Integrated Research on Chinese Web Novel AI Platforms (Lingxie Creation / Wawa Writing / Xingyue Writing / FeelFish / Qimao AI Assistant / Zhongwen Xiaoyao)

1. Introduction

1.1 Research Scope and Topic-Selection Bias

This integrated research covers the niche track of Chinese web novel AI creation tools. One topic-selection bias must be stated up front: the task brief originally suggested covering "Xiezuoshi, Metasharp Writing Cat, and Biling AI", but the actual results of three rounds of searching did not match the suggestioncars, and this was handled as follows:

The six products that genuinely fit the definition of a "Chinese web novel AI platform" are the ones for which materials were actually obtained in the search: Lingxie Creation (SoloEnt), Wawa Writing, Xingyue Writing, FeelFish Flying Fish Creation, Qimao AI Assistant, and Zhongwen Xiaoyao. This report takes these six as its main subject.

One point that must be admitted: the material density on these six is highly uneven. Lingxie Creation has one longer description of features and pricing (source: a long CSDN tech-community article), while Wawa Writing, Xingyue Writing, and FeelFish have only fragmentary descriptions from aggregate review sites. This report therefore adopts a "Lingxie Creation as the primary specimen, the rest as comparisons" approach, and uniformly flags low-credibility sources.

1.2 Platform Overview

Of the six products, only Qimao AI Assistant and Zhongwen Xiaoyao are "the web-novel platform's own AI"; the other four are third-party tools aimed at authors. This distinction will recur throughout the later Harness analysis — the L6 governance of platform-embedded AI is necessarily stronger than that of standalone tools, because it simultaneously bears distribution and moderation responsibility.

1.3 Positioning and Pricing

1.3.1 Lingxie Creation (SoloEnt) Pricing

The prices come from a long CSDN tech-community article, labeled as "yearly early-bird price". No independent official website or company-registration information could be verified, so all are marked [To be verified].

What is notable is its price structure: the difference between the four tiers lies in the "built-in AI credit", not in feature toggles. This is the same business logic as Sudowrite's "identical features across all tiers, differing only in credits", and is the opposite of NovelAI's "feature tiering (Xialong only on Opus)".

1.3.2 Pricing of the Other Platforms

This is the starkest contrast between this report and the overseas-platform report: NovelAI and Sudowrite have public, clearly tiered, cross-verifiable pricing pages, while pricing information for Chinese web-novel AI tools is largely missing. The likely reason is that these tools are mostly sold via "yearly early-bird" and "community group-buy" arrangements, and their pricing pages are not publicly indexed. For a buyer this means no comparable pricing is possible at the selection stage; one must go through an actual hands-on trial process.

1.4 Open Form and Integration Methods

1.5 This Group's Position in the AI Harness System

Of the six tools, only Lingxie Creation's design touches the core of L1 and L4 among the six Harness layers; the other five mostly remain at the "feature collection" level. The engineering value of this group can be summarized in one sentence: Lingxie Creation productized the "project constitution" concept, which is tantamount to bringing software engineering's AGENTS.md into web-novel creation.

This point occupies a special place among this group's 9 documents: the overseas platforms (NovelAI's Lorebook, Sudowrite's Story Bible) and the world-model (Claude Book's Bible / State / Timeline) all did the same thing, but an author writing in Chinese and publishing on a Chinese platform needs a state layer that can handle Chinese web-novel terminology (satisfaction beats, pacing, cheat/golden finger, chapter outline) — this is precisely the reason for this group's tools' existence.

2. Glossary

3. Feature Description

3.1 Lingxie Creation SoloEnt: A Project-Constitution-Driven Workbench

Lingxie Creation is aimed at authors of long novels and web novels, positioned as an AI creation workbench. It has five core capabilities:

3.1.1 Three-Column Creation Interface

The engineering significance of this layout is often underestimated: it solves the context visibility problem. Only when the author can see the settings and the main text at the same time can they judge "what constraints the AI currently sees". In a single-column editor, the settings are hidden state; when the AI cannot see the constraints, it improvises freely, and problems usually surface only dozens of chapters later.

3.1.2 SoloEnt.md Project Constitution

The project constitution centrally records seven categories of information:

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

The seventh item is the easiest to overlook and the one that best reflects engineering maturity. It maps directly to a failure mode unique to Agent writing: if the knowledge graph stores the whole-novel information, the model will "accidentally" reveal the chapter-30 secret by chapter 3. Writing "information that must not be disclosed early" into the constitution is equivalent to adding an "information disclosure boundary" to the L6 governance layer.

3.1.3 AI Editor

Analyzes chapters from an editorial perspective across multiple dimensions, scoring, rating, and giving revision suggestions; the checks cover:

• Opening appeal
• Clarity of the protagonist's goal
• Concentration of satisfaction beats and conflict
• Pacing
• Whether character behavior matches the settings
• Chapter-ending follow-read momentum

The choice of these six items is noteworthy — they are all observable indicators of commercially oriented web-novel writing, not literary indicators. This is a different facet of the same recognition that Claude-Code-Novel-Writer states: "mechanical signals only reveal anomalies, they do not judge literary quality". Lingxie does not pretend to judge literary quality; it anchors evaluation on engineering signals specific to web novels.

3.1.4 Book Deconstruction

Analyzes existing works along six dimensions: writing style, structure, pacing, characters, story boundaries, and cheat/golden finger. Its workflow significance is reverse-engineering a hit into executable constraints — the analysis results ultimately land in the project constitution.

3.1.5 Skill-Pack Marketplace and Multi-Model

The skill-pack marketplace offers four types of reusable assets — Prompt / Workflow / Rules / Skills; it supports multiple models — Claude, GPT, Gemini, GLM, Kimi, DeepSeek — and different Agents can be assigned different models.

"Assigning different models to different Agents" is a rare L3 design among this group's platforms: it means the orchestration layer is not a fixed pipeline, but can choose the most appropriate inference resource for each role — this is isomorphic to Claude Book's approach of using Opus for PLANNER / WRITER, Sonnet for REVIEWER, and Ministral for the Perplexity Gate, except that the model selection is exposed to the user at the UI layer.

3.2 Wawa Writing: A Three-Tier Master-Outline—Chapter-Outline Architecture

Wawa Writing (Waveform Intelligence) focuses on full-chain web-novel creation; the features most frequently mentioned are the "master-outline—chapter-outline three-tier architecture" and "strong long-form character consistency".

In a horizontal comparison in the same source, Wawa Writing and Biling AI are contrasted as two different routes:

The same source also gives "novel creation"-dimension ratings: Metasharp Writing Cat ★★☆☆☆, Biling AI ★★★★☆, Wawa Writing ★★★★★. These ratings come from an aggregate review site, have a marketing advertorial nature, are marked [To be verified], and are not cited as factual conclusions; they are used here only to illustrate the comparison frame of the source.

From a Harness perspective, the "master-outline—chapter-outline" layering is itself an L1 context strategy: first let the model see the master outline to fix the direction, then let the model see the chapter outline to fix that chapter's task, avoiding attention dilution from stuffing in the entire outline at once. This is the same idea as Sudowrite's Chapter Beats and Claude Book's beats.

3.3 Xingyue Writing: A Compliance-and-Risk-Control-Oriented Aid

Xingyue Writing focuses on web-novel duplicate-checking, offending-word risk control, and platform pitfall detection; the source explicitly notes it has "no core creation capability".

The engineering positioning of this kind of tool is clear: it does not take on L1—L4, but instead is an L6 gate before submission. After the Measures for Labeling Artificial-Intelligence-Generated Synthetic Content took effect on 2025-09-01 and Fanqie began mandating AI declarations from 2025-09-23, "self-check before submission" went from optional to a required step in the process. Duplicate-checking and offending-word detection therefore gained their own product space.

Credibility of materials is low (single source); functional details and pricing are both [To be filled].

3.4 FeelFish Flying Fish Creation: An Ultra-Long-Context Orientation

The two features of FeelFish Flying Fish Creation that are mentioned are "ultra-long-text context" and "multi-AI division of labor and collaboration"; its drawback is "heavy client application".

The "ultra-long-text context" selling point must be viewed with caution: this group's core thesis holds that the key to long-form capability lies not in how large the context window is, but in whether state is externalized. Claude Book's measured conclusion is that — using a state/current/ snapshot "when writing chapter 15, the model can precisely retrieve what happened in chapters 1—14, and does not need a 100K-token context". In other words, cramming earlier text in by relying on a huge window is the most costly route.

Credibility of materials is low (single source); no pricing or version information, [To be filled].

3.5 Qimao AI Assistant: The Platform-Embedded Minimal Capability Set

Qimao's "AI Assistant" cooperates with Baidu ERNIE Bot and supports only 3 functions:

1. Writing inspiration
2. Story setting
3. Character naming

At the same time, Qimao required authors to label AI usage earlier than Fanqie did.

Taken together, these two facts form a complete picture: extreme contraction in capability, one step ahead in governance. This stands in contrast to Fanqie — Fanqie provides a more complete toolset (book-opening inspiration, outline generation, writer's-block tips, AI query, AI naming, expansion / rewriting / custom description, continuation), and therefore bears heavier governance responsibility (expansion / continuation disabled for guaranteed books, mandatory declaration).

For the platform side, this is a computable trade-off: the closer AI capability gets to "generating the main text", the higher the moderation and compliance cost. By stopping its capabilities at "inspiration and naming", Qimao keeps L6 risk to a minimum.

3.6 Zhongwen Xiaoyao: The Platform Self-Developed Large-Model Route

Chinese Online released the Zhongwen Xiaoyao large model in October 2024, and opened it to some authors in June 2025. Officially claimed capabilities include:

• One-click generation of ten thousand characters
• Writing a whole novel from a single image
• Comprehending a million-character novel in one read

The third claim — "comprehending a million-character novel in one read" — points to the same capability dimension as Yuewen's Miaobi Tongjian "deep understanding of ten-million-character web novels" and the Manju Assistant's "deep understanding of a million-character novel in as fast as 5 minutes": long-text comprehension, not long-text generation. This group classifies this capability at the highest tier of L1: it is not about "how many characters it can read", but "whether, after reading, it can answer questions about foreshadowing and details".

None of the three claims has any third-party verification or technical-detail disclosure, so they are marked [To be verified].

3.7 Analysis of the Three Excluded Tools

3.7.1 Metasharp Writing Cat

The two sets of pricing data conflict, so this is marked [To be verified]. Reason for exclusion: its core scenarios are practical genres such as papers, reports, and official documents, which do not involve web novels' long-range state-management problem; the "four-step writing method" targets "producing one complete long text in a single pass", whereas web novels are "serialized over hundreds of chapters that must remain consistent" — these are different engineering problems.

3.7.2 Biling AI

According to another source, its novel-creation functions include: 200+ novel AI generators, one-click novel-outline generation, AI writing of the full work, viral-novel deconstruction, character setting and plot-structure design, in-site saving, and one-click continuation; the outline generation will "directly tell the user what emotional pain points the first chapter should express, what the editor will focus on, and what the review highlights are".

Reason for exclusion: its positioning conflicts across multiple sources, making it impossible to confirm the true boundary of its web-novel capabilities; and it is rated "no long-term memory over long works; extremely prone to continuity errors after 200,000 characters" — which is precisely the L4 capability gap this group's core thesis points to.

3.7.3 Xiezuoshi

Three rounds of searching returned absolutely nothing; no citable information was obtained. This report makes no description of it at all, only recording it in the information-gap statement after section 7.

4. Platform Architecture

图 4-1｜灵蟹创作三栏工作台架构：项目宪法为共同输入，每章回写形成闭环

数据来源：基于本文分析绘制的示意图。

4.1 Two Architecture Forms: Standalone Workbench vs Platform-Embedded

This divide explains why Chinese web-novel AI tools present a pattern of "the capable ones ignore compliance, the compliant ones are weak in capability": standalone tools have no moderation duty, so they are free to focus on generation; platform-embedded AI has a moderation duty, so it necessarily contracts its capabilities. Fanqie is the only platform that did both — it has both a complete toolset and mandatory declaration plus guaranteed-book restrictions (see 04-fanqie-ai.md).

4.2 Lingxie Creation's Three-Column Workbench Architecture

┌─────────────┬─────────────────┬──────────────────┐
│ 左栏        │ 中栏            │ 右栏             │
│ 人物        │                 │                  │
│ 大纲        │   正文编辑      │   AI Agent 对话  │
│ 章节        │                 │                  │
│ 资料        │                 │                  │
└─────────────┴─────────────────┴──────────────────┘
       │                │                  │
       └────────────────┴──────────────────┘
                        │
                        v
              SoloEnt.md 项目宪法
        （世界观规则 / 人设边界 / 主线支线进度 /
          时间线 / 已埋伏笔 / 文风视角 / 不能公开的信息）
                        │
        ┌───────────────┼───────────────┐
        v               v               v
    AI 责编         AI 拆书        技能包市场
  （评分评级）    （维度分析）   （Prompt/Workflow/
                                  Rules/Skills）

Key point of the architecture: the project constitution sits below the three columns and above the Agents, and is the common input to all AI capabilities. This corresponds fully in responsibility to Claude Book's bible/ directory (which never changes during generation); the difference is that Lingxie made it a single visible, editable root file, whereas Claude Book made it a directory.

Each has engineering trade-offs: a single file is easy to read and edit but grows with length; a directory structure is extensible but requires additional assembly logic. For Chinese web-novel authors (mostly non-engineers), the maintainability of a single file is significantly higher.

4.3 Model Integration Layer: Multi-Model and BYOK

BYOK's Harness significance has two layers:

1. Cost guardrail: with the author using their own key, cost is predictable)Skip, and the platform does not bear the risk of inference-cost fluctuation.
2. Data boundary: integrating Ollama / LM Studio local models means the main text never leaves the machine. This is two implementations of the same need as NovelAI's XSalsa20 client encryption — NovelAI uses encryption, Lingxie uses local models.

"Different Agents can be assigned different models" pushes the model-choice authority down to the orchestration layer, a coupling design between L2 and L3.

4.4 Context-Assembly Data Flow

Taking one "write a new chapter" operation in Lingxie Creation as an example:

[作者选中章节，发出写作指令]
        |
        v
[读取 SoloEnt.md 项目宪法] --> [世界观规则 + 人设边界 + 已埋伏笔]
        |
        +--> [读取左栏：本章细纲 / 相关人物卡 / 资料]
        |
        +--> [读取中栏：上一章正文尾部（近邻上下文）]
        |
        +--> [剔除：宪法中标记为"不能提前公开"的信息]
        |
        v
[按角色分配模型] --> [生成章节草稿]
        |
        v
[AI 责编评分评级] --> [作者采纳 / 打回]
        |
        v
[回写宪法：主线支线进度 / 时间线 / 新埋伏笔 / 已回收伏笔]

The final step, "writing back to the constitution", is the most critical link in the entire chain and the one most easily omitted in a self-built workflow. Without the write-back, the state layer is read-only, and long-form consistency cannot be maintained.

5. Harness Design

5.1 L1 Context Engineering Layer

Against the four assembly strategies this group has identified (all-at-once stuffing, conditional triggering, state externalization, pruning agent), Chinese web-novel AI platforms generally remain at the "all-at-once stuffing" or "layered stuffing" stage; only Lingxie Creation achieves partial "state externalization" through its project constitution.

One clear gap: no Chinese web-novel AI tool implements a keyword-trigger mechanism like NovelAI's Lorebook. As a result, in the scenario of "many setting entries but only three used in this chapter", Chinese tools either stuff everything in (wasting context) or rely on the author to pick manually (depending on human memory).

5.2 L2 Tools and Execution Layer

This group's L2 has one striking feature: the "scripting of deterministic tasks" ring is missing. Claude / Codex workflows wrap naming, word-count statistics, and knowledge-graph queries as scripts (see 05-claude-fiction.md), because these are things "the model is not good at and is not reliable for". Chinese web-novel AI tools still leave these tasks to the model, with the result that:

• Naming is prone to repetition (especially in web novels, where duplicate names are a hard defect);
• Word-count statistics are unreliable (web novels are billed by character count, so errors directly affect revenue);
• Foreshadowing ledgers rely on the model "remembering" rather than "storing".

This is the most substantive gap between Chinese tools and self-built, engineered workflows.

5.3 L3 Orchestration and Control Layer

The six-step pipeline recommended by Lingxie Creation (source: CSDN):

Two commendable aspects of the pipeline design:

1. Step 3 precedes step 4: building the state layer before starting generation is consistent with this group's general engineering recommendation of "build the state layer first, then talk about generation".
2. Step 6 brings submission into the pipeline: making L6 compliance the finishing node of the pipeline rather than an after-the-fact remedy fits the compliance reality after 2025-09-01.

The other design point in the orchestration layer is assigning different models to multiple Agents, as described above. Note that: if multi-Agent collaboration lacks a responsibility boundary of the "verify only, do not write" kind, a review Agent overstepping to rewrite can cause state inconsistency. Whether Lingxie defines such boundaries is not covered by any public documentation found, so this is [To be filled].

5.4 L4 Memory and State Layer

This is the focus of this report, and the part with the highest value among Chinese web-novel AI tools.

5.4.1 Lingxie Creation's Six Categories of State

Against the complete state layer this group has identified (six categories of information), Lingxie Creation covers all six categories, and additionally adds "writing style and narrative perspective" and "information disclosure boundary". Among this group's 9 documents, this is the most complete state design apart from Claude Book's four-directory Bible / State / Story / Timeline.

The only thing missing is "characters' immediate state" — i.e., the "character location, possessions, knowledge, relationships" managed by Claude Book's state/ directory. This item is precisely the highest-frequency-changing, most-error-prone information in a long serialization.

5.4.2 The Missing State Versioning and Snapshots

Claude Book's key design is a state/current/ symlink plus chapter-NN/ archival, making state auditable, rollback-able, and capable of running multiple progress tracks in parallel. No equivalent mechanism was found in any Chinese web-novel AI tool [To be filled].

The practical consequences of this gap:

The spontaneous countermeasure of Chinese practitioners is to build a custom snapshot skill: saving the current knowledge graph, main text, and outline, with support for restoring snapshots and toggling among multiple snapshots (source: the tulancn blog). This shows the need is real, just not yet productized.

5.4.3 Comparison with the Three Paradigms

Conclusion: on L4, Chinese web-novel AI platforms generally sit at the stage of "have a constitution, no snapshot; have all-at-once stuffing, no triggering". Lingxie Creation takes the manual version of paradigm A to a productizable degree, but paradigms B and C remain blank.

5.5 L5 Evaluation and Observation Layer

Lingxie's AI Editor is the only systematic L5 implementation among this group's Chinese tools. Two aspects of its design are worth affirming:

1. Indicator selection fits web-novel commercial logic: it does not judge "literariness", but "whether the reader will keep reading" — this is observable and attributable.
2. It clearly defines capability boundaries: the official statement is that "the AI Editor is for reference only, does not equal the platform's real editorial opinion, and does not guarantee acceptance". This statement belongs to the same class of mature boundary declarations as Claude-Code-Novel-Writer's "mechanical signals do not judge literary quality".

An industry-wide blank: apart from Yuewen (daily active users, token consumption, weekly usage rate observable), no Chinese web-novel AI tool has published quantitative effectiveness data. Third-party evaluation data is mostly advertorial, and this report does not cite it.

5.6 L6 Governance and Security Layer

The compliance background that must be made explicit: the Measures for Labeling Artificial-Intelligence-Generated Synthetic Content took effect on September 1, 2025, and its Article 6, Paragraph 4 requires platforms to provide a labeling function and remind users to declare proactively. Fanqie's mandatory "whether AI was used" declaration in the author backend from 2025-09-23 is the direct implementation of that clause (see 04-fanqie-ai.md).

For authors using standalone Chinese tools (Lingxie, Wawa, FeelFish), this means: the tool itself provides no labeling capability, and the legal obligation of AI labeling is borne entirely by the author at the submission stage. This is a risk item that cannot be ignored during tool selection.

One more industry fact should be recorded here: Zongheng Chinese Web and Qimao required authors to label AI usage earlier than Fanqie did; Jinjiang Literature City's February 2025 trial-run notice permits only three scenarios: text proofreading, creative-element assistance, and creative rough-outline assistance.

5.7 Summary of the Six-Layer Capabilities

Response to the core thesis: Lingxie Creation's value lies in productizing the "project constitution", giving Chinese web-novel authors, for the first time, a state layer that can be maintained without engineering ability. Its shortcomings (no snapshot, no triggering, no characters' immediate state) show that the L4 build-out of Chinese web-novel AI tools is still mid-course — a big step forward from "storing nothing", but not yet at the completeness of Claude Book's four directories.

5.8 Special Topic on Long-Range State Management: The Particular Pressures of the Chinese Web-Novel Scenario

Chinese web novels place pressure on L4 via four amplifiers that overseas short- / medium-form scenarios lack:

Therefore, the minimum requirement that Chinese web-novel AI tools face on L4 is: complete in both the setting dimension and the temporal dimension, and it must be able to evolve automatically (or semi-automatically) with each chapter.

Against this requirement, how this group's six tools measure up:

Engineering advice for Chinese long-form authors (based on this group's cross-platform synthesis, not any vendor's official solution):

1. Use the project constitution as the authoritative source, but add a per-chapter-evolving characters' immediate state to it — who is where, what they know, what they hold. This is the layer that is missing from Lingxie's constitution but present in Claude Book's state/.
2. Create chapter archive copies (chapter-NN/) for the constitution, so state can be rolled back.
3. Put "information that must not be disclosed early" in its own section, and explicitly screen it out before each generation.
4. Before submission you must run through offending-word risk control + AI declaration, because the tool itself provides no labeling capability.

6. Practical Cases

Note: Cases 1 and 2 in section 6 are application-scenario inferences based on combinations of already-published features, used to illustrate capability boundaries rather than vendor-disclosed cases; Case 3 is verifiable public fact; section 6.4 covers known issues that recur in public evaluations. Among Chinese web-novel AI tools, apart from the scope of Qimao AI Assistant's capabilities, which is cross-corroborated by financial media, no official cases with quantitative indicators were found.

6.1 Case 1: A Million-Character Serialization Driven by a Project Constitution

Background: a planned 2-million-character xianxia (fantasy cultivation) long work, containing three major factions, a four-tier power system, over 30 named characters, one main plot + five side plots. The core risk is forgetting, by chapter 300, an artifact planted back in chapter 12.

Solution (an inference based on Lingxie Creation's already-published features):

1. Deconstruct: use AI Book Deconstruction to analyze 3—5 hits of the same type, extracting writing style, structure, pacing, and cheat/golden-finger settings to form writing constraints;
2. Build the skeleton: use DeepSeek to generate the worldview and master outline, and write them into the project constitution after human review;
3. Set up the constitution: in SoloEnt.md, record in sections the world-building rules, character personality and behavioral boundaries, main/side plot progress (initially 0), timeline (initially empty), planted foreshadowing (initially empty), writing style and narrative perspective, and information that cannot be disclosed early;
4. Produce chapter by chapter: complete a chapter → AI Editor score → author revision → write back to the constitution (update progress, timeline, add new foreshadowing, mark recovered foreshadowing);
5. Stage review: every 50 chapters, use AI Book Deconstruction to run a structural-and-pacing health check on what has been written.

Effect and boundaries: within this group's Chinese tools, this workflow is already the best reachable solution; in particular, the "information that cannot be disclosed early" section directly prevents the Agent from prematurely leaking later-stage settings — a design that overseas general-purpose tools (Sudowrite, NovelAI) all lack.

The boundaries lie in two places: first, the characters' immediate state is not in the constitution, so details like "does the protagonist know X at this point" still require manual checking; second, there is no versioned snapshot, so if at chapter 300 you find chapter 250 went in the wrong direction, the rollback cost is extremely high.

6.2 Case 2: AI Editor Review and Pre-Submission Self-Check

Background: the author has finished a chapter and is about to publish, needing to judge before submission whether "this chapter is gripping enough".

Solution (an inference based on already-published features):

1. Use the AI Editor to score the chapter on six dimensions: opening appeal, protagonist-goal clarity, satisfaction-beat and conflict concentration, pacing, whether character behavior matches the settings, and chapter-ending follow-read momentum;
2. Rewrite specifically toward the lowest-scoring dimension, e.g. if "chapter-ending follow-read momentum" is low, add suspense at the end;
3. Use an Xingyue-Writing-type tool for duplicate-checking and offending-word risk control;
4. Check "whether AI was used" on the submission platform to complete the declaration.

Effect and boundaries: the AI Editor turns "is this chapter good?" from a subjective feeling into six locatable dimensions, making both what to change and why explicit. But the official statement must be remembered: the AI Editor's opinion does not equal the platform's editorial opinion and does not guarantee acceptance.

Step 4 cannot be omitted: since 2025-09-01, AI labeling has been a legal obligation; Fanqie has mandated declaration since 2025-09-23, and failure to declare truthfully may block moderation. And the tool chain itself provides no labeling capability whatsoever.

6.3 Case 3: The Capability Contraction of Platform-Embedded AI (Qimao)

Background (verifiable public fact): Qimao's "AI Assistant" cooperates with Baidu ERNIE Bot and supports only 3 functions: writing inspiration, story setting, and character naming; at the same time, Qimao required authors to label AI usage earlier than Fanqie did.

Analysis: this is a clear L6-driven product decision. Restricting AI capability to "not generating the main text" eliminates the governance problem of "AI generating low-quality content" at the source — inspiration, settings, and naming are not part of the work's main text, need no labeling, and do not affect contract review.

Contrast with Fanqie: Fanqie chose the opposite route — providing a more complete toolset (including expansion, continuation, and rewriting) — and therefore had to build matching governance capabilities: expansion / continuation disabled for guaranteed books, mandatory AI declaration, and machine determination of low-quality content. Fanqie's MAU exceeds 240 million (QuestMobile, through end of June 2025); with such volume, it cannot be avoided.

Conclusion: the capability boundary of platform-embedded AI is not a technical problem, but a balancing problem between governance cost and commercial return. When selecting, authors should understand: the more capable a platform-embedded AI is, the stricter its moderation tends to be.

6.4 Counterexample: Long Works with No Long-Term Memory, Continuity Errors After 200,000 Characters

A problem that recurs in public evaluations: general-purpose Chinese AI writing tools (the evaluation named Biling AI, but the conclusion is general for similar tools) are rated "no long-term memory over long works; extremely prone to continuity errors after 200,000 characters"; they are also rated as having "a relatively noticeable homogenization tendency".

Diagnosis (an analysis based on this group's six-layer framework):

The value of this counterexample is that it gives the capability watershed for Chinese web-novel AI tools: 200,000 characters. Below this line, how many features there are determines the experience; above it, only the state-layer design determines whether the work can hold together. This is fully consistent with this group's core thesis.

7. Summary

7.1 Strengths

1. The project constitution is the most valuable original design among Chinese web-novel AI tools. By concentrating world-building rules, character-boundary settings, main/side plot progress, timeline, planted foreshadowing, writing-style perspective, and information that cannot be disclosed early into a single root file, it lets even non-engineer authors maintain a state layer.
2. Evaluation indicators fit web-novel commercial logic. The six items the AI Editor checks (opening appeal, protagonist-goal clarity, satisfaction-beat and conflict concentration, pacing, character-behavior consistency, chapter-ending follow-read momentum) are all observable and attributable, bypassing the pseudo-question of "AI judging literariness".
3. "Information that cannot be disclosed early" is a distinctive L6 design, confronting head-on the biggest trap in Agent writing — premature disclosure of future information.
4. BYOK and local-model integration let authors control their own cost and data boundary; Ollama / LM Studio integration means the main text can remain entirely off the machine.
5. Assigning different models to multiple Agents pushes model choice down to the orchestration layer, a rare L3 design among Chinese tools.
6. Compliance risk-control tools gained their own product space (duplicate-checking, offending-word risk control, platform pitfall detection), answering the real process needs after 2025-09-01.

7.2 Limitations and Known Shortcomings

1. Credibility of materials is generally low: Lingxie Creation's core materials come from a CSDN advertorial, and Wawa Writing, Xingyue Writing, and FeelFish have only fragments from aggregate review sites — all marked [To be verified].
2. Developer entity, founding date, and official-URL are largely missing: Lingxie Creation's business-registration information was not found, marked [To be filled].
3. Pricing information is extremely opaque: five of the six have no public price list, making comparable pricing impossible.
4. L4 lacks two key components: no characters' immediate state (location / possessions / knowledge / relationships), and no versioned snapshots and rollback.
5. No entry-level triggering mechanism: there is no equivalent of NovelAI's Lorebook among Chinese tools.
6. Deterministic tasks are not scripted: naming, word-count statistics, and foreshadowing ledgers still depend on the model, with insufficient reliability.
7. L6 governance is largely blank: apart from the platform-embedded type, there is no AIGC labeling, no training-data authorization statement, and no copyright-ownership statement.
8. No public quantitative-effectiveness data: apart from Yuewen, no Chinese web-novel AI tool discloses verifiable effectiveness indicators.
9. No commercial platform publishes AGENTS.md / SKILL.md: this group found no official Harness config file from any Chinese web-novel AI platform.

7.3 Applicability Boundaries

7.4 Selection Recommendations

• Long Chinese web-novel serialization: first choice is Lingxie Creation SoloEnt (project constitution + AI Editor); the backup is the self-built workflow described in 05-claude-fiction.md (if the team has engineering capability). If already within the Yuewen system, directly use the Writer Assistant's Miaobi Tongjian (see 03-yuewen-miaobi.md).
• Submitting to Fanqie: must use Fanqie's official toolset to go through the declaration process (see 04-fanqie-ai.md); AI labeling must be completed by yourself for third-party-tool output.
• Only need inspiration and naming: the scope of Qimao AI Assistant's capabilities is sufficient, and it carries the lowest governance risk.
• Compliance prioritized over output: prefer platform-embedded AI (Fanqie, Qimao), because the labeling and moderation responsibility sits on the platform side; standalone tools leave that responsibility to the author.
• Ultra-long works above 1,000,000 characters: no single Chinese tool can support this; it is recommended to use a project constitution as the authoritative source and supplement characters' immediate state and chapter snapshots yourself (referring to the Bible / State / Timeline four-directory design in 05-claude-fiction.md).
• Academic / workplace writing: Metasharp Writing Cat and Biling AI are more suitable, but they are not web-novel tools and should not be mixed in.

Finally, it must be emphasized: for all the AGENTS.md / SKILL.md form descriptions involved in this report, apart from the GitHub community project Claude-Code-Novel-Writer, no official Harness config file from any Chinese web-novel AI platform was found. The engineering recommendations in this report about state layers and pipelines are all syntheses and proposed drafts based on public features, and do not constitute any official standard.

Information-Gap Statement

1. "Xiezuoshi": three rounds of searching returned absolutely nothing; this report makes no description of it at all.
2. Lingxie Creation (SoloEnt) developer entity, founding date, and official URL: the source is a long CSDN tech-community article; no independent official website or business-registration information was found; all prices and feature descriptions in this report are marked [To be verified].
3. Wawa Writing, Xingyue Writing, and FeelFish developers, versions, and feature details: only fragmentary descriptions from aggregate review sites, with low credibility, marked [To be verified].
4. Public pricing for the six platforms: apart from Lingxie Creation (source questionable), five are all [To be filled]; no public price lists were found.
5. Whether each platform conforms to the Measures for Labeling AI-Generated Synthetic Content: apart from Fanqie (mandatory AI declaration) and Qimao and Zongheng (labeling required earlier than Fanqie), no AIGC-labeling implementation description from any standalone Chinese tool was found.
6. Quantitative-effectiveness data from Chinese web-novel AI platforms: apart from Yuewen (DAU growth more than doubled, average daily token consumption grew over 90%, author weekly usage rate >75%), no other platform discloses verifiable indicators; third-party evaluation data is mostly advertorial, and this report does not cite it.
7. Commercial-platform AGENTS.md / SKILL.md: no Chinese web-novel AI platform publishes such files. Any spec-form descriptions in this report are community practices or proposed drafts, not official standards.
8. Whether Lingxie Creation defines Agent responsibility boundaries (e.g. "verify only, do not write"): no public documentation found, [To be filled].
9. Zhongwen Xiaoyao technical details and opening scope: only official claims ("one-click generation of ten thousand characters", "write a novel from a single image", "comprehend a million-character novel in one read"), with no third-party verification or technical disclosure, marked [To be verified].
10. Metasharp Writing Cat pricing conflict: one source says "free / 24 yuan per month / 48 yuan per month (yearly 450 yuan per year)", another says "free tier 100,000 characters per month / advanced tier 29 yuan per month for 500,000 characters / pro tier 59 yuan per month for 1,000,000 characters"; the two sets of data contradict each other and are presented side by side.
11. Biling AI positioning conflict: the two claims of "academic-writing platform" and "web-novel creation platform" come from different sources, and the official-site entry points are inconsistent; this report treats it as "positioning conflict, excluded from its own section".
12. Official customer cases of Chinese web-novel AI tools: no official case with quantitative indicators was found; Cases 1 and 2 in section 6 are scenario inferences, as explicitly marked.

8. References

1. CSDN《2026 年写小说好用的 AI 工具推荐》（灵蟹创作 SoloEnt） — CSDN，2026。https://tianqi.csdn.net/6a70503110ee7a33f295946a.html
2. 塔猴《笔灵 AI 写作全能创作平台详解》 — 塔猴。https://www.tahou.com/article/211029564132177925
3. AI Express《2026 年 AI 写作工具横评》 — AI Express。https://www.aiexpress.news/wiki/ai-writing-tools-2026
4. AI 产品库《10 款中文 AI 论文写作润色降重工具实测对比》 — AI 产品库。https://aiproducthub.cn/?p=6514/
5. 夜雨聆风《2026 年主流 AI 写作与文档处理工具梳理》 — 夜雨聆风。https://www.yeyulingfeng.com/634408.html
6. hqwc.cn《2026 年度 5 款番茄小说过稿 AI 工具全维度测评》 — hqwc.cn。http://www.hqwc.cn/news/939376.html
7. 中新经纬《创作心血变 AI 养料？网文作者"揭竿而起"，番茄小说忙澄清》 — 中新经纬，2025。https://www.jwview.com/jingwei/html/07-26/601732.shtml
8. 海克财经《番茄小说的 AI 难题》（新浪财经） — 海克财经，2025。https://finance.sina.com.cn/search/2025-10-09/doc-infthsqh9655363.shtml
9. 番茄小说官方公告《AI 写作工具功能上线通知》 — 番茄小说，2024。https://fanqienovel.com/writer/zone/article/7327136545129906238
10. 极客公园《AI 能不能写出〈庆余年〉？》 — 极客公园。https://so.html5.qq.com/page/real/search_news?docid=70000021_10168f9fb6122152
11. 中国青年报《"妙笔通鉴""漫剧助手"发布，AI 赋能网文创作和 IP 改编》（腾讯网） — 中国青年报，2025。https://new.qq.com/rain/a/20251017A08J6J00
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. GitHub — forsonny/Claude-Code-Novel-Writer（Multi-Agent Novel Writer v4.1） — 2026。https://github.com/forsonny/Claude-Code-Novel-Writer
14. 博客《如何用 Agent 写一本小说（其二）》 — 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
15. BuildFastWithAI《NovelAI Review 2026》 — BuildFastWithAI。https://buildfastwithai.com/ai-tools/novelai
16. Sudowrite 官方博客《How Sudowrite Works: Story Bible, Muse, and the Tools》 — Sudowrite。https://sudowrite.com/blog/how-sudowrite-works
17. 美篇《新手如何开始网文写作》 — 美篇。https://www.meipian.cn/5gn73n9w
18. 今日头条《小说写作技巧笔记：从入门到进阶》 — 今日头条。https://m.toutiao.com/article/7666250379276927531
19. 今日头条《长篇小说大纲完整规划方法》 — 今日头条。https://m.toutiao.com/article/7638890425910968851
20. ProcessOn《〈别说你懂网文〉读书笔记》 — ProcessOn。https://www.processon.com/view/61064ee81e0853746619fa4a