Yuewen Miaobi / Writer Assistant AI Platform Research

1. Introduction

1.1 Release History

Two time points are worth emphasizing:

1. 2023-07-19 is a watershed for the Chinese web-novel industry — after this, "platform self-developed large models" became the standard for top web-novel platforms.
2. The 2025-02-05 independent deployment of DeepSeek-R1 marks a turn in the technical route: from "self-developed large models" to "self-developed + independently deployed third-party strong-reasoning models". Independent deployment means data does not leave the domain — a key decision for L6 governance.

1.2 Product Matrix and Positioning

Yuewen's official position (source: GeekPark): "AI handles the manual labor in the creative process, not the intellectual work that concerns the soul."

This position directly determines the product shape: Miaobi Tongjian does not generate chapters — it only does understanding, retrieval, and reminders. Among the 8 platforms in this group, this is the most restrained product positioning and the one most aligned with this group's core thesis — Yuewen chose to invest all its resources in L1 (ten-million-character understanding) and L4 (foreshadowing and detail retrieval), rather than in generation.

1.3 Pricing

The absence of pricing information is itself a signal: Yuewen treats AI capability as a basic feature of Writer Assistant rather than a standalone item for sale. This is completely different from the "subscription-based tool" business logic of NovelAI / Sudowrite — Yuewen's AI is an investment to improve platform author retention and productivity, not a monetization product.

1.4 Openness

2. Glossary

3. Feature Description

3.1 Miaobi Tongjian

Core positioning: it does not generate chapters; it acts as the author's "second brain".

Two major values (source: GeekPark):

Capability specifications:

• Has deep understanding of ten-million-character web novels, officially opened to web-novel writers;
• Helps writers quickly build worldbuilding and generate plot frameworks;
• Over 2000 writers participated in co-creation before release.

Why the "second brain" positioning is right: this group's core thesis points out that the core challenge of long-running serialized novels is "not collapsing characters or forgetting foreshadowing even after hundreds of thousands of characters". Yuewen did not try to have AI write for the author, but instead had AI remember for the author — solving an L4 problem the L4 way rather than masking it with generative capability. This is the clearest product judgment in this group.

Evaluation by Platinum author Liudan PaShui (author of "Shao Song"):

"Simple Q&A can check out characters missing their follow-up results, and with targeted completion I can raise the overall completeness of the story by a level."

This evaluation precisely describes how the "second brain" works: AI handles discovery, humans handle repair. This is the correct division of labor between L5 (evaluation and discovery) and human decision-making.

3.2 Copyright Assistant

The significance of Copyright Assistant: among the 8 platforms in this group, it is the only AI capability aimed at the "post-completion stage of a work". All other platforms focus on the creation process; Yuewen covers the entire IP lifecycle.

3.3 Manga Drama Assistant

Manga Drama Assistant is the only AI application in this group with complete quantitative evidence of effectiveness. Three figures (90 days → 10–13 days, 100,000–300,000 RMB, 100+ studios) constitute a credible industry-level efficiency-improvement sample.

3.4 Writer Assistant Basic AI Capabilities

Rolled out progressively since 2023:

The three areas upgraded after integrating DeepSeek-R1 on 2025-02-05:

1. Smart Q&A
2. Getting inspiration
3. Description polishing

3.5 Writer Assistant Claw / WriteClaw

The "integrated QQ conversation capability" item is worth noting: it extends the creation tool from "within the editor" to "within IM", meaning authors can call AI in everyday communication contexts. This is an entry-point migration at the L2 tool layer — a different solution to the same goal as Sudowrite's plugin ecosystem and Claude's coding tools.

4. Platform Architecture

图 4-1｜阅文妙笔平台架构：三助手分工 × 双模型路由 × 正版 IP 数据底座

数据来源：基于本文分析绘制的示意图。

4.1 Three-Assistant Division of Labor Architecture

                [阅文海量正版 IP 数据库]
                          |
        +-----------------+-----------------+
        |                                   |
  [作家助手 / 妙笔通鉴]              [版权助手]
   创作源头：                        价值挖掘：
   - 世界观搭建                      - 精准选书
   - 情节框架生成                    - 内容梳理
   - 伏笔挖掘                        - 匹配改编需求
   - 细节检索                              |
        |                                   |
        |                                   v
        |                            [下游 IP 改编方]
        v
  [漫剧助手]
   形态衍生：
   - 内容理解 - 创作建议
   - 视觉风格 - 素材制作
        |
        v
   [漫剧工作室]

The three assistants cover the entire IP lifecycle: creation (Writer Assistant) → value mining (Copyright Assistant) → form derivation (Manga Drama Assistant). This is the only architecture among the 8 platforms in this group that spans the full "creation—distribution—adaptation" chain.

4.2 Model Layer

Independent deployment is the platform's most important technical decision. It means:

• The author's manuscripts do not leave the Yuewen domain, avoiding the root cause of disputes like Fanqie's 2024 "AI training supplementary agreement";
• It can be deeply customized for web-novel corpora;
• The cost structure is fixed and does not grow linearly with API call volume.

From a Harness perspective, the dual-model strategy of "self-developed + third-party independent deployment" corresponds to model routing at the L2 tool layer: vertical tasks go to the self-developed Miaobi, and reasoning-intensive tasks go to DeepSeek-R1.

4.3 Data Asset Layer

The licensed IP database is the foundation of Yuewen's L6 layer — because data ownership is clear, it can legitimately be used for model understanding and adaptation recommendations, without facing the "training AI on authors' works" dispute that Fanqie encountered in 2024.

5. Harness Design

5.1 L1 Context Engineering Layer

Yuewen's L1 capability specifications are the strongest in this group:

"Ten-million-character deep understanding" and "reading a million characters in 5 minutes" are two metrics of different natures: the former is capacity (how much it can hold), the latter is throughput (how quickly it reads). Together, they show that Yuewen has solved both capacity and efficiency at the L1 layer.

In comparison with other platforms:

It should be noted: the measurements are not fully comparable (characters vs tokens; 1 Chinese character ≈ 1.5–2 tokens), but the order-of-magnitude difference is real. Yuewen is the only platform that directly addresses the long-form problem on the "capacity" dimension; the Claude workflow tackles it from the other side — using externalized state so that capacity is no longer the bottleneck.

5.2 L2 Tools and Execution Layer

A feature of the tool layer: it does not build "generate entire chapters" type tools. Yuewen deliberately avoids the ethical and compliance risks of AI ghostwriting, placing all tools on "understanding, retrieval, suggestion, and visualization".

This restraint pays off directly at the L6 level: Yuewen has never seen the kind of collective author boycott event like Fanqie's 2024 "AI training supplementary agreement".

5.3 L3 Orchestration and Control Layer

Three-assistant division pipeline:

[创作源头] 作家助手 / 妙笔通鉴
      |
      v
[价值挖掘] 版权助手
      |
      v
[形态衍生] 漫剧助手

This pipeline covers the entire IP lifecycle and is the only orchestration design in this group spanning creation—distribution—adaptation.

Orchestration within creation (inferred from publicly disclosed features):

Key observation: Yuewen places "foreshadowing mining" in the revision period rather than the writing period. This is a pragmatic choice — during the writing period the model does not know how the story will develop and cannot judge which foreshadowing is unrecovered; only in the revision period, with the whole book complete, does global retrieval become possible.

5.4 L4 Memory and State Layer

Foreshadowing mining and detail retrieval are Yuewen's unique contribution at the L4 layer. The L4 implementations of the other platforms in this group all manage "settings" (Story Bible, Lorebook, Project Constitution); only Yuewen manages "consistency issues within already-written content" — a post-hoc state check that complements proactive state maintenance.

The limitation is equally clear: this is a checking-type capability, not a tracking-type one. It can tell you, when you ask, "someone introduced in chapter 37 never appeared again later", but it will not proactively remind you while you are writing chapter 100. To become proactive reminders, a persistent foreshadowing ledger (like Claude Book's Timeline + State) is needed.

5.5 L5 Evaluation and Observation Layer

Yuewen is the only platform in this group that discloses verifiable business metrics:

The five metrics form a complete L5 observation matrix:

"Weekly usage rate >75%" is the single most persuasive metric in this group — it shows AI has moved from "trying it out" to "daily reliance".

The other side of the evaluation layer is content quality assessment: Yuewen relies on "final human editorial review" (because Qidian Chinese Net has a high signing threshold and manual editorial review, works typically only go live for paid access after 200,000 characters, so it has been almost unaffected by AI-generated text). This contrasts with Fanqie's "machine judging low-quality content" — manual review vs machine judgment are the two L5 routes on Chinese web-novel platforms.

5.6 L6 Governance and Security Layer

"Licensed IP database + independent deployment + written authorization" constitutes Yuewen's threefold safeguard at L6. Compared with Fanqie, Yuewen's governance is proactive (data ownership was clear from the start), while Fanqie's is reactive (opening first, then patching the rules). This explains why Yuewen has not seen a collective author boycott.

AIGC labeling is a clear gap for this platform: the _Measures for the Labeling of AI-Generated Synthetic Content_ took effect on 2025-09-01, but no specific implementation details from Yuewen on AI labeling were found. This contrasts with Fanqie's mandatory reporting.

5.7 Six-Layer Capability Summary

A response to the core thesis: Yuewen uses "ten-million-character deep understanding" to directly counter long-context forgetting at the L1 layer, and "foreshadowing mining + detail retrieval" to directly counter "continuity errors" at the L4 layer. It is the only commercial platform in this group that treats long-form consistency as a core product proposition (rather than an add-on feature). Its "second brain" positioning — not generating, only remembering and retrieving — is the most precise commercial response to this group's core thesis.

6. Real Cases

6.1 Case 1: Miaobi Tongjian's Foreshadowing Mining

Background: In long-running web-novel serials, a character disappearing after being introduced is the most frequent form of "continuity error". Authors often remember the main-line characters but forget some supporting character introduced in chapter 37.

Approach: Platinum author Liudan PaShui (author of "Shao Song") commented after using Miaobi Tongjian:

"Simple Q&A can check out characters missing their follow-up results, and with targeted completion I can raise the overall completeness of the story by a level."

Effect:

• Discovery: simple Q&A can list "characters who have appeared but lack follow-up plot";
• Repair: the author makes targeted completions;
• Result: overall story completeness rises by a level.

Engineering interpretation: this is a typical L5 discovery → human decision → L4 repair loop. The key is that AI only does discovery (global retrieval, which machines excel at) and not repair (creative judgment, which humans excel at). This division is safer than "having AI rewrite directly", and is more consistent with Yuewen's stance that "AI handles manual labor, not intellectual work".

6.2 Case 2: 100,000 Average Subscriptions in 165 Days

Background: the growth cycle for a newcomer author in the web-novel industry is usually several years; reaching 100,000 average subscriptions (the average subscription count across all VIP chapters) is the level of top-tier works.

Approach and effect: newcomer author Heshouyuemanchi reached 100,000 average subscriptions in just 165 days, setting an industry record (per Yuewen's official account).

Verifiable context:

Causal boundary to be stated explicitly: official materials present the 165-day 100,000-average-subscriptions achievement alongside Miaobi's capabilities, but provide no direct evidence that this author used Miaobi, nor any causal argument. This group records it as the phenomenon that "the platform's AI capabilities and the acceleration of a newcomer's growth occurred in the same period", and does not claim causality.

6.3 Case 3: Manga Drama Assistant's IP Form Derivation

Background: the traditional production cycle for adapting a web novel into a manga drama is about 90 days with relatively high cost, limiting IP-derivative capacity.

Approach: Manga Drama Assistant was launched in October 2025, offering full-workflow support from content understanding, creation suggestions, and visual style to material production; it relies on Yuewen's massive licensed IP database.

Effect:

Engineering interpretation: this is the only industry-level case in this group with complete quantitative results. Its success depends on three conditions stacked together — licensed data (L6) + ten-million-character understanding (L1) + a full-workflow toolchain (L2). Missing any one makes it unattainable: without licensed data it cannot be legally adapted, without ten-million-character understanding one must read the original manually, and without a full-workflow toolchain only a single step is sped up.

7. Summary

7.1 Strengths

1. L1 capacity is the strongest in this group: ten-million-character deep understanding of web novels + reading a million characters in 5 minutes, the only platform that directly solves the long-form problem on the capacity dimension.
2. The clearest product positioning: the "second brain" does not generate chapters, concentrating resources on memory and retrieval, precisely hitting the core pain point of long-form works.
3. Foreshadowing mining and detail retrieval are the only explicit consistency capability among the commercial platforms in this group.
4. The only platform that discloses verifiable business metrics: weekly usage rate >75%, DAU growth >100%, tokens +90%.
5. The three assistants cover the entire IP lifecycle, the only architecture spanning creation–distribution–adaptation.
6. Proactive governance: licensed IP database + DeepSeek-R1 independent deployment + written authorization, avoiding training-authorization disputes.
7. Manga Drama Assistant has complete quantitative results: 90 days → 10–13 days, adopted by 100+ studios.

7.2 Limitations and Known Shortcomings

1. Pricing is completely opaque: no price list was found, making accessibility hard to judge.
2. AIGC labeling implementation not disclosed: the _Measures for the Labeling of AI-Generated Synthetic Content_ took effect on 2025-09-01, but no specific implementation details were found.
3. L4 is checking-type rather than tracking-type: foreshadowing mining requires the author to proactively ask; there is no proactive reminder or versioned ledger.
4. No timeline mechanism, no versioned snapshots: an obvious gap versus the Claude workflow.
5. No generation-type tools: insufficient for authors who want AI to directly produce drafts.
6. The causal chain behind 100,000 average subscriptions in 165 days has not been officially established, so it should not be cited as proof of capability.
7. Limited openness: mainly aimed at the signed-author system; accessibility for non-signed authors [To be filled].

7.3 Applicability Boundaries

7.4 Selection Recommendations

• Recommended scenario: you are a Chinese long-form web-novel author whose most painful problem is "forgetting the foreshadowing planted earlier while writing". Miaobi Tongjian’s foreshadowing mining and detail retrieval are the most direct solution in this group.
• Combination suggestion: Yuewen’s positioning is "checking and memory", not "generation and orchestration". If you need a complete generation pipeline, combine it with the Lingxie Creation described in 07-chinese-webnovel-ai.md (Project Constitution + AI editor) — the former handles consistency, the latter handles output.
• IP adaptation scenario: if you care about web-novel derivatives (manga drama, anime, film/TV), Yuewen’s three-assistant architecture is the only full-chain solution available in this group.
• Compliance note: when using Yuewen’s AI capabilities, you still need to confirm the AI reporting requirements of your target publishing platform yourself (especially Fanqie’s mandatory reporting), because Yuewen does not provide an AIGC labeling capability.

Information Gap Statement

1. Pricing of Yuewen Miaobi / Writer Assistant AI: Yuewen has never disclosed standalone pricing for authors; no price list was found, marked [To be filled].
2. Manga Drama Assistant pricing: only "over 100 studios pay to use it" was disclosed; the specific price was not, marked [To be filled].
3. AIGC labeling implementation: no public statement from Yuewen on compliance with the _Measures for the Labeling of AI-Generated Synthetic Content_ was found, marked [To be filled].
4. Scale of the licensed IP database: sources only describe it as "massive", with no specific number, marked [To be filled].
5. Accessibility for non-signed authors: no statement found, marked [To be filled].
6. Developer API: no public information found, marked [To be filled].
7. Parameter scale and technical details of the Miaobi large model: no public statement found, marked [To be filled].
8. Causal relationship between 100,000 average subscriptions in 165 days and Miaobi: official materials present them side by side but provide no causal argument; this group does not claim causality and has noted this in the main text.
9. Per-work cost of Manga Drama Assistant before the transformation: not disclosed, marked [To be filled].
10. Whether L4 has proactive reminder capability: no statement found; treated as "requires the author to query proactively".

8. References

1. 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
2. China Youth Daily "'Miaobi Tongjian' and 'Manga Drama Assistant' released: AI empowers web-novel creation and IP adaptation" (Tencent) — China Youth Daily, 2025. https://new.qq.com/rain/a/20251017A08J6J00
3. Tencent News "Manga dramas grow over 45% annually: IP and AI dual engines ignite a hundred-billion market" — Tencent News, 2026. https://new.qq.com/rain/a/20260319A03C2U00
4. GeekPark "Can AI write 'Celebrating Years'?" — GeekPark. https://so.html5.qq.com/page/real/search_news?docid=70000021_10168f9fb6122152
5. Yuewen Group official website — Yuewen Group. https://www.yuewen.com/
6. Sohu Culture "2024–2025 Biennial List, Preface 1: The AI era, the 'grand literary view', and the gentleness of the world's dark side" — Sohu Culture, 2025. https://cul.sohu.com/a/1058758339_121124749
7. Haike Finance "Fanqie Novel's AI dilemma" (Sina Finance) — Haike Finance, 2025. https://finance.sina.com.cn/search/2025-10-09/doc-infthsqh9655363.shtml
8. Fanqie Novel official announcement "Notice of the launch of AI writing tool features" — Fanqie Novel, 2024. https://fanqienovel.com/writer/zone/article/7327136545129906238
9. China News Service "Has creative effort become AI fodder? Web-novel authors 'rise up' while Fanqie Novel hurriedly clarifies" — China News Service, 2025. https://www.jwview.com/jingwei/html/07-26/601732.shtml
10. 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
11. Hands-on article "Writing novels with Claude Code unexpectedly revealed a team cheat mode" — 2026. https://m.aitntnews.com/newDetail.html?newId=20196
12. Toutiao "Web-novel writing technique notes: from beginner to advanced" — Toutiao. https://m.toutiao.com/article/7666250379276927531
13. Meipian "How beginners start writing web novels" — Meipian. https://www.meipian.cn/5gn73n9w