DSH Novel Quality Analyzer

A community novel analyzer plugin for DeepSeek Harness that checks chapters and story excerpts for vocabulary diversity, clichés, repeated words, and repeated sentences.

Current release: version 1.1.1

Analysis runs locally in your browser. No API key, LLM request, or analysis subscription is required.

This plugin provides writing-style diagnostics. It does not judge plot, literary quality, originality, plagiarism, or AI authorship.

Demo

Excerpt analysis

The analyzer displays the style score, word counts, and vocabulary diversity alongside your excerpt.

Short-sample handling

Excerpts with fewer than 50 words show their metrics without assigning a style score.

Installation

Requirements:

• Node.js 22.22.2 or later; Node 24 LTS is recommended.
• DeepSeek Harness with an existing web-capable profile.
• A modern browser supporting Intl.Segmenter.

Install or update the plugin:

dsh plugin --profile web add dsh-novel-quality-analyzer@latest

Replace web with your existing profile name.

Inspect the profile configuration and launch it:

dsh --profile web --dump-config
dsh --profile web

Installing the plugin into a new base-only profile does not install the Web UI. For initial Harness setup, follow the official quick start.

Alternatively, install from GitHub:

dsh plugin --profile web add github:umairrrkhan/dsh-novel-quality-analyzer

The generated browser bundle is committed, so Git installation does not require building the plugin first.

How to use

1. Open a conversation in DeepSeek Harness.
2. Click Analyze writing in the composer toolbar.
3. Paste a chapter or excerpt into Story excerpt.
4. Results update after a 250 ms pause.
5. Review the flagged phrases and repetitions in context.
6. Click Copy report, or manually copy the full report.
7. Clear or close the panel to discard the excerpt.

The plugin does not read chat history or modify your composer draft.

Features

Feature	What it reports
Word statistics	Total words and unique words
Vocabulary diversity	Moving-average diversity over a default 100-word window
Cliché detection	Every matching occurrence, with bounded position evidence
Content-word repetition	Frequent words, excluding common function words
Repeated sentences	Substantial sentences repeated after token normalization
Sentence metrics	Average sentence length and sentences exceeding 35 words
Reading time	Estimate based on 200 words per minute
Style score	A transparent heuristic with individual penalties
Report export	A copyable plain-text diagnostic summary

The interface includes English and Chinese labels. Text analysis is English-oriented.

Understanding the results

Vocabulary diversity

The plugin reports both the raw unique-word ratio and moving-average diversity.

The moving metric uses a sliding 100-word window to reduce the effect of excerpt length. Smaller samples use a shorter effective window.

Clichés

The current phrase list includes:

• suddenly
• unbeknownst
• out of nowhere
• unknown to them
• before he knew it
• little did they know
• little did he know
• little did she know

All matching occurrences are counted. Up to five example ranges are returned per phrase.

Ranges use JavaScript UTF-16 offsets with an exclusive end position.

Repetition

Content words are flagged when they:

• Contain at least four letters.
• Occur at least six times by default.
• Are not in the common function-word exclusion list.

Up to eight repeated-word examples are returned.

Repeated-sentence detection considers sentences containing at least eight words and returns up to five examples.

Names, dialogue, recurring motifs, and deliberate repetition may be appropriate. Review every flag in context.

Style score

The score is calculated as:

100
− cliché penalty
− content-repetition penalty
− low-diversity penalty
− duplicate-sentence penalty

The result is clamped to 0–100. Individual penalties are included in the analysis output.

A high score means few configured style flags. It does not establish that a story is well written.

Samples below 50 words are not scored. Detected Chinese, Japanese, or Thai text is also not scored because the tokenizer does not reliably segment those languages.

Limits and privacy

• Default maximum: 200,000 UTF-16 string units per excerpt.
• Oversized input is rejected rather than silently truncated.
• Analyze longer novels chapter by chapter.
• Unicode letters, accents, and apostrophes are supported.
• Standalone numbers and punctuation are not counted as words.
• Other languages have not been validated.
• Analysis runs synchronously after debounce; slow devices may pause near the input limit.
• The plugin does not send excerpts to a service or save them to persistent storage.
• Exported reports may contain short snippets from repeated sentences.
• Other Harness features have their own data-handling behavior.

JavaScript API

The analyzer can also be used outside the Harness interface:

npm install dsh-novel-quality-analyzer

import {
  analyzeStory,
  DEFAULT_OPTIONS,
  formatReport,
} from 'dsh-novel-quality-analyzer/analysis';

const result = analyzeStory(storyText);

console.log(result.words);
console.log(result.score);
console.log(result.penalties);
console.log(formatReport(result));

Optional settings:

const result = analyzeStory(storyText, {
  maxCharacters: 200_000,
  diversityWindow: 100,
  repetitionMinimum: 6,
  minimumScoredWords: 50,
});

Options must be positive integers. The API supports a maximum character limit of 1,000,000 and a maximum diversity window of 1,000 words.

Invalid input, unknown options, or oversized text throw TypeError or RangeError. Results are JSON-compatible, and score may be null.

Development

git clone https://github.com/umairrrkhan/dsh-novel-quality-analyzer.git
cd dsh-novel-quality-analyzer
npm ci
npm run check

Useful commands:

npm run build
npm test
npm run benchmark
npm pack --dry-run

Edit:

• lib/analyzer.js — shared analysis implementation.
• src/client.js — Harness browser interface.

Run npm run build after editing. Do not manually edit the generated client.js.

The package has zero runtime npm dependencies. The Harness supplies React and UI primitives; build and test dependencies are development-only.

Validation

Version 1.1.1 passed:

• 28 automated tests on Node 22.22.2.
• 28 automated tests on Node 24.
• Generated browser-bundle consistency checks.
• Package-content checks and artifact upload.
• A fresh public npm installation and analyzer smoke test.
• Verification that the installed README matches the release source.

View the passing GitHub Actions run.

Testing uses actual React with a simulated browser and Harness services. Full live-Harness acceptance testing remains unverified.

DeepSeek Harness is a developer preview with changing APIs. The integration was checked against its published conversation and UI-primitives interfaces at version 0.2.0-rc.2.

Reports and release notes

• Implementation and quality report
• Release verification
• Changelog
• Public npm package

Version 1.1.1 fixes the unsupported Node 22 test-command flag and includes the updated npm README.

GitHub and npm display separate README copies. Editing this file updates GitHub; updating npm’s packaged README requires publishing another package version.

License

MIT.

Maintained by umairrrkhan. This is an independent community plugin, not an official DeepSeek product.