Tianming and the two bugs its own README says to fix first
Tianming is an AI novel-writing system with 15-dimensional fact snapshots, 12 types of change declarations, and 6 generation gates. Stays coherent through 3000 chapters without relying on context or model memory, only on per-chapter state writebacks.
At a glance
- What is it?
- Tianming is a Windows desktop system for long-form Chinese web novels that refuses to depend on a context window: every chapter passes six gates, emits a machine-readable change declaration, and writes the result back into a 15-dimension fact snapshot. The repository is written in Chinese, ships an open-source build with two known bugs, and states MIT while asking that commercial use be arranged with the author.
- Who is it for?
- Tianming suits a writer who is generating hundreds of chapters in Chinese and needs state bookkeeping that survives past the context window, because the change declaration and fact snapshot carry continuity forward where a prompt cannot. It does not suit anyone who needs a licence that permits commercial use without a conversation with the author, since the MIT statement and the commercial notice contradict each other, and it does not suit a non-Windows machine at all.
- Can I use it commercially?
- Not without permission. GitHub finds no licence file in the repository, and without a licence all rights are reserved by default: you may read the code but not reuse it. Check the README, or ask the authors, before using it.
- Is it still maintained?
- Yes. The repository last received commits 11 days ago.
- What is it written in?
- Mainly C#, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on November 15, 2026, and from our analysis. They are not legal advice.
Star history
- zy-zmc/tianming-novel-ai-writer456 ★ · +21
| Repository | 2026-09-20 | 2026-11-15 |
|---|---|---|
| zy-zmc/tianming-novel-ai-writer | 435 | 456 |
21 Stars in the last 56 days, measured from the daily GitHub snapshot. Compare with other projects
Editorial analysis
The homepage field points at a different repository
The repository is zy-zmc/tianming-novel-ai-writer, a C# WPF application. Its homepage field points at github.com/zy-zmc/tianming-skill, which is a separate project. The README explains the split in a block near the top: if you want the underlying protocol system and prompt engineering rather than the application, go to the Skill repository, which is aimed at people who want to understand the mechanism and port it elsewhere.
So there are two artifacts under one owner. The application manages a book; the Skill repository documents the protocol. Worth knowing before you follow the homepage link expecting an installer, because what you will find there is the prompt layer.
The README itself is written entirely in Chinese, and it says so with a bluntness worth repeating in translation: if you cannot read Chinese, this is not for you. Every architectural term below comes from that document.
MIT in the text, a commercial notice under it, no LICENSE file
The licence section links MIT License to a LICENSE file, and the licence field on the repository is empty. There is no LICENSE entry among the top-level files, which are Core/, Docs/, Framework/, Modules/, Scripts/, Services/, Storage/, README.md, 开源说明.md and a handful of directories with Chinese names.
The contradiction is not about the missing file. It is the paragraph directly underneath, labelled a commercial notice: the code is open under MIT, but any commercial use, including repackaging for sale, commercial SaaS deployment, or embedding in a paid product, requires contacting the author first. MIT grants those uses unconditionally. A notice that withdraws them is not an MIT licence; it is a different agreement wearing MIT's name, and it is worth settling with a lawyer rather than with a README.
For what it is worth, the repository states that the open-source cleanup is finished: the encryption protection system was removed and private domains were replaced with placeholders.
Two version numbers in the tree, and no release at all
The repository has no GitHub releases, yet the root carries a directory called 1.4.6-天命 and a text file called v2.8.7-升级说明.txt, which is an upgrade note for version 2.8.7. Those two names disagree by more than a major version, and neither is a tag you can install.
The README explains part of it: the author deleted the old source code rather than re-download and re-upload it, and asks readers to look at the section on missing items because of that. The result is a repository where the version number lives in three unrelated places, none of which a package manager can read. Nothing pins a build, so there is no way to tell from outside whether your copy is the one the tutorials describe.
The tutorials, in turn, are both labelled as covering the old version. The author notes that development moves faster than the videos can follow.
Windows, WPF, and a project file named in Chinese
The stack is .NET 8.0 with WPF, MVVM with dependency injection, Microsoft Semantic Kernel as the AI engine and ONNX Runtime for local embeddings. WPF is the decisive constraint: this is a Windows desktop application, and the stated requirement is Windows 10 build 19041 or newer with the .NET 8.0 SDK installed.
Two commands get you to a first run:
git clone https://github.com/zy-zmc/tianming-novel-ai-writer.git
dotnet run --project Core/App/天命.csprojThe project file is named 天命.csproj, so the path argument is not plain ASCII and has to survive your shell's encoding. On first launch you configure an AI API key in the model manager; the project states that no key is bundled and that OpenAI and Anthropic compatible interfaces are supported.
The source tree is five directories with a stated dependency direction: Modules to Services to Framework, with Core on the other side of Framework.
Six gates stand between a draft and the file system
Each chapter runs a closed loop rather than a pipeline: read before writing, AI writes, gates validate, the text is stored, state is written back, and the next chapter reads the fresh values. The repository description claims coherence out to chapter 3000 without relying on context or model memory; no measurement is offered for that claim.
Six gates decide whether text is allowed to land. Protocol parsing requires a `---CHANGES---` delimiter followed by a complete JSON declaration, and missing fields are rejected outright. Reference validation checks that every character, location and faction id cited in the declaration exists in the design data. Then consistency, unknown-entity detection, description consistency for hair and eye colour, and a blueprint check that named characters actually appear.
Two of those have numeric thresholds. More than five unknown entities rejects the chapter, and so do more than three of them being extras with no change record. A viewpoint character who appears only once gets a warning that narrative intensity is insufficient, which is softer than a rejection.
The open-source build ships two bugs the author says block generation
The README has a section headed 重点缺失, missing items, and it is the most useful part of the document. The author's position is that you can hand these to an AI to fix, and that the first two must be fixed before generation is stable.
One is a mapping bug between the Chinese and English forms of CHANGES, which means the model performing validation cannot recognise the declaration and the gates are validating nothing. The other is missing injection of the rewrite-feedback dimension, so a rejected chapter does not come back with what went wrong. Two further items are listed as improvements rather than blockers: the Beta build's UTF-8 handling was replaced with Console.OutputEncoding and the open-source version did not receive that change, described as about ten lines of code, and the state machine needs strengthening because it records state too early and the model then misjudges.
In other words, the published build is not the build the author uses.
Both ONNX models are in the repository, and deleting them disables two features
Two local models ship inside the repository rather than being downloaded at runtime. They are copied into the output directory at build time through Services.props, so no manual download or placement is needed.
One is bge-small-zh-v1.5, a small Chinese BERT for semantic search and vector recall, and it is also the mechanism behind the long-range recall that pulls a chapter 30 fragment forward to chapter 500. The other is chinese-roberta-tiny, a masked language model used for word-level substitution that the README frames as reducing the AI flavour of generated paragraphs, paired with a semantic guard that checks similarity before and after so the substitution cannot drift the meaning.
Both unload from memory after a default idle period of ten minutes and reload on demand. Delete either Resources directory and the application degrades on purpose: vector search and the word-choice feature switch off, everything else keeps working.
The README says the tutorials are out of date, and names QQ as support
Support runs through a QQ group, listed with a join link and the number 414086347, described as a bug-collection group for the intelligent writing system. A separate note explains that the official version is finished testing and the group is now an open-source mutual-aid group, and that the open-source build needs no key activation while the commercial one does, with 开源说明.md covering how the key system works.
The same passage says not to bother joining for the sake of it. The more telling line is the admission that both Bilibili tutorials, one for the official version and one for the open-source build, cover the previous release, and that the author posts shorter quick-start videos that are also out of date because the project moves faster than the videos.
One module deserves a flag for anyone considering this for work that is not their own. The design module includes a built-in crawler that scrapes web-novel platforms and analyses structure from what it collects. The document says nothing about rate limits, robots.txt or platform terms, so treat that feature as a decision to make deliberately.
Editorial conclusion
Tianming suits a writer who is generating hundreds of chapters in Chinese and needs state bookkeeping that survives past the context window, because the change declaration and fact snapshot carry continuity forward where a prompt cannot. It does not suit anyone who needs a licence that permits commercial use without a conversation with the author, since the MIT statement and the commercial notice contradict each other, and it does not suit a non-Windows machine at all. Before you build it, check two things: that the CHANGES English-Chinese mapping bug listed under the missing-fixes section is not still present in your copy, and that the 天命.csproj path resolves, since the project file name is in Chinese.
Frequently asked questions
What operating system does the Tianming novel writer need?
Windows 10 build 19041 or newer, plus the .NET 8.0 SDK. The interface is built on WPF over .NET 8.0, with MVVM and dependency injection, and Microsoft Semantic Kernel as the AI engine.
How do I build and run Tianming from source?
Clone the repository and run dotnet run --project Core/App/天命.csproj. The project file name is in Chinese characters, and the source tree is split across Core, Framework, Services, Modules and Storage with the dependency direction Modules to Services to Framework.
Does Tianming require me to bring an AI API key?
Yes. The project states that you configure your own key in model management on first run, that OpenAI and Anthropic compatible interfaces are supported, and that no key is included in the repository.
Can I use the Tianming source commercially?
The README states MIT and links a licence file, then adds a commercial notice asking that repackaging for sale, commercial SaaS deployment and embedding in paid products be arranged with the author first. No licence file appears among the top-level entries.
What are the six generation gates in Tianming?
Protocol parsing, reference validation, consistency validation, unknown-entity detection, description consistency and a blueprint appearance check. Text is written to storage only after passing all six, and more than five unknown entities in a chapter is one of the stated rejection conditions.
Which local models does Tianming bundle with the source?
Two ONNX models, copied into the output directory at build time by Services.props: bge-small-zh-v1.5 for semantic search and vector recall, and chinese-roberta-tiny for word substitution and the semantic guard. Removing either Resources directory disables that feature and leaves the rest working.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/zy-zmc-tianming-novel-ai-writer)Star history chart for your README
[](https://hysenlabs.com/projects/zy-zmc-tianming-novel-ai-writer)Was this useful?