Two files. One teaches AI how you actually write. The other checks if it learned.
English · العربية
What's Inside · Quick Start · The 8 Angles · Install · Why
Built for writers who use AI but don't want to sound like it. Built especially for Arabic speakers moving between languages, dialects, and registers that AI flattens into generic English.
nafas.md — The voice framework.
A guided process for collecting your writing samples and identifying your patterns. Covers rhythm, metaphor, emotional register, code-switching. Includes a kill list of 100+ AI vocabulary words backed by academic research.
ghirbal.md — The sieve.
An 8-angle review protocol modeled on the Damascene star (النجمة الشامية). Pass any AI-generated text through it. Each angle checks a different dimension: voice, rhythm, body, purity, substance, structure, soul, ground. Rates output as ردّه (send it back), قريب (close), or تمام (ships).
Step 1: Build your nafas.
Open nafas.md. Follow the prompts. Collect 5,000+ words of your actual writing (emails, posts, WhatsApp messages, articles). Fill in every section. Save the result.
Step 2: Write something. Give your AI the completed nafas file as context. Ask it to draft whatever you need.
Step 3: Run the ghirbal.
Pass the draft through ghirbal.md. The 8 angles will catch what's off. Fix it, run again. 3 passes max.
النجمة الشامية — The Damascene Star
| # | العربية | English | What it checks |
|---|---|---|---|
| 1 | الصوت | Voice | Does it sound like the person? |
| 2 | الإيقاع | Rhythm | Alive or metronomic? |
| 3 | الجسد | Body | Grounded or floating? |
| 4 | النقاء | Purity | Clean of AI contamination? |
| 5 | المادة | Substance | Does it actually say something? |
| 6 | البنية | Structure | Does the architecture feel human? |
| 7 | الروح | Soul | Is there a human in this text? |
| 8 | الأرض | Ground | Rooted in lived experience? |
Copy both files into your commands directory:
cp nafas.md ~/.claude/commands/nafas.md
cp ghirbal.md ~/.claude/commands/ghirbal.mdThen invoke with /nafas to build your voice file, /ghirbal to review output.
For the voice file itself, once you've built it using the nafas framework, save it where it loads automatically as context:
cp my-nafas.md ~/.claude/rules/common/my-nafas.mdPaste the contents of nafas.md into a custom GPT's instructions, or upload both files as knowledge files. Use "Run the ghirbal on this" as a prompt to trigger the review.
Both files work as system prompts or context files. Feed your completed nafas file as persistent context. Run the ghirbal as a separate review pass (ideally a separate agent or conversation, so it grades independently).
AI writes like a committee. Technically correct and completely empty. It avoids risk, hedges every claim, and produces text that could have been written by anyone about anything. The voice disappears.
For Arabic speakers, the problem is worse. AI flattens diglossia (the spectrum between dialect and fusha), drops code-switching patterns that carry identity, and strips out the oral tradition rhythms that shape how many of us actually write.
Nafas and the ghirbal are tools for pushing back. Build your voice reference. Then hold the AI accountable to it.
The banned vocabulary and structural patterns are backed by peer-reviewed research and large-scale corpus analysis:
- Words flagged as 16x-182x overrepresented in AI text vs human writing (GPTZero, 3.3M text corpus)
- Vocabulary spikes of 59-137% in post-ChatGPT scientific papers (Gray, Nature 2024)
- Trigram patterns up to 85,000x overrepresented (Paech et al., ICLR 2026)
- Syntactic template repetition analysis (Northeastern University, 2024)
- Burstiness and perplexity benchmarks from GPTZero and Originality.ai
- Wikipedia's "Signs of AI Writing" guide (15,000 words, maintained by WikiProject AI Cleanup)
- Stylometric analysis confirming AI cannot fully replicate human writing variability (University College Cork, 2025)
If you've found AI vocabulary words, phrase patterns, or structural tells that aren't in the kill list, open a PR. Include the source and the overrepresentation data if you have it.
If you've built a nafas file for a language other than Arabic or English and want to share the language-specific guidance, that's welcome too.
This work didn't come from nowhere. The kill list, the detection heuristics, and the structural patterns all build on research and tools that other people made public:
Antislop by Sam Paech — the deepest slop analysis I've seen. Trigram patterns overrepresented up to 85,000x in AI text. The antislop sampler catches patterns at generation time; we catch them at review time. Same enemy, different angle.
EQ-Bench Creative Writing Bench and the Slop Score — the methodology for weighting diagnostic patterns. Their finding that "not just X but Y" carries 25% diagnostic value is why we treat it as fatal.
GPTZero — the burstiness and perplexity research, and the 3.3M text corpus that gave us the overrepresentation multipliers. When we say "delve" is 50x overrepresented, that number comes from their work.
Wikipedia: Signs of AI Writing — a 15,000-word community guide maintained by WikiProject AI Cleanup. Volunteer editors cleaning AI contamination out of an encyclopedia, one article at a time. They wrote the field guide.
Andrew Gray — "Chatbot Influence on Scientific Writing" (Nature, 2024). The paper that proved "meticulous," "commendable," and "intricate" spiked 59-137% in scientific literature after ChatGPT.
Justin Blackman — Brand Ventriloquism. A professional ghostwriter who figured out that the most useful question you can ask a client is "mark every word you'd never say." The negative list in nafas comes from that insight.
If nafas shapes how your project writes, you can add this badge to your README:
[](https://github.com/karam2022/nafas)
MIT
The goal is not perfect text. The goal is text that breathes.