yontu

Turkish humanizer, TDK linter and style guide for AI agents. Türkçe metni yontar: biçim verir, inceltir.

yontu is a Turkish humanizer (Türkçe humanizer) and TDK spelling and punctuation linter, shipped as an agent skill, an offline CLI and an MCP server. It flags AI and translation patterns in Turkish text, checks spelling and punctuation against the Türk Dil Kurumu (TDK) rules, and fits wording to the genre the text is written for. It works with Claude Code, Codex, Cursor, Gemini CLI and any MCP client. English humanizers do little for Turkish, so every rule here was written for Turkish.

Türkçe README

Why

Most humanizers target English. Their rules come from lists such as Wikipedia's "Signs of AI writing", built on English words like "delve" and "tapestry". Run them on Turkish and they find almost nothing. A few patterns, such as the em dash, carry over, but only with a Turkish rule behind them (TDK allows the dash for dialogue only). Most Turkish AI text fails in other ways:

• -mAktAdIr at the end of every sentence, and long -ArAk chains
• calques such as önemli bir rol oynamaktadır and bu bağlamda
• wrappers that replace a case suffix: "konusunda", "hususunda", "bir şekilde"
• chatbot leftovers: Umarım bu e-posta sizi iyi bulur, Umarım yardımcı olur
• spelling and punctuation errors that a careful Turkish writer would not make: herşey, a comma before "ve", Türkçe'de

Turkish also changes shape with context. An academic abstract, an ad, a blog post, a README, a commit message and a court petition each have their own sentence length, formality, pronouns and connectives. A rule that fixes one breaks another: -mAktAdIr is normal in an abstract and a tell in a blog post. Word choice matters as much: hızlı bir şekilde is usually just "hızla".

yontu encodes these as rules with sources, per-genre profiles and a word list, and gives the agent a workflow that keeps the meaning intact.

What's in the box

Before and after

From skills/yontu/references/ornekler.md (README genre):

yontu: Türkçe Metinler İçin Modern Ve Güçlü Bir Linter

yontu, Türkçe metinlerinizi analiz ederek yazım hatalarını, noktalama sorunlarını ve yapay zekâ izlerini tespit eden, modern, güçlü ve kullanıcı dostu bir araçtır. Projenin temel amacı, yazım kalitenizi bir üst seviyeye taşımaktır!

Kullanım: Analiz işlemini gerçekleştirmek için aşağıdaki komutu kolayca çalıştırabilirsiniz: yontu lint metin.md

yontu: Türkçe metinler için linter

yontu, Türkçe metindeki yazım hatalarını, noktalama sorunlarını ve yapay zekâ izlerini bulan
bir linterdir.

Kullanım: `yontu lint metin.md`

Kept: the name, the three things it finds, the command. Dropped: the adjective triad, the exclamation, "kolayca". The title moved to sentence case. Nothing was added.

Quick start

Claude Code (skill, MCP server, hooks)

/plugin marketplace add senrecep/yontu
/plugin install yontu@yontu

This is the only install path that also adds hooks. Commands: /yontu:turkcelestir, /yontu:yontu-lint.

Any agent that reads skill folders

npx skills add senrecep/yontu

This copies the skill only: no MCP server, no hooks.

CLI

pipx install yontu          # puts yontu and trlint on PATH
uvx yontu lint metin.md     # no install

From a checkout, without installing: python3 skills/yontu/scripts/yontu lint metin.md.

MCP server

claude mcp add yontu -- uvx yontu mcp

Other clients:

{
  "mcpServers": {
    "yontu": { "command": "uvx", "args": ["yontu", "mcp"] }
  }
}

Per agent

Config files, exact paths and what was tested on which version: integrations/.

Install profiles

install.sh is a dry run without --apply and writes only the skill folder. In Claude Code, install one of the two plugins, not both.

Context cost: about 150 tokens always on for the full plugin before its MCP tools, about 130 for lite, and about 3,900 tokens for a whole blog rewrite. Numbers, method and what loads when: docs/CONTEXT.md.

Tools

The server lists the first three tools by default. Set YONTU_MCP_TOOLS=full in its env to list all eight. It answers the other five either way, but most clients show the model only listed tools.

Prompts: turkcelestir (translate or adapt into Turkish), yogunlastir (tighten a text).

CLI

yontu and trlint are the same command.

yontu lint FILE [--register SLUG] [--tur ID] [--layer tdk,ai-iz] [--min-level warning] [--format json]
yontu lint FILE --fix [--write]     # safe tdk fixes only; --write saves the file
yontu compare OLD NEW               # fidelity gate
yontu stats FILE                    # sentence length, Ateşman readability, density
yontu rules [--layer tdk]           # list rules
yontu explain AI-E6                 # one rule with source and examples
yontu profiles [ID]                 # genres and registers

FILE can be - for standard input. lint and compare exit with 1 when they report an error, and with 2 on a usage error.

Without --register or --tur, yontu guesses the register from the content and falls back to notr, where every style rule is info.

Project settings go in .yontu.json (keys register, tur, enable, rules). Silence a rule in place with <!-- yontu-disable-next-line KS-01 -->, or a range with <!-- yontu-disable KS-01 --> and <!-- yontu-enable KS-01 -->.

Profiles

--register picks the register, --tur the genre. A genre implies its register.

hukuk and edebi get the tdk layer and the fidelity gate only. The skill does not rewrite them unless asked.

Rule layers

Every tdk rule names the TDK Yazım Kılavuzu section or Güncel Türkçe Sözlük entry it comes from. The links are in skills/yontu/references/kaynaklar.md. Autofix runs only on deterministic rules with near-zero false positives (herşey → "her şey", Türkçe'de → "Türkçede"). Dictionary redirects are never autofixed.

Every ai-iz rule carries an evidence label: veri (measured on a corpus), hakem (agreed by reviewers) or hipotez (observed, not measured). A hipotez rule never goes above info, and a hakem rule never goes above warning. The v0.1 rule set has no veri rules yet; see ROADMAP.md.

Run yontu rules for the full list and yontu explain ID for any rule.

Limitations

• Not an AI detector. yontu reports patterns. A finding does not mean a text was machine-written, and a clean report does not mean it was not. Passing a detector is not a goal.
• Heuristics. The ai-iz, stil and yogunluk rules are regular expressions and counts over surface forms, with no morphological analysis. Expect false positives, mostly at info.
• Uncalibrated thresholds. Sentence bands and density thresholds come from the literature and from review, not from a human baseline corpus. Until calibration they stay at info or warning.
• No dictionary. The linter catches known misspellings, not every misspelling.

Contributing

See CONTRIBUTING.md. New rules need a source, an evidence label and fixtures. Turkish examples of clean human writing in any genre are the most useful contribution: they are what calibration needs.

Credits

• Türk Dil Kurumu: Yazım Kuralları and Güncel Türkçe Sözlük.
• turkish-humanify by Oguzhan Durmaz (MIT) and turkce-humanizer by bushrabeg (MIT), whose ideas were rewritten and checked against TDK.
• Wikipedia's "Signs of AI writing", for the structure of the pattern catalog.
• Full source list: skills/yontu/references/kaynaklar.md.

License

MIT. See LICENSE.