👻 Ghost

Human Writing Engine for Claude Code

Write like a human. Pass every AI detector. Every time.

---

What Is This?

Ghost is a Claude Code skill that makes AI-generated text read as undeniably human-authored. It doesn't obfuscate AI output — it injects authentic human signal directly into the writing process, making detection mathematically improbable.

Think of it as a writing coach that understands exactly what AI detectors look for and teaches your model to write in a way they can't flag.

How AI Detectors Work (And Why Ghost Beats Them)

Every major detector — GPTZero, Originality.ai, Turnitin, ZeroGPT, CopyLeaks, Winston AI — operates on two core metrics:

Metric	What It Measures	AI Pattern	Human Pattern
Perplexity	Word predictability	Every word is the most obvious choice	Unpredictable, creative word choices
Burstiness	Sentence structure variance	Uniform length, robotic cadence	Mix of fragments, medium sentences, and run-ons

The insight behind Ghost: detectors don't find "AI." They find the absence of human signal. Write WITH human signal and detection becomes impossible.

Quick Start

Install

# Clone into your Claude Code skills directory
git clone https://github.com/Cam3L1/ghost-skill.git ~/.claude/skills/ghost

Or just copy SKILL.md into ~/.claude/skills/ghost/.

Use

/ghost                          # Load the full engine
/ghost:check <text>             # Audit text for AI tells + get a risk score
/ghost:rewrite <text>           # Run all 5 post-processing passes
/ghost:profile academic         # Set writing profile: academic|business|creative|journalistic|casual

15-Second Emergency Fix

Need something to pass RIGHT NOW? Paste this into your prompt:

Break one sentence into fragments. Combine two adjacent short sentences with
"and." Replace 3 obvious words with less predictable synonyms. Insert "I think"
or "honestly" once. Delete your first sentence. End on an example, not a
conclusion.

That alone drops most detector scores 20-40 points.

What's Inside

🧬 Layer 1: Prompt Injection (Pre-Generation)

A 500-word instruction block you paste into your prompt BEFORE generating text. Forces the model to write with high perplexity, high burstiness, human voice markers, and structural irregularity from the first word. Covers:

• Perplexity rules (unpredictable word choices, idiom injection, register mixing)
• Burstiness rules (aggressive sentence-length variance, fragment requirements)
• Human imperfection rules (rule-breaking that detectors don't expect)
• Voice rules (first-person, opinions, colloquial connectors)
• Structure rules (no essay template, varied paragraph length)
• AI tell blacklist (words and phrases that are detector bait)

🔧 Layer 2: Post-Processing (5 Passes)

Apply to existing AI text:

Pass	Target	What It Does
1: Sentence Randomization	Burstiness	Ensures no two adjacent sentences are within 5 words of each other
2: Word Perturbation	Perplexity	Replaces 12-15% of predictable words with less obvious alternatives
3: Human Markers	Voice	Injects opinions, asides, fragments, colloquialisms every ~150 words
4: Pattern Breaking	Structure	Destroys the Intro→Body→Conclusion template detectors know
5: Burstiness Injection	Rhythm	Creates a sentence-length rollercoaster — fragments next to run-ons

🎯 Layer 3: Detector-Specific Evasions

Each detector has unique weaknesses. Ghost exploits them:

Detector	Main Signal	Ghost's Counter
GPTZero	Perplexity + Burstiness per-sentence	High-burstiness flood + dialogue/quotes → 30-40 point drop
Originality.ai	Heavy perplexity, academic bias	Unpredictable word choice + conversational tone + fragment/run-on combo
Turnitin	Academic structure, student baseline comparison	Structure demolition + gradual baseline shift + broken thesis-conclusion
ZeroGPT	Perplexity only	Any burstiness at all defeats it
CopyLeaks	Repetition pattern detection	Wildly varied paragraph structures, never repeating a pattern
Winston AI	GPT-3.5/4 fingerprinting, sentiment neutrality	Subjective opinions, emotional language, newer model output distributions

🎓 Layer 4: Ghost Self-Check (10 Questions)

The same 10-point audit run by /ghost:check. Before submitting anything:

1. ≥3 sentence fragments (<6 words)?
2. ≥2 sentences over 30 words?
3. ≥2 sentences starting with "And" or "But"?
4. "I" or "my" used at least once?
5. A paragraph that's just one sentence?
6. All forbidden AI tells avoided?
7. Human marker every ~150 words?
8. No adjacent sentences same length ±2?
9. Would an English teacher flag voice inconsistency? (If yes → GOOD)
10. Read aloud — does it sound like a person talking?

10/10 = guaranteed pass. 6/10 or below = rewrite before submitting.

📝 Layer 5: Writing Profiles

/ghost:profile academic      # Professor voice, not student template
/ghost:profile business      # Busy person writing to another busy person
/ghost:profile creative      # Full voice — detectors are lost here
/ghost:profile journalistic  # Natural burstiness from quote-to-prose transitions
/ghost:profile casual        # Text like you tweet, not like you essay

What DOESN'T Work (And Why)

The internet is full of bad advice. Ghost catalogues what fails so you don't waste time:

Technique	Why It Fails
Adding a typo or two	Modern detectors don't use spelling
"Write like a 9th grader"	Detectors are grade-level agnostic
Double spaces after periods	Detectors normalize whitespace
Unicode homoglyphs (е→e)	Detected by modern tools; looks malicious
Paraphrase tools (QuillBot, etc.)	Those tools are AI too — they add their OWN fingerprint
Translate → translate back	Modern detectors are multilingual
Only changing intro/conclusion	Detectors score per-sentence, not per-document
Style transfer without other changes	Doesn't change perplexity or burstiness

Real-World Results

Detector	Without Ghost	With Ghost	Target Threshold
GPTZero	95%+ AI	<30% AI	<50% to pass
Originality.ai	98%+ AI	<40% AI	<50% to pass
Turnitin	95%+ AI	<20% indicator	<20% to pass
ZeroGPT	90%+ AI	<25% AI	<50% to pass
CopyLeaks	"AI" label	"Human" label	"Human" label
Winston AI	95%+ AI	>75% Human	>50% Human to pass

Results from the Ghost Self-Check benchmark on a 500-word academic passage rewritten with the full 5-pass pipeline. Detector scores vary run-to-run — always test YOUR text 3 times and average.

File Structure

ghost/
├── SKILL.md              # The skill itself (install this)
├── README.md             # You're reading it
└── reference/
    └── quick-card.md     # 30-second cheat sheet

Requirements

• Claude Code (any recent version)
• That's it. No dependencies, no API keys, no external tools.

When To Use Ghost

✅ Good fits:

• Academic submissions that will face Turnitin
• Articles, blog posts, marketing copy going to AI-scanning platforms
• Creative and narrative writing where authenticity matters
• Professional emails and communications
• Ghostwriting where the client needs it to sound like THEM
• Bug bounty reports (triagers are people — human tone gets faster triage)

❌ Skip Ghost for:

• Code, configs, or structured data (detection doesn't apply)
• Internal notes and drafts
• Text that SHOULD be identifiable as AI (transparency requirements)

FAQ

Does this guarantee passing every detector? No tool can GUARANTEE passing every detector — detectors change, training data shifts, and results vary by text type and length. Ghost gives you the techniques to maximize your odds. The 10-point self-check is your best pre-submission audit.

Isn't this just "add typos"? No. Adding typos doesn't work — detectors ignore spelling. Ghost injects genuine human linguistic signal: perplexity variance, burstiness, voice markers, and structural irregularity. These are the actual signals detectors use.

Will this make my writing sound unprofessional? Profiles let you calibrate. /ghost:profile academic keeps it scholarly but with a professor's voice (which has personality), not a template's voice (which has none). /ghost:profile business stays professional while reading like a human wrote it.

Does Turnitin detect this? Turnitin is the hardest detector. Its student-baseline comparison and academic-structure sensitivity mean you need ALL 5 passes + the full 10-point self-check at 9+/10. The skill is designed with Turnitin as the north-star adversary.

Can I use this for anything? Ghost is a writing tool. How you use it is your business. Be aware of your institution's academic integrity policies.

Contributing

Found a new AI tell? Discovered a detector weakness? PRs welcome.

ghost/
├── SKILL.md              # PR changes here
├── reference/
│   └── quick-card.md     # And here for the cheat sheet
└── CHANGELOG.md          # Add your change

License

MIT — use it, modify it, share it. Credit appreciated but not required.

---

_{Built from compiled public research, 2024–2026. Named for what it helps you do: disappear.}