avoid-ai-writing — Multilingual Skills

Language-specific versions of the avoid-ai-writing skill by Conor Bronsdon (MIT licence), extended to languages beyond English.

Each SKILL-XX.md file is grounded in statistically documented, language-native AI writing patterns — not translated from English. Every flagged word or phrase must be traceable to a real source showing it is over-represented in LLM output for that specific language.

Available Languages

Planned: Spanish (SKILL-ES.md) — PRs welcome.

What a skill does

You give it text. It:

1. Flags every AI writing pattern it finds (severity P0 / P1 / P2)
2. Rewrites the text with those patterns removed (default mode)
3. Runs a second pass on its own rewrite to catch patterns it introduced
4. Optionally runs in detect-only mode — flags without rewriting

Patterns covered include: chatbot openers, generic conclusions, semantic inflation, synonym chains, structural uniformity, excessive hedging, EU/management jargon substituting for content, and language-specific statistical tells (e.g. calques from English in Romanian, Nominalstil clusters in German).

How to Install

A · Claude Code (CLI and VS Code / JetBrains extensions)

Skills live in .md files. The name: field in the frontmatter becomes the slash command.

Global install — available in every project:

# macOS / Linux
mkdir -p ~/.claude/skills
cp SKILL-DE.md ~/.claude/skills/avoid-ai-writing-de.md
cp SKILL-RO.md ~/.claude/skills/avoid-ai-writing-ro.md
cp SKILL-IT.md ~/.claude/skills/avoid-ai-writing-it.md
cp SKILL-FR.md ~/.claude/skills/avoid-ai-writing-fr.md
cp SKILL-SV.md ~/.claude/skills/avoid-ai-writing-sv.md

# Windows (PowerShell)
New-Item -ItemType Directory -Force "$env:USERPROFILE\.claude\skills"
Copy-Item SKILL-DE.md "$env:USERPROFILE\.claude\skills\avoid-ai-writing-de.md"
Copy-Item SKILL-RO.md "$env:USERPROFILE\.claude\skills\avoid-ai-writing-ro.md"
Copy-Item SKILL-IT.md "$env:USERPROFILE\.claude\skills\avoid-ai-writing-it.md"
Copy-Item SKILL-FR.md "$env:USERPROFILE\.claude\skills\avoid-ai-writing-fr.md"
Copy-Item SKILL-SV.md "$env:USERPROFILE\.claude\skills\avoid-ai-writing-sv.md"

Project install — available only in the current repository:

mkdir -p .claude/skills
cp SKILL-DE.md .claude/skills/
cp SKILL-RO.md .claude/skills/
cp SKILL-IT.md .claude/skills/
cp SKILL-FR.md .claude/skills/
cp SKILL-SV.md .claude/skills/

After installing, invoke in any Claude Code conversation:

/avoid-ai-writing-de
/avoid-ai-writing-ro
/avoid-ai-writing-it
/avoid-ai-writing-fr
/avoid-ai-writing-sv

B · Claude.ai Web — Projects (recommended for web use)

Claude Projects let you attach a persistent system prompt to a conversation thread.

1. Open claude.ai → Projects → New Project
2. Click Project Instructions
3. Paste the full content of SKILL-DE.md or SKILL-RO.md
4. Name the project (e.g. "AI Writing Reviewer — German")
5. Every new conversation inside this project applies the skill automatically

Tip: Create one project per language. You can switch between them from the sidebar.

C · ChatGPT Web — Custom GPTs

1. Go to chatgpt.com → Explore GPTs → Create
2. Switch to the Configure tab
3. Paste the full content of SKILL-XX.md into the Instructions field
4. Set a name, e.g. "AI Writing Detector – German"
5. Save → Only me (private) or share the link

The YAML frontmatter block at the top of the file is harmless — ChatGPT will treat it as plain text context. You can remove the --- block if you prefer a clean paste.

D · Cursor

As a project rule (recommended):

1. Create the file .cursor/rules/avoid-ai-writing-de.mdc in your project
2. Add this header, then paste the skill content below it:

   ---
   description: Detect and remove AI writing patterns in German text
   globs:
   alwaysApply: false
   ---
   

3. Strip the original YAML frontmatter from the pasted content (Cursor has its own)

As a global rule:
Cursor Settings → Rules for AI → paste the skill content.

E · GitHub Copilot / Codex (VS Code)

1. Create .github/copilot-instructions.md in your repository
2. Paste the skill content there

Copilot will apply these instructions to all Copilot Chat conversations in that workspace. For per-file context, add the skill as a referenced file in a Copilot Chat prompt:

#file:.claude/skills/avoid-ai-writing-de.md

Review this text: [your text]

F · Windsurf, Zed, or any other AI editor

Paste the skill content into the system prompt / custom instructions field of your AI assistant configuration. Most editors expose this as "AI Instructions", "System Prompt", or "Custom Rules".

G · Ad-hoc (any chat interface)

Copy the full file content and paste it at the start of your conversation, then add your text:

[paste full SKILL-XX.md content]

---

Now review the following text:

[your text here]

How to Use

In Claude Code

You can also add a context hint to adjust tolerance:

/avoid-ai-writing-de linkedin

[text]

Available context profiles per language:

In web tools (Claude Projects, Custom GPTs)

Just paste your text. No slash commands needed — the system prompt is always active.

To switch modes:

• Rewrite (default): paste text and send
• Detect only: start your message with "nur prüfen:" (DE), "doar marcare:" (RO) or "bara markera:" (SV)
• Context hint: start with "context: academic" or the equivalent

How to Add a New Language

The one-line prompt

Paste this into Claude Code (or any AI with access to this repository):

Create SKILL-ES.md for Spanish. Follow the process in CLAUDE.md exactly.
Research sources first — do NOT write from memory or translate from English.
Search in Spanish, not in English. Use SKILL-DE.md as the structural reference.

Replace ES / Spanish with your target language and ISO 639-1 code.

What the process produces

The process defined in CLAUDE.md requires:

1. Web research in the target language — at least 3 independent source types:

   • Wikipedia community page for the target language
   • Native SEO / content blogs on AI text detection
   • Academic papers on AI text detection in multilingual settings
   • Practitioner reports (editors, journalists, content agencies)

2. Language-native vocabulary table (Tier 1 / 2 / 3) — no pattern is included unless a source shows it is statistically over-represented in LLM output for that language

3. 40+ documented patterns — structured as: pattern name · severity · before/after example

4. A culturally authentic full example — drawn from the context where AI text appears most densely in that language (e.g. LinkedIn for German, NGO press releases for Romanian, management memos for French)

5. sources/XX-sources.md — all citations with URLs, dates, and which patterns they support; gaps explicitly noted

Quality checklist (before opening a PR)

• All Tier-1 words have a cited source
• No Tier-1 word is just "bad style" — it must be a documented LLM statistical tell
• The full example is language-native, not a translation
• Context profiles match real text types of that language region
• Tolerance matrix is filled in
• Severity grades (P0/P1/P2) are assigned in all pattern tables
• Version is set to 1.0.0

Project Structure

/
├── SKILL-DE.md             ← German (v2.1.0, source-verified)
├── SKILL-RO.md             ← Romanian (v1.0.0, source-verified)
├── SKILL-IT.md             ← Italian (v1.0.0, source-verified)
├── SKILL-FR.md             ← French (v1.0.0, source-verified)
├── SKILL-SV.md             ← Swedish (v1.0.0, source-verified)
├── CLAUDE.md               ← Authoring instructions (how to create new skills)
├── README.md               ← This file
└── sources/
    ├── DE-sources.md       ← Citations for German patterns
    ├── RO-sources.md       ← Citations for Romanian patterns
    ├── IT-sources.md       ← Citations for Italian patterns
    ├── FR-sources.md       ← Citations for French patterns
    └── SV-sources.md       ← Citations for Swedish patterns

Design Principles

• Statistically grounded, not opinionated. Patterns are included only when sources show they are over-represented in LLM output for that language. General bad style is not a KI tell.
• Language-native. English patterns are not translated. Each language has its own statistical anomalies (e.g. Romanian calques from English, German Nominalstil clusters, French management jargon and impersonal distancing formulas).
• Context-aware. Technical writing, academic prose, and legal documents have different baselines. The skill adjusts its strictness by context profile.
• Severity-graded. P0 patterns are always removed. P1 patterns are usually removed. P2 patterns are evaluated in context.

License

MIT — based on avoid-ai-writing (v3.4) by Conor Bronsdon.
Multilingual adaptations by Jürgen Kraus.