MIT-licensed DSH plugin that turns 👍/👎 into an editable taste profile, then drafts chapters, scoped rewrites and plot diagnostics from it — manuscripts stay local and only the few-hundred-byte rules file enters the writing context.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-21
License
MIT
01What can it help you accomplish?
Calibrate the model to your own prose taste without writing critiques
An editable author-preference profile holding only rules (the author's own file is 714 bytes), each rule ticked on intake, deletable, and traceable back to the paragraphs it came from
Authors already drafting fiction or web novels inside DeepSeek Harness who are tired of re-tuning prompts every session
Draft a new chapter as a batch of candidates, then merge them into a final
One file per draft (`第13章-A-保守精修.txt`-style names, following your own convention), a pick-list of the paragraphs you approved with your one-line reason, `〔此处需过渡〕` markers where candidates do not join, and an automatic backup of the manuscript before it is written
Serial novelists who need dozens of chapters to stay consistent
Revise only the paragraphs you annotated
A paragraph-by-paragraph before/after, a backup on accept, and a hard refusal to write back if any unannotated paragraph changed by a single character or the paragraph count moved
Authors clearing AI-ish tone, padding or spelled-out emotion without losing the rest of the chapter
See where a 100k-character book is collapsing
Tension curve (manual 1–5 marks in solid, estimated dotted values with their basis), outline drift, foreshadowing debt, word-count imbalance, event density and rating decline — all computed locally
Long-form authors who are past the point where memory still works
02How to install into DeepSeek Harness
Prerequisites
- DeepSeek Harness — the README states it as the prerequisite; the command-line DSH 0.2.2-rc.1 build needs Node 24.9 or newer, while the official Desktop bundles its own runtime
Installation steps
- 01
DSH Web: `dsh plugin --profile web add dsh-novel-craft`, then restart dsh once (the host half is loaded at startup).
$ dsh plugin --profile web add dsh-novel-craft
- 02
Alternative: download the latest plugin package from GitHub Releases and, in the download directory, run `dsh plugin --profile web add "$PWD/dsh-novel-craft-0.3.1.tgz"`.
$ dsh plugin --profile web add "$PWD/dsh-novel-craft-0.3.1.tgz"
- 03
Official DSH Desktop: add `dsh-novel-craft` on the Plugins page and fully quit and reopen the app — or start Desktop once, quit, then run `dsh plugin --profile desktop add dsh-novel-craft` with the `dsh` bundled in Desktop before opening it again.
$ dsh plugin --profile desktop add dsh-novel-craft
- 04
Not ready to install: open `docs/preview.html` from the repository in a browser to click through a snapshot rendered by the real components, all seven tabs included.
Verify the integration
- After the restart the 🎴 card-draw workbench shows up at the bottom of the sidebar — that is the host half having loaded.
03DSH integration and capability boundaries
Native DSH plugin: installed with the `dsh plugin` CLI (or from the Desktop Plugins page), it patches its own row into the profile through `cordis.patch.yml` and appears as the 🎴 card-draw workbench at the bottom of the sidebar
Taste calibration (👍/👎 → rules)
Good/bad marks made while reading candidate paragraphs→An editable preference profile containing rules only, plus the evidence excerpts and a rule→source table behind every rule
writes `.dsh-novel-craft/作者偏好档案.md`, `marks.json`, `证据摘录.md` and `规则来源.json` inside the project directoryspends one auxiliary model call on the marked paragraphs (≤40 paragraphs × 240 chars, ≤20KB total) when you press 提炼规律Chapter board and new-chapter pipeline
Chapter number, chapter brief (goal / must happen / must not happen) and how many directions (10 by default)→Scene-decision directions to pick and rename, one candidate file per draft, a pick-list of approved paragraphs, and a merged final
writes each candidate to disk as soon as it finishes, under the project candidate folderbacks up the manuscript before writing into the body textAnnotation-scoped revision
Paragraph annotations (AI tone / verbose / emotion spelled out / flat / out of character / ledger / information gap / other, plus one line of text)→A revision touching only the annotated paragraphs, with a per-paragraph before/after and a byte-exact guard on everything else
stores annotations in `.dsh-novel-craft/批注/第N章.json`refuses to write back when unannotated paragraphs change or the paragraph count shiftsWriting pack with a context budget
Chapter brief, preference profile, previous chapter ending, plot recap from per-chapter summaries, characters, item/ledger state, plot beats, open foreshadowing, banned-AI-tone list and writing requirements→A ten-section writing pack — the only file the drafting session reads — with a default 9,000-character budget and an itemised cut log
writes `写作包.md` into the round directory, falling back to `.dsh-novel-craft/写作包/`the `pack` call with `save:false` only previews and writes nothingLocal plot diagnostics
Chapter bodies, outlines and per-chapter plot summaries→Tension curve, outline drift, foreshadowing debt, word-count imbalance, event density and rating decline
Loopback HTTP API for the host half
Requests from the browser half on 127.0.0.1→23 routes under `/novel-craft/api/` covering state, marks, distill, chapters, annotations, packs, revisions, tension checks and the new-chapter stages
binds a loopback-only endpoint that answers 503 as a whole while `enabled: false`
04Who is it for? When not to use it?
Good for
- Authors already drafting fiction or web novels inside DeepSeek Harness who are tired of re-tuning prompts every session
- Serial novelists who need dozens of chapters to stay consistent
- Authors clearing AI-ish tone, padding or spelled-out emotion without losing the rest of the chapter
- Long-form authors who are past the point where memory still works
Not for
- It only runs on DeepSeek Harness — the README lists "never used dsh and do not plan to install it" as an explicit non-fit, and states DeepSeek Harness as the prerequisite.
- It does not write the manuscript for you, and it does not let a model read the whole book — plot checks are local string statistics that cannot spot semantic repetition. The UI and the banned-AI-tone list are built for Chinese fiction, so non-Chinese projects get the workflow but not the tone list. It also does not guarantee platform approval and ships no cover, typesetting, publishing or scraping.
05Compatibility, maintenance and safety notes
- It only runs on DeepSeek Harness — the README lists "never used dsh and do not plan to install it" as an explicit non-fit, and states DeepSeek Harness as the prerequisite.
- No version pin: `dsh` is a peer dependency declared as `*` and resolved from the running profile, because DSH packages ship independently. DSH Web is verified on 0.2.2-rc.1 and the command-line build needs Node 24.9 or newer.
- The plugin itself is MIT and free, but every distill, backfill, revision or draft call spends your own DSH model quota; set `distillProvider` / `distillModel` in its settings to route those calls to a cheaper model.
- It does not write the manuscript for you, and it does not let a model read the whole book — plot checks are local string statistics that cannot spot semantic repetition. The UI and the banned-AI-tone list are built for Chinese fiction, so non-Chinese projects get the workflow but not the tone list. It also does not guarantee platform approval and ships no cover, typesetting, publishing or scraping.
MIT · actively maintained (latest release v0.3.1, 2026-11-08; open-sourced 2026-09, 775 assertions across 7 test files)
06Frequently asked questions
How do I install it in DeepSeek Harness?
In DSH Web run `dsh plugin --profile web add dsh-novel-craft` and then restart dsh once, because the host half is loaded at startup. On DSH Desktop, add `dsh-novel-craft` on the Plugins page and fully quit and reopen the app — or start Desktop once, quit, and run `dsh plugin --profile desktop add dsh-novel-craft` with Desktop's own `dsh`. Either way the 🎴 card-draw workbench shows up at the bottom of the sidebar.
Do I have to write prompts or paragraph-by-paragraph critiques?
No. While reading candidates you only press G for good and B for bad (j/k to move, space to clear, n/p to switch drafts), then hit 提炼规律 and tick the rules you accept. Nothing enters the profile unless you tick it, so a bad summary is always yours to reject.
Does my manuscript get sent to the model?
Not while drafting: the writing pack carries the rules, the previous chapter's ending (≤900 characters) and setup notes, with a default 9,000-character budget. Three buttons you press yourself do send text, each with a hard cap — distil (≤40 paragraphs × 240 chars, ≤20KB), backfill sources (160 chars per rule) and annotation-scoped revision (800 chars per paragraph plus 80 either side, ≤20 per run).
Which DSH version and platform does it need?
DeepSeek Harness, and no pinned DSH version: `dsh` is a peer dependency declared as `*` and resolved from your profile at runtime. DSH Web is verified on 0.2.2-rc.1; the command-line build needs Node 24.9 or newer while the official Desktop ships its own runtime. `node scripts/check-dsh-compat.mjs next` re-checks the next DSH release against a real host.
How does it sit next to other DSH plugins?
It is a standalone DSH plugin: it inserts its own row into the profile via `cordis.patch.yml` and uses DSH's public slots, client service and LLM service without copying platform source, so it does not patch other plugins. Model calls default to your current DSH default model; set `distillProvider` / `distillModel` in its settings to route them elsewhere.
07Related DSH workflows
dsh-market
by dsh-market
插件市场 (plugin marketplace): 打开 Settings → Plugin Market 浏览/搜索/一键安装社区插件;安装源限制为 curated awesome-dsh-plugin 注册表(其余拒绝),默认禁用构建脚本(pnpm>=10 需显式开启)。
dsh-guizang-design
by op7418
Expose five content/design skills, a gallery, and multi-skill workflows inside DSH.
oh-story-dsh
by zenstory-ai
Fiction/short-drama writing workbench: 23 skills, roles, markdown/JSONL preview, live file-follow.
dsh-skills-manager
by michengai
DSH 技能集合管理器,统一管理多个 skill。
08Data and sources
**dsh-novel-craft 是一个跑在 DeepSeek Harness(dsh)上的开源 AI 小说写作插件**(MIT 协议,npm 一行安装)。
cordis.patch.yml # 安装时把插件行插进 profile 组合
dsh plugin --profile web add dsh-novel-craft
装完**重启一次 dsh**(宿主半区在启动时装载)。侧栏底部会出现「🎴 抽卡工作台」。
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-21. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
