources only write "Sudowrite", with no registered company name or founding date.
11. Lingxie Creation (SoloEnt) developer entity, founding time, and official-site URL: the source is a CSDN marketing long-form article; no independent official site or business-registration information was found; all prices marked [To be verified].
12. Direct verification of AI Dungeon's official pricing page: Forward Future marks a 2026-08-30 verification, but the original latitude.io official page text was not found.
13. MidReal's L5 evaluation mechanism: no public search results, marked [To be filled].
14. Quantitative-effectiveness data for Chinese web-novel AI platforms: apart from Yuewen (DAU +100%, tokens +90%, weekly usage rate >75%), no official verifiable metrics were disclosed for the remaining platforms; third-party evaluation data is mostly marketing copy and was not cited.

10. References

1. Fanqie Novels official announcement, "AI Writing Tools Feature Launch Notice" — Fanqie Novels, 2024. https://fanqienovel.com/writer/zone/article/7327136545129906238
2. Haike Caijing, "Fanqie Novels' AI Problem" (Sina Finance) — Haike Caijing, 2025. https://finance.sina.com.cn/search/2025-10-09/doc-infthsqh9655363.shtml
3. Sohu Culture, "2024—2025 Biennial List · Preface One: The AI Era, the 'Great View of Literature', and the Tenderness of the World's Dark Side" — Sohu Culture, 2025. https://cul.sohu.com/a/1058758339_121124749
4. Chinanews Jingwei, "Creative Effort Turned into AI Fodder? Web-Novel Authors 'Rise Up', Fanqie Novels Rushes to Clarify" — Chinanews Jingwei, 2025. https://www.jwview.com/jingwei/html/07-26/601732.shtml
5. China Youth Daily, "'Miaobi Tongjian' and 'Comic Drama Assistant' Released: AI Empowers Web-Novel Creation and IP Adaptation" (Tencent) — China Youth Daily, 2025. https://new.qq.com/rain/a/20251017A08J6J00
6. Tencent News, "Comic Drama Annual Growth Over 45%: IP and AI Dual-Drive Ignite a Hundred-Billion Market" — Tencent News, 2026. https://new.qq.com/rain/a/20260319A03C2U00
7. Geek Park, "Can AI Write a Novel Like 'Joy of Life'?" — Geek Park. https://so.html5.qq.com/page/real/search_news?docid=70000021_10168f9fb6122152
8. Baidu Baike, "Writer Assistant Miaobi Edition" — Baidu Baike. https://baike.baidu.com/item/%E4%BD%9C%E5%AE%B6%E5%8A%A9%E6%89%8B%E5%A6%99%E7%AC%94%E7%89%88/63227250
9. Yuewen Group official site — Yuewen Group. https://www.yuewen.com/
10. Hands-on article, "Writing a Novel with Claude Code: I Accidentally Discovered a Squad Power-Up Mode" (reprinted from "Xixiaoyao Tech Talk") — 2026. https://m.aitntnews.com/newDetail.html?newId=20196
11. GitHub — forsonny/Claude-Code-Novel-Writer (Multi-Agent Novel Writer v4.1) — 2026. https://github.com/forsonny/Claude-Code-Novel-Writer
12. HackerNoon, "Claude Book: A Multi-Agent Framework for Writing Novels with Claude Code" — HackerNoon. https://hackernoon.com/claude-book-a-multi-agent-framework-for-writing-novels-with-claude-code
13. Blog, "How to Write a Novel with Agents (Part II)" — tulancn. https://tulancn.github.io/%E5%AD%A6%E4%B9%A0/study/%E5%A6%82%E4%BD%95%E7%94%A8Agent%E5%86%99%E4%B8%80%E6%9C%AC%E5%B0%8F%E8%AF%B4%EF%BC%88%E5%85%B6%E4%BA%8C%EF%BC%89
14. AI Text Tools, "NovelAI: Features, Pricing & Reviews 2026" — AI Text Tools. https://aitexttools.net/tools/novelai
15. RECATOOLS, "NovelAI Review" — RECATOOLS. https://recatools.com/ai-directory/novelai/
16. BuildFastWithAI, "NovelAI Review 2026" — BuildFastWithAI. https://buildfastwithai.com/ai-tools/novelai
17. ToolNest, "NovelAI" — ToolNest, 2026. https://toolnest.ai/project/novelai
18. Sudowrite official blog, "How Sudowrite Works: Story Bible, Muse, and the Tools" — Sudowrite. https://sudowrite.com/blog/how-sudowrite-works
19. AI Wiki, "Sudowrite" — AI Wiki. https://aiwiki.ai/wiki/sudowrite
20. Roleforge, "Best AI Dungeon Master Tools 2026" — Roleforge. https://roleforge.ai/blog/best-ai-game-master-tools-compared
21. Apple App Store, "MidReal - AI Stories" — MidReal Inc. https://apps.apple.com/mr/app/midreal-ai-stories/id6566178525
22. Toolify, "MidReal" — Toolify. https://www.toolify.ai/tw/tool/midreal
23. CSDN, "AI Tools Recommended for Writing Novels in 2026" (Lingxie Creation SoloEnt) — CSDN, 2026. https://tianqi.csdn.net/6a70503110ee7a33f295946a.html
24. AI Express, "2026 AI Writing Tools Roundup" — AI Express. https://www.aiexpress.news/wiki/ai-writing-tools-2026
25. Tahou, "A Detailed Look at Biling AI, the All-Round Creation Platform" — Tahou. https://www.tahou.com/article/211029564132177925
26. Meipian, "How Beginners Start Web-Novel Writing" — Meipian. https://www.meipian.cn/5gn73n9w
27. Toutiao, "Novel-Writing Technique Notes: From Beginner to Advanced" — Toutiao. https://m.toutiao.com/article/7666250379276927531
28. Toutiao, "A Complete Planning Method for Long-Novel Outlines" — Toutiao. https://m.toutiao.com/article/7638890425910968851