humanizer-lint is an open-source natural-text and Markdown prose linter that deterministically identifies style patterns often associated with AI-generated writing in Indonesian and English; it is a style validator, not an authorship detector.

Try in your browser

Use the official demo: humanizer.aiclub.ai.id.

humanizer-lint.pages.dev remains available as a mirror.

It works locally with zero runtime dependencies. Instead of rewriting text through another language model, it reports paragraph- and sentence-level violations with specific correction instructions. Use it to lint LLM-generated documentation, remove robotic AI phrasing from Markdown, check burstiness and lexical diversity, or add a deterministic prose gate to an agent workflow.

It can serve as a CLI text humanizer guardrail, automated writing style linter, or CI/CD documentation linter. The bundled rules flag overused AI clichés such as delve and tapestry without rewriting the source.

What it does

humanizer-lint analyzes prose after removing or neutralizing common non-prose Markdown structures. It then combines lexical rules, sentence-length statistics, structural checks, Indonesian morphology checks, and an optional voice profile into a score from 0 to 100.

The project includes:

• 86 reviewable lexical and structural regex rules: 42 Indonesian and 44 English.
• Sentence-length burstiness, adjacent rhythm delta, short/long sentence ratios, and adaptive MATTR.
• Repetition checks based on paragraph symmetry, repeated sentence starters, and compression ratio.
• Indonesian nominalization, passive-voice, loanword, and translation-style opening checks.
• Optional voice profiles generated from one to five writing samples.
• Human-readable terminal output and pure JSON output.
• Standard exit codes for agents, scripts, and CI jobs.

It does not estimate perplexity, classify who wrote the text, rewrite prose automatically, scan directories, or provide SARIF output.

Terminal demo

Given this input:

In today's fast-paced world, teams delve into a robust landscape. Moreover, this pivotal approach fosters seamless collaboration across every project. In conclusion, the solution unlocks valuable insights for everyone.

Run the analyzer:

humanizer-lint --file draft.md --lang en

The current CLI reports:

Status: FAIL
Score: 39
Language: en

Metrics:
  wordCount: 29
  sentenceCount: 3
  meanSentenceLength: 9.6667
  stdDev: 0.4714
  burstinessCV: 0.0488
  adjacentDelta: 0.5
  shortSentenceRatio: 0
  longSentenceRatio: 0
  mattr: 1
  compressionRatio: 0.6972
  nominalizationRatio: 0
  passiveRatio: 0
  voiceProfileDeviation: 0.5319

Violations:
  SLOP_PHRASE | Paragraph 1, Sentence 1 | Name the exact action or subject instead.
  SLOP_PHRASE | Paragraph 1, Sentence 2 | Describe the actual consequence without inflated significance.
  SLOP_PHRASE | Paragraph 1, Sentence 2 | Name the concrete behavior or outcome.
  SLOP_PHRASE | Paragraph 1, Sentence 1 | Name the field or conditions directly.
  SLOP_PHRASE | Paragraph 1, Sentence 3 | Name what becomes possible and how.
  SLOP_PHRASE | Paragraph 1, Sentence 2 | Connect the sentences by their actual relationship.
  SLOP_PHRASE | Paragraph 1, Sentence 3 | End on the final concrete point.
  SLOP_PHRASE | Paragraph 1, Sentence 1 | Name the current condition precisely.
  SLOP_PHRASE | Paragraph 1, Sentence 2 | Describe the missing step, delay, or transition precisely.
  SLOP_PHRASE | Paragraph 1, Sentence 1 | Name the failure mode or capacity the system handles.
  SLOP_PHRASE | Paragraph 1, Sentence 3 | State the finding itself.
  MONOTONOUS_RHYTHM | Paragraph 1, Sentences 1-3 | Sentence lengths are uniform (10, 10, 9 words). Split one sentence into a short statement or merge two into a longer clause.
  LOW_BURSTINESS | Whole text | Vary sentence lengths by splitting one uniform sentence and merging another where the meaning supports it.
  LOW_ADJACENT_DELTA | Whole text | Change neighboring sentence lengths by splitting or merging only where the content allows it.

A failed analysis exits with code 1. A passing fixture returns:

Status: PASS
Score: 100
Language: en

and exits with code 0.

Why humanizer-lint

Many AI humanizer services send text to another model and return a rewritten version. That can alter facts, quotations, formatting, or technical meaning. humanizer-lint takes the smaller and more predictable approach: it identifies the prose that needs attention and leaves the edit under your control.

Capability	humanizer-lint	Web rewriting service	General grammar checker
Detects configured LLM clichés	Yes, with rule IDs and locations	Usually rewrites them	Usually not
Measures sentence rhythm	Yes	Usually not exposed	Limited
Preserves source text	Yes, analysis only	No, returns a rewrite	Depends on product
Handles common Markdown structures	Yes, through deterministic sanitization	Product-dependent	Product-dependent
Runs offline after installation	Yes	Usually no	Product-dependent
Supports Indonesian and English lexical rules	Yes	Product-dependent	Product-dependent
Produces JSON	Yes	Product-dependent	Product-dependent
Proves human authorship	No	No	No

If you are looking for an AI writing detector linter, a ChatGPT tone remover CLI, or an open-source alternative to AI humanizers, note the boundary: this tool detects configured style signals and rhythm problems. It does not claim to determine whether a human or model wrote the text.

Install

CLI

Run the published CLI without a global install:

npx humanizer-lint --file tulisan.md --lang id

Or install it globally:

npm install -g humanizer-lint

Install as an agent skill

The recommended installation for coding agents is through the skills CLI:

npx skills add yanstd888/humanizer-lint --skill humanizer-lint

The public listing, installation command, and install count are available on skills.sh.

Development (from source)

Clone the repository and use npm link only when developing the local source:

git clone https://github.com/yanstd888/humanizer-lint.git
cd humanizer-lint
npm test
npm link
humanizer-lint --help
humanizer-fingerprint --help

No dependency installation or build step is required.

Quick start

Analyze one file:

humanizer-lint --file draft.md --lang auto

Return machine-readable JSON:

humanizer-lint --file draft.md --lang id --json

Analyze stdin without creating a temporary file:

cat draft.md | humanizer-lint --stdin --lang auto --json

Use a custom voice profile:

humanizer-lint --file draft.md --lang en --profile profiles/custom-voice.json --json

Without npm link, replace humanizer-lint with node scripts/analyze.mjs and replace humanizer-fingerprint with node scripts/fingerprint.mjs.

Contoh / Example

Skor mengukur kecocokan dengan sinyal gaya yang diperiksa oleh linter: skor lebih tinggi berarti lebih sedikit pelanggaran yang ditemukan. Ini bukan skor probabilitas penulis AI atau manusia.

Before

Di era digital saat ini, penggunaan teknologi menjadi sangat penting bagi setiap bisnis. Tidak hanya itu, solusi ini juga memberikan manfaat yang sangat besar bagi semua pihak. Dengan demikian, organisasi dapat meningkatkan efisiensi secara signifikan dalam berbagai aspek. Pada akhirnya, pendekatan ini membuka peluang baru yang lebih optimal bagi masa depan.

Output dari tests/fixtures/ai-before-id.md:

Status: FAIL
Score: 38
Language: id

Violations:
  SLOP_PHRASE | Paragraph 1, Sentence 1 | Mulai langsung dari subjek atau fakta konkret.
  SLOP_PHRASE | Paragraph 1, Sentence 3 | Sebutkan aspek yang relevan.
  SLOP_PHRASE | Paragraph 1, Sentence 4 | Tutup langsung dengan akibat atau keputusan akhir.
  MONOTONOUS_RHYTHM | Paragraph 1, Sentences 1-3 | Sentence lengths are uniform (13, 14, 11 words).
  LOW_BURSTINESS | Whole text | Vary sentence lengths by splitting one uniform sentence and merging another where the meaning supports it.
  LOW_ADJACENT_DELTA | Whole text | Change neighboring sentence lengths by splitting or merging only where the content allows it.

After

Pemilik toko memakai aplikasi stok untuk melihat barang yang habis pada hari yang sama. Semua melihat angka yang sama. Saat satu produk menipis, petugas memesan ulang sebelum rak kosong dan mencatat alasan setiap perubahan. Laporan harian menunjukkan barang mana yang laku, barang mana yang tertahan, dan kapan petugas harus menghubungi pemasok lagi agar toko tidak kehabisan barang yang dicari pelanggan. Pemilik toko kini membeli barang berdasarkan catatan, bukan tebakan.

Output dari tests/fixtures/human-after-id.md:

Status: PASS
Score: 100
Language: id

Untuk contoh ini, skor naik dari 38 menjadi 100 setelah frasa template dihilangkan, subjek dibuat konkret, dan panjang kalimat divariasikan.

CLI reference

Analyzer

humanizer-lint (--file <path> | --stdin) [--lang id|en|auto] [--profile <path>] [--json]

Option	Required	Description
--file <path>	One input required	Read one UTF-8 text or Markdown file.
--stdin	One input required	Read the complete document from standard input.
`--lang id	en	auto`
--profile <path>	No	Load a JSON voice profile.
--json	No	Print JSON only. Without it, print the terminal report.
--help	No	Print command usage.

Exactly one of --file or --stdin is required.

Voice fingerprint

humanizer-fingerprint --input <file-or-dir> --output <profile.json> [--name <name>] [--lang id|en|auto]

Option	Required	Description
--input <path>	Yes	Read one file or up to five .md and .txt files from one directory.
--output <path>	Yes	Write the generated profile atomically.
--name <name>	No	Set the profile name. Default: custom-voice.
`--lang id	en	auto`
--help	No	Print command usage.

Exit codes

Code	Meaning
0	PASS, or SKIP for text too short for statistical scoring.
1	Analysis completed and returned FAIL.
2	Invalid arguments, unreadable input, or malformed profile.

Detectors and violation types

The JSON rule catalogs contain stable IDs. Runtime statistical checks use named violation types.

Violation type or rule range	Language	What it detects
ID001-ID040, EN001-EN040 / SLOP_PHRASE	Indonesian and English	Configured clichés, generic transitions, inflated wording, and repeated LLM vocabulary such as delve, tapestry, testament, and their Indonesian equivalents.
ID041-ID042 / DEFINITIONAL_OPENING	Indonesian	Formulaic paragraph openings and rigid definition templates.
EN041-EN044 / STRUCTURAL_CLICHE	English	Participial tails, binary formulas, repeated double-hyphens, and similar structural templates.
MONOTONOUS_RHYTHM	Both	Three adjacent sentences whose lengths differ by at most three words.
LOW_BURSTINESS	Both	Sentence-length coefficient of variation below the configured threshold.
MISSING_SENTENCE_RANGE	Both	Text with at least eight sentences that lacks enough short or long sentences.
LOW_ADJACENT_DELTA	Both	Consecutive sentence lengths that change too little.
LOW_LEXICAL_DIVERSITY	Both	Adaptive MATTR below the language threshold.
LOW_ENTROPY_REPETITION	Both	Low DEFLATE compression ratio in text of at least 150 words.
SYMMETRIC_PARAGRAPHS	Both	Three consecutive paragraphs with the same sentence count.
REPETITIVE_STARTERS	Both	A paragraph with at least three sentences beginning with the same token.
EXCESSIVE_NOMINALIZATION	Indonesian	Nominalization density above the configured threshold after whitelist filtering.
PASSIVE_BLOATING	Indonesian	Passive morphology density or consecutive translation-style subordinate openings.
VOICE_PROFILE_DEVIATION	Both	Custom-profile mean sentence length that differs from the target by more than 30 percent.

Metrics and scoring

Metric	Definition
burstinessCV	Population standard deviation divided by mean sentence length.
adjacentDelta	Mean absolute difference between neighboring sentence lengths.
mattr	Moving-average type-token ratio over an adaptive 10- to 50-token window.
compressionRatio	Raw DEFLATE byte length divided by original UTF-8 byte length.
shortSentenceRatio	Share of sentences containing six words or fewer.
longSentenceRatio	Share of sentences containing 26 words or more.
voiceProfileDeviation	Mean relative deviation from available profile targets. It is not an authorship probability.

The score starts at 100 and applies these caps:

Category	Maximum deduction
Burstiness and sentence-range checks	25
Monotonous rhythm windows	20
Lexical and structural clichés	30
Indonesian morphology	15
Vocabulary, entropy, and custom-profile deviation combined	10

A normal-length document passes only when the score is at least 80, no SLOP_PHRASE or DEFINITIONAL_OPENING remains, burstiness and adjacent-delta gates pass, and the sentence-range requirement passes when applicable.

Documents below 20 words or three sentences return SKIP. Lexical slop phrases are still reported.

Markdown sanitization

Before tokenization, the analyzer:

• removes YAML frontmatter, fenced code blocks, headings, and Markdown tables;
• keeps Markdown link anchor text while removing the URL;
• replaces inline code and bare URLs with TOKEN_ENTITY;
• removes short or unfinished list fragments while keeping complete list prose;
• removes blockquote markers while preserving the quoted prose;
• preserves paragraph boundaries for locations and structural checks.

This is deterministic sanitization, not a full Markdown or MDX abstract syntax tree parser. HTML blocks and arbitrary MDX components do not receive special parsing.

Voice profiles

Generate a profile from one to five samples:

humanizer-fingerprint \
  --input writing-samples \
  --output profiles/custom-voice.json \
  --name editorial-voice \
  --lang en

The profile records mean sentence length, standard deviation, burstiness, adjacent delta, MATTR, average paragraph sentence count, and punctuation rates. Supplying it with --profile adds voiceProfileDeviation; a mean sentence-length drift above 30 percent creates a five-point profile penalty inside the 10-point vocabulary and entropy cap.

JSON output

--json writes a single JSON object to stdout. The stable top-level fields are:

{
  "status": "FAIL",
  "score": 39,
  "language": "en",
  "metrics": {
    "wordCount": 29,
    "sentenceCount": 3,
    "burstinessCV": 0.0488,
    "adjacentDelta": 0.5,
    "mattr": 1
  },
  "violations": [
    {
      "type": "LOW_BURSTINESS",
      "location": "Whole text",
      "actual": 0.0488,
      "target": 0.4,
      "fixInstruction": "Vary sentence lengths by splitting one uniform sentence and merging another where the meaning supports it."
    }
  ]
}

The real result may contain additional metric properties and violations.

Use as an agent skill

After installation, an agent can use the skill as a closed verification loop:

1. Draft or edit Indonesian or English prose without changing supported facts.
2. Run the analyzer against the exact complete draft.
3. Read every violation and edit only the affected prose or named whole-document gate.
4. Re-run the analyzer, stopping after three iterations.
5. Return only a PASS, or an honestly labeled short-text SKIP with no unresolved lexical violations.

The skill is indexed on skills.sh and its current listing passes the platform's Snyk security audit.

Used by

The skills.sh badge at the top of this README is the public source for anonymous installation activity. GitHub's native Used by sidebar is generated automatically from its dependency graph; repository owners cannot set that count manually.

Because the CLI is not published to npm, skill installations are represented by skills.sh rather than GitHub package dependents. The GitHub panel will appear when downstream repositories declare a dependency that GitHub can detect.

CI and repository automation

This repository keeps .github intentionally small. The included .github/workflows/ci.yml workflow runs the test suite and package dry-run on supported Node.js versions. A CI badge is useful because it reflects an actual workflow; issue templates, Dependabot, and release automation are omitted until the repository needs them.

To use the analyzer in another GitHub Actions workflow without an npm release, check out this repository into a subdirectory and run the script directly:

name: prose-check

on:
  pull_request:
    paths:
      - "**/*.md"
      - "**/*.txt"

jobs:
  lint-readme:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/checkout@v4
        with:
          repository: yanstd888/humanizer-lint
          path: .humanizer-lint
      - uses: actions/setup-node@v4
        with:
          node-version: "22"
      - run: node .humanizer-lint/scripts/analyze.mjs --file README.md --lang auto --json

The analyzer accepts one file per invocation. Invoke it once for each document you want to gate.

Security properties

Production scripts use only Node.js built-ins and have no runtime dependencies. The test suite rejects production use of child_process, eval, dynamic Function construction, and fetch. The fingerprint writer writes to a temporary file and renames it into place.

Text remains local unless the surrounding agent or workflow sends it elsewhere.

Frequently asked questions

Is humanizer-lint an AI content detector?

No. It detects configured lexical, structural, rhythm, and morphology signals. A pass or failure does not prove who wrote the text.

Does humanizer-lint rewrite or auto-fix prose?

No. It returns correction instructions but does not modify the source file. This prevents silent factual or formatting changes.

Can it remove robotic AI phrasing from Markdown?

It can identify the exact paragraph or sentence that matches a configured phrase or style gate. A human or coding agent then makes the smallest supported edit and runs the analyzer again.

Can I install it from npm?

Not currently. Install the agent skill with npx skills add yanstd888/humanizer-lint --skill humanizer-lint, or clone the repository and use npm link for the local CLI binaries.

Which languages are supported?

The configured lexical and language-detection paths support Indonesian and English. Do not treat the current statistical thresholds as validated defaults for other languages.

Does it send text to an API?

No. Production code has no network call. The skills installer and Git operations need network access only to download the repository.

Why did a short document return SKIP?

Documents below 20 words or three sentences do not contain enough material for the configured statistical gates. Lexical slop phrases are still checked.

Contributing

Contributions should remain deterministic, zero-dependency, and covered by a focused regression test.

1. Fork the repository.
2. Create a focused branch.
3. Add or update one minimal test for non-trivial behavior.
4. Run npm test and npm pack --dry-run.
5. Open a pull request that explains the user-visible effect.

Every repository change updates the package and skill version, then publishes a matching Git tag and GitHub release after verification.

Project layout

.github/
  workflows/
    ci.yml
LICENSE
README.md
SKILL.md
package.json
profiles/
  default-human.json
rules/
  en-slop.json
  id-slop.json
scripts/
  analyze.mjs
  fingerprint.mjs
  utils/
    metrics.mjs
    sanitizer.mjs
    tokenizer.mjs
tests/
  fixtures/
  test-runner.mjs

License

Licensed under the MIT License.