skill

turkish-native

Agent skill for writing native Turkish and removing translationese from AI and localized copy.

oguzhankayan/turkish-nativecommunitylistedunreviewednot running here yet

Use skill

Set up

Skills are installed by placing their folder where your agent loads skills.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below; one-click install arrives with accounts.

README

synced Oct 17, 2026

Turkish Native

Write Turkish as Turkish, not translated English.

Türkçe README · Skill · Examples · Release · Contributing · Roadmap

Turkish Native is an open-source agent skill for writing and editing natural Turkish. It targets a specific failure mode in AI-generated and localized copy: Turkish that is grammatically understandable but still carries foreign sentence structure, unnatural collocations, literal SaaS or sales calques, internal product jargon, or model-written marketing formulas.

A sentence can be grammatically correct and still sound translated.

Why this exists

Spell checkers and grammar tools catch mechanical errors. Turkish Native works one layer above them: native construction. It asks an agent to rebuild the sentence from meaning instead of preserving the source-language skeleton.

That means checking syntax before vocabulary: subordinate clauses, nominalization, word order, case suffixes, possessives, verb–noun pairings, product language, register, factual fidelity, and page-level domain consistency.

Examples

Artificial / translated	Natural Turkish
`Kim açtı, kim düzenledi gör.`	`Kimin açtığını, kimin düzenlediğini gör.`
`Sayfayı WhatsApp'ta gönder.`	`Sayfayı WhatsApp'tan gönder.`
`16 işletme bulundu, 12'si telefonlu.`	`16 işletme bulundu, 12 telefon numarası bulundu.`
`Aynı işletme için tekrar kayıt oluşturulmaz.`	`Aynı işletme ikinci kez eklenmez.`
`Kart istemiyoruz.`	`Kredi kartı gerekmez.`
`Buket için randevu almak istiyorum.`	`Buket siparişi vermek istiyorum.`

The skill also protects against overcorrection. Natural phrases such as Sen satışa bak., Bulamazsa uydurmaz., Site kurmak, or Linki gönder. should stay natural when the context supports them.

Scope

Use Turkish Native for marketing, product/UI copy, SaaS, localization, documentation, support, legal/privacy/security prose, institutional communication, generated business pages, and reviews of AI-written Turkish.

It is not a spell checker, AI detector, loanword ban list, or replacement for professional legal review. It does not claim that Turkish has only one correct phrasing.

Installation

Skills CLI

npx skills add oguzhankayan/turkish-native --global

Claude Code plugin

/plugin marketplace add oguzhankayan/turkish-native
/plugin install turkish-native@turkish-native

Manual

Copy SKILL.md and the bundled references/ directory into the skill directory used by your agent.

Usage

Use $turkish-native to review this Turkish copy.
Flag translationese and unnatural collocations, then rewrite it naturally.

Translate this into native Turkish with $turkish-native.
Preserve facts and product terminology, but do not preserve English sentence structure.

Use $turkish-native to rewrite the prose in src/content/tr/home.ts.
Keep keys, variables, URLs, code, and route names unchanged.

How it works

Turkish Native uses a meaning-first workflow: identify the audience and register, separate facts from wording, fix structural problems before isolated vocabulary, rebuild foreign sentence architecture, then check native flow, factual fidelity, domain fit, and over-editing.

The taxonomy contains 57 patterns in six groups:

1–10: sentence architecture,
11–20: verb choice and lexical fit,
21–30: product, UI, and localization,
31–40: marketing and model-written copy,
41–51: register, fidelity, and final checks,
52–57: domain and page integrity.

Detailed rules, examples, and false-positive guards live in references/. The compact runtime instructions live in SKILL.md.

Evaluation

The repository ships with 105 regression cases across evals/cases/, including positive controls that should remain unchanged. Every numbered pattern has evaluation coverage.

python3 -m pip install -r requirements-dev.txt
python3 scripts/validate-package.py
python3 scripts/validate-benchmark.py

The package validator checks package structure, continuous pattern numbering, version consistency, unique case IDs, positive-control count, and per-pattern eval coverage. The benchmark validator separately checks the held-out corpus schema, unique inputs, scoring dimensions, and pattern/group associations. For benchmark methodology, see docs/BENCHMARKING.md.

Project status

Current release: v0.5.0, pre-1.0. Pattern names and the evaluation schema may still change before 1.0.

See docs/PHILOSOPHY.md and ROADMAP.md for design rationale and next steps.

Relationship to Humanizer

blader/humanizer is a general-purpose skill for broad AI-writing patterns. Turkish Native solves a narrower Turkish-specific problem: translationese and non-native construction.

Turkish Native is not a fork and does not reuse Humanizer's pattern taxonomy. Humanizer inspired the packaging approach.

Contributing

Contributions are welcome. Read CONTRIBUTING.md before opening a PR.