yucezerey/turkish-anti-ai-skill

What it does

Large language models tend to produce Turkish that reads like translated English: English sentence architecture under Turkish words, right-branching modifiers, mechanical ve where converbs should appear, and every sentence ending in -DIr / -mAktAdIr. This project packages rules for that problem as an Agent Skill plus a deterministic Python detector.

It operates in three modes: rewrite repairs existing text while forbidding invented facts, numbers, or anecdotes; write generates Turkish from scratch; and audit reports AI/Turkish-language-defect findings without changing the text. Five registers (blog/essay, technical, corporate/SaaS, academic, literary) are calibrated against style dimensions, so a KVKK-style text is not given inverted literary sentences and a blog post is not forced into bureaucratic endings.

Install and use

The universal path for Claude Code, Cursor, Codex, Gemini CLI, Windsurf, and OpenCode is:

npx skills add yucezerey/turkish-anti-ai-skill

GitHub CLI users (v2.90+) can install with gh skill install yucezerey/turkish-anti-ai-skill. Claude Code also supports plugin marketplace installation via /plugin marketplace add yucezerey/turkish-anti-ai-skill and /plugin install turkish-anti-ai-skill@yucezerey, or a direct clone into ~/.claude/skills/ (installation options).

After installing, address the agent in natural language and the skill triggers from its own description, for example: "Şu blog yazısını insanlaştır ve AI kokusunu sil: <metin>" for rewrite mode, or "Bu kurumsal metni değiştirme, sadece içindeki AI-Türkçesi ve TDK hatalarını raporla" for audit mode (usage).

CLI detector

Beyond the agent skill, the repository includes a standalone scanner that needs no external libraries. Run it against a Markdown or text file:

python3 scripts/detect_tr_ai.py ornek-metin.md
python3 scripts/detect_tr_ai.py ornek-metin.md --json

It measures AI clichés, TDK errors, -mAktAdIr density, and sentence-length variation (Burstiness CoV), so the same check can run locally or in CI (detector section). scripts/build_dist.py rebuilds the single-file dist/turkish-anti-ai-skill.full.md for web interfaces that cannot load a folder.

Using it in other tools

Cursor can consume the repository's .cursor/rules/*.mdc rule, and Codex, Gemini CLI, Windsurf, and other agents can read the universal ~/.agents/skills/ directory or the dedicated Gemini path. For ChatGPT Custom GPT, Claude Projects, and Google AI Studio, where folders cannot be uploaded, copy the concatenated dist/turkish-anti-ai-skill.full.md into Project Instructions, System Prompt, or Knowledge (web setup).

Constraints

• The skill format follows the Agent Skills specification, so target clients must support that format or one of the provided adapters (AGENTS.md, CLAUDE.md, GEMINI.md, Cursor rules).
• Register calibration is a rule set, not a translator: it does not promise to preserve meaning beyond the explicit anti-fabrication lock, and the README does not publish accuracy figures for the evals.
• The CLI is Python 3 with no third-party dependencies, but it is a detector and does not rewrite text by itself.
• MIT licensed.

HotRepo's take

On the new Python list for 2026-11-06, the repository recorded 116 stars and 8 forks in a window starting 2026-11-05 20:08 UTC and observed through 2026-11-06 06:45 UTC, with no all-language daily rank recorded. That is a single snapshot, so it supports only the fact of an early spike, not a trend. What stands out technically is the split between a rule-based agent skill and an independently runnable detector, which makes the project useful even outside chat clients.