deai

name	deai
description	Reduce AI detection rate in writing (Chinese & English). Use when de-AIing, humanizing, or auditing text for AI traces in any format (LaTeX, Typst, Markdown, plain text).

De-AI Writing Assistant (Chinese & English)

Detect and remove AI writing traces to reduce AI detection rate. Supports Chinese and English, multiple document formats, and both automated checking and LLM-guided rewriting.

Core Principle

AIGC detectors recognize sentence-structure fingerprints (skeleton repetition), not individual words. The same skeleton used 3+ times triggers detection far more than any single word choice. Fix skeletons first, vocabulary second.

Steps

1. Parse $ARGUMENTS to determine: target file/text, language (--lang zh|en|auto), mode (--mode check|rewrite|audit), and format.
2. If --lang auto or unspecified, auto-detect language by checking CJK character ratio.
3. Choose the appropriate workflow:

Mode: check (default) -- Automated Detection

1. Run the detection script:

   python3 $SKILL_DIR/scripts/deai_check.py <file> --analyze --lang <zh|en|auto>
   

2. Review the report: section-by-section density scores, prioritized trace listing.
3. Present findings to user with actionable suggestions.

Mode: rewrite -- LLM-Guided Rewriting

1. First run check to identify AI traces.
2. Read $SKILL_DIR/references/REWRITING_GUIDE.md for rewriting strategies.
3. Read the language-specific pattern file:
   • Chinese: $SKILL_DIR/references/PATTERNS_ZH.md
   • English: $SKILL_DIR/references/PATTERNS_EN.md
4. For each flagged section (highest density first): a. Apply skeleton-level transformations (vary sentence structure, break parallelisms). b. Apply vocabulary-level fixes (replace AI-typical words/phrases). c. Preserve all technical content, data, citations, and math.
5. Second-pass audit: Re-run detection on the rewritten text. If traces remain, fix them.
6. Output in diff-comment format showing original vs. revised with change explanations.

Mode: audit -- Two-Pass Full Audit

1. Run check to get baseline scores.
2. Perform rewrite on all flagged sections.
3. Re-run check on rewritten output.
4. Generate before/after comparison report with density score improvements.

Mode: style-profile -- Style Extraction

Analyze reference articles and output a JSON style profile. The profile captures 8 dimensions of writing style and can be used during rewrite mode to match the target author's voice.

1. Run the style profiler:

   python3 $SKILL_DIR/scripts/style_profile.py ref1.txt ref2.txt --output profile.json
   

2. The output JSON includes: sentence patterns (mean/std/CV), word choice (formality, preferred words), rhetoric, structure, narrative perspective, emotion, rhythm, and punctuation profile.
3. When using rewrite mode, optionally pass --style-profile profile.json to guide the LLM toward the target style.

Model Profile: --profile

Control which model-specific detection rules are loaded:

• generic (default): Universal patterns only, no model-specific tics
• deepseek: Adds DeepSeek-specific patterns (量子/赛博 buzzwords, 三-fixation, parenthetical overload, tree numbering)
• claude: Skips all model-specific patterns
• mixed: Loads all model-specific patterns

Detection Philosophy

What We Check (Priority Order)

1. Structural skeleton repetition (highest signal) -- same sentence pattern used 3+ times
2. Template expressions -- phrases that fit any paper/article in any field
3. Over-confident claims -- absolute assertions without evidence
4. Empty phrases -- vague adjectives replaceable by specific data
5. Vague quantification -- "many studies" without citations
6. AI vocabulary -- words that appear 5-20x more often in AI text

What We Preserve

• All technical content, data, and reasoning
• LaTeX/Typst commands, math environments, citations, labels
• Necessary causal connections and logical flow
• Chapter-level introductory paragraphs (style-guide recommended)
• Statistically qualified claims (with p-values, percentages)

Rewriting Rules

1. Zero fabrication: Never add data, metrics, claims, or citations
2. Skeleton over lexicon: Change sentence structure, not just swap synonyms
3. Content preservation: Technical content is sacred; only modify how it's expressed
4. Vary everything: No two adjacent paragraphs should share the same structure
5. Implicit causality: Delete connector words when context makes causality obvious
6. Specific over vague: Replace vague claims with concrete data when available; mark as [PENDING] if not

Output Format

% ============================================================
% De-AI Edit (Line X - [Section])
% ============================================================
% Original: [AI-trace text]
% Revised:  [Humanized text]
%
% Changes:
% 1. [Type]: [Details]
% 2. [Type]: [Details]
%
% Warning: [PENDING VERIFICATION] if any claims need evidence
% ============================================================

Change Types

1. skeleton_vary -- Changed sentence structure to break repetition
2. delete_empty -- Removed vague adjective/adverb
3. add_specific -- Replaced vague claim with concrete data
4. split_sentence -- Divided sentence >50 words (ZH) or >40 words (EN)
5. downgrade_claim -- Added academic hedging
6. delete_connector -- Removed redundant causal connector
7. replace_template -- Replaced template expression with specific content
8. replace_vocabulary -- Replaced AI-typical word with natural alternative
9. vary_parallelism -- Broke mechanical parallelism (firstly/secondly/finally)
10. merge_fragments -- Combined fragmented sentences for natural flow

Density Score Interpretation

Score	Chinese	English	Action
>15%	>10%	Critical	Rewrite section immediately
10-15%	5-10%	High	Rewrite soon
5-10%	3-5%	Medium	Review and revise
<5%	<3%	Low	Minor polish only

Script Reference

Script	Purpose
deai_check.py <file> --analyze	Full document analysis with density scores
deai_check.py <file> --section <name>	Check specific section
deai_check.py <file> --score	Section scores only
deai_check.py <file> --fix-suggestions	JSON suggestions for programmatic fixing
deai_check.py <file> --lang zh|en|mixed|auto	Specify language
deai_batch.py <file> --all-sections	Batch analysis of all sections
deai_batch.py <file> --chapter <file>	Process specific chapter file
style_profile.py ref1.txt ref2.txt	Extract 8-dimension style profile as JSON
style_profile.py *.md -o profile.json	Save style profile to file

Troubleshooting

• If section detection fails, use --section to specify manually.
• For mixed-language documents, use --lang mixed to run both Chinese and English patterns simultaneously.
• If false positive rate is high, check if claims are statistically qualified (p-values, percentages).
• For non-academic text (blog posts, emails), focus on vocabulary and template patterns; structural skeleton analysis is less relevant.