Humanize AI Text Skill: Detect and Rewrite the 16 Patterns That Trigger AI Detectors

23,000+ downloads and 97 stars — humanize-ai-text is one of the most practically-used writing skills on ClawHub. It's a Python-based detection and transformation system built on a rigorous foundation: Wikipedia's comprehensive "Signs of AI Writing" guide, which catalogs the linguistic patterns that distinguish AI-generated text from human writing.

The skill doesn't just do surface-level paraphrasing. It has a taxonomy of 16 specific pattern categories, from citation artifacts to overused vocabulary to structural habits, and applies targeted transformations to each.

The Problem It Solves

AI writing detectors have become standard in education, publishing, and corporate contexts. GPTZero, Turnitin, Originality.ai, and similar tools scan for the statistical fingerprints of language models. These fingerprints are more specific than most people realize — it's not just that AI sounds "formal" or "robotic," it's that it uses particular phrases, structures, and patterns with unnatural frequency.

If you're using Claude for first drafts and want them to survive detector scrutiny, or if you've received AI-generated text from a collaborator that needs editing, you need to know exactly what to change. humanize-ai-text makes this systematic.

The 16 Pattern Categories

The detection system is organized into four tiers by signal strength:

Critical (Immediate AI Detection)

These are automatic red flags for any detector. The citation bugs in particular are ChatGPT artifacts that appear in copy-pasted ChatGPT output and are instantly identifiable.

High Signal

The AI vocabulary list is surprisingly specific. "Delve" in particular is a strong Claude signature — it appears at roughly 10x the frequency in AI text vs. human text.

Medium Signal

Style Signal

The Three Scripts

detect.py — Scan for AI Patterns

# Scan a file
python scripts/detect.py essay.txt
 
# JSON output (for piping)
python scripts/detect.py essay.txt -j
 
# Score only
python scripts/detect.py essay.txt -s
 
# Stdin
echo "This delves into the pivotal landscape..." | python scripts/detect.py

Output includes:

• Total issue count by category
• AI probability rating (low / medium / high / very high)
• Which patterns are auto-fixable vs. requiring manual revision

transform.py — Auto-Rewrite

# Transform file in-place
python scripts/transform.py essay.txt
 
# Write to new file
python scripts/transform.py essay.txt -o output.txt
 
# Aggressive mode (more changes)
python scripts/transform.py essay.txt -a
 
# Quiet mode (no progress output)
python scripts/transform.py essay.txt -q

What it auto-fixes:

• Citation bugs → removed
• Markdown formatting → stripped to plain text
• Chatbot sentences → removed or rewritten
• Copula avoidance → "is/has" substituted
• Common filler phrases → simplified
• Curly quotes → straight quotes

What requires manual review:

• AI vocabulary (context-dependent replacement)
• Structural patterns (depends on document purpose)
• Promotional language (requires knowing the intended register)

compare.py — Before/After Diff

# Show what changed
python scripts/compare.py text.txt -o clean.txt

The compare script shows exactly what the transformation changed, which is useful for understanding the pattern and learning to avoid it in future drafts.

Practical Workflow

# Step 1: Detect what's in the text
python scripts/detect.py my-draft.txt
 
# Step 2: Auto-fix the obvious patterns
python scripts/transform.py my-draft.txt -o clean-draft.txt
 
# Step 3: Review what changed
python scripts/compare.py my-draft.txt -o clean-draft.txt
 
# Step 4: Manually fix any remaining high-signal patterns
# (AI vocabulary, structural issues — requires human judgment)

How to Install

clawhub install humanize-ai-text

Practical Tips

1. Run detect first, don't just transform — Understanding what's flagged before auto-fixing helps you catch patterns in your drafting process and avoid them in the future.

2. Aggressive mode for heavy AI text — Default transform is conservative. Use -a for content that's heavily AI-flavored (multiple high-signal categories).

3. AI vocabulary is the hardest to auto-fix — Words like "delve," "tapestry," and "underscore" require context-appropriate substitution. The script flags them; you need to replace them manually with words that fit your actual sentence.

4. Don't forget structural signals — "Despite these challenges" as a section header and "In conclusion" as a opener are clichés that detectors recognize. These won't always auto-fix; they need structural rethinking.

5. Use JSON output for batch processing — For reviewing multiple documents, -j output lets you script comparison across files and prioritize the highest-risk documents first.

Considerations

• Auto-fixes may change meaning — Copula replacement ("serves as" → "is") is usually fine but occasionally loses nuance. Review the diff after transformation.
• Not a guarantee against detection — Detectors evolve. This skill is based on currently-known patterns; new model versions may introduce new signatures.
• Pattern categories have some overlap — The same sentence might trigger multiple categories. The probability score accounts for this, but don't double-count individual fixes.
• Aggressive mode can over-edit — -a removes more patterns but may make the text feel stilted if overdone. Use conservatively on otherwise well-written text.

The Bigger Picture

humanize-ai-text makes explicit what was previously intuitive: there is a specific, enumerable set of patterns that distinguish AI writing from human writing. The Wikipedia "Signs of AI Writing" guide that this skill is based on is a living document, updated as researchers and editors catalog new patterns.

The skill turns that taxonomy into a practical tool — not just a list of things to avoid, but a scanner that finds them and a transformer that fixes the mechanical ones. At 23,000+ downloads, it's found its audience among writers, editors, and content professionals who use AI as a drafting tool and need reliable output.

View the skill on ClawHub: humanize-ai-text