mathews-tom/humanize

Humanize: AI Pattern Detection and Removal

Remove AI-generated writing patterns from text. Produce natural, human-sounding output that preserves meaning.

This is not a generic rewriter. It targets specific, documented AI-writing patterns catalogued by Wikipedia's WikiProject AI Cleanup from thousands of observed instances.

Workflow

Five phases. Each phase has a clear input, transformation, and output. Do not skip phases.

Phase 1: Detection Scan

Read the input text. Load references/detection-patterns.md. Scan for two categories of signals:

A. Lexical patterns (the 24 catalogued AI-writing patterns):

| Category | Patterns | Priority | | ----------------- | ------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------- | | Content inflation | Significance puffing, notability claims, superficial -ing analyses, promotional language, vague attributions, formulaic challenges sections | HIGH — loudest AI tells | | Vocabulary | AI-frequency words, copula avoidance, filler phrases, excessive hedging | HIGH — statistically detectable | | Structure | Rule of three, negative parallelisms, elegant variation, false ranges, inline-header lists | MEDIUM — structural fingerprints | | Style | Em dash overuse, boldface overuse, title case headings, emoji decoration, curly quotes | MEDIUM — formatting tells | | Communication | Chatbot artifacts, knowledge-cutoff disclaimers, sycophantic tone, generic conclusions | LOW — obvious, usually caught by author |

B. Statistical regularity signals (see references/statistical-signals.md):

| Signal | What to look for | | ------------------------------- | ---------------------------------------------------------------------------------------------- | | Sentence length uniformity | Sentences clustering within a narrow word-count range | | Low clause density variation | Every sentence has the same number of clauses | | Flat information density | Every sentence carries roughly the same amount of detail | | High-frequency phrase templates | Stock collocations and common bigrams/trigrams dominating the text | | Excessive transition markers | Formal connectives appearing more than 8 per 1,000 words | | Structural symmetry | Paragraphs and sentences following balanced, mirror-like patterns | | Uniform inter-sentence cohesion | Every sentence tightly follows the previous with no topic shifts or digressions | | Generic function word usage | Connectors and prepositions used in textbook-standard distribution with no personal tendencies |

Output a detection report using the detection report template (see Output Format).

Instance severity rating:

| Severity | Criteria | | -------- | --------------------------------------------------------------------------------------------------------------- | | HIGH | 3+ patterns co-occurring in a single paragraph, or any paragraph saturated with AI vocabulary (5+ signal words) | | MEDIUM | 1-2 patterns in a paragraph, or a statistical signal present across 3+ consecutive sentences | | LOW | Isolated single instance of any pattern, or a borderline statistical signal |

Phase 2: Structural Rewrite

Transform document structure to break AI-typical organization:

• Convert uniform paragraph lengths to varied blocks
• Merge or split sentences to break rhythmic uniformity
• Reorder clauses where meaning permits
• Convert formulaic list structures to narrative where appropriate
• Remove tripartite constructions unless the content genuinely has three parts

Do not change factual content. Do not add information. Do not remove cited sources, data, or technical terms.

Phase 3: Vocabulary and Style Pass

Apply pattern-specific rewrites from the detection report:

• Replace AI-frequency vocabulary with natural alternatives
• Restore simple copulas (is/are/has) where the text uses elaborate substitutes
• Remove filler phrases and excessive hedging
• Cut promotional language and significance inflation
• Replace vague attributions with specific ones (or remove if no source exists)

Load the appropriate style profile from references/style-guide.md based on the target domain. Apply domain-specific voice calibration.

Phase 4: Entropy and Variation

Human writing has burstiness — irregular rhythm, varied sentence lengths, uneven information density. AI text is statistically smooth. This phase breaks that smoothness.

Load references/statistical-signals.md for target ranges. Apply:

• Sentence length variance: mix short declarative with longer explanatory. Target visible variance across any 5-sentence window.
• Clause density variation: alternate simple sentences (one clause) with compound/complex (2-3 clauses). Do not settle on a uniform clause count.
• Information density variation: let some sentences carry heavy detail while others are light — a summary statement, a reaction, a pivot. Uniform density reads as generated.
• Phrase template breaking: replace stock collocations with specific phrasings. "Play a role in" -> name the specific action. "In terms of" -> delete or restructure.
• Inter-sentence cohesion variation: not every sentence should tightly follow the previous. Allow small topic expansions, brief asides, or contextual jumps that a thinking human would make.
• Function word personalization: vary connector usage. Use "but" in one place, "still" in another, nothing in a third. Do not default to the same conjunction pattern throughout.
• Paragraph length variance: mix single-sentence paragraphs with 4-5 sentence blocks.
• Controlled imperfection: fragments at impact positions, parenthetical asides, concessive turns. Sparingly — seasoning, not structure.

Phase 5: Validation and Output

Two checks before delivering:

Semantic check: Compare rewrite against original. Every factual claim, data point, argument, and technical term in the original must be present in the rewrite. If anything was lost, restore it.

Self-audit: Ask internally: "What still sounds AI-generated about this text?" If residual patterns remain, fix them. One pass only — do not loop indefinitely.

Output the final text followed by a brief changes summary.

Output Format

Full Rewrite / Targeted Fix / Style Shift

[Humanized text]

---
Changes: [2-4 bullet summary of what was changed and why]
Patterns detected: [list of pattern numbers/names found]
Domain: [detected or specified domain]

For short texts (under 100 words), skip the changes summary unless the user requests it.

Detection Only

## Detection Report

**Domain:** [detected or specified]
**Overall severity:** [HIGH / MEDIUM / LOW]
**Patterns found:** [count]

### Findings

| Location | Pattern | Severity | Evidence |
|----------|---------|----------|----------|
| Para 1 | #7 AI vocabulary | HIGH | "delve", "intricate", "pivotal" in same sentence |
| Para 2 | #8 Copula avoidance | MEDIUM | "serves as" instead of "is" |
| Para 1-4 | Sentence length uniformity | MEDIUM | All sentences 18-22 words, SD < 3 |
| ... | ... | ... | ... |

### Statistical Signals

| Signal | Status | Detail |
|--------|--------|--------|
| Sentence length variance | FLAG | SD ~3 words (human typical: 7-15) |
| Transition frequency | OK | 5 per 1,000 words |
| ... | ... | ... |

### Summary
[1-2 sentences: overall assessment and highest-priority patterns to fix first]

Reference Files

| File | Purpose | Load When | | ------------------------------------ | ---------------------------------------------------- | ------------------------------------ | | references/detection-patterns.md | 24 AI-writing patterns with examples | Always (Phase 1) | | references/statistical-signals.md | 12 statistical regularity signals with target ranges | Phase 1 (scan) and Phase 4 (targets) | | references/style-guide.md | Domain-specific voice profiles and calibration rules | Phase 3 (match to domain) | | references/transformation-rules.md | Structural rewrite strategies and entropy techniques | Phase 2 and Phase 4 | | examples/academic.md | Before/after pairs for academic writing | When domain is academic | | examples/blog.md | Before/after pairs for blog/casual writing | When domain is blog or social | | examples/professional.md | Before/after pairs for professional/business writing | When domain is professional |

Domain Detection

If the user does not specify a domain, infer from:

1. Vocabulary density and jargon type
2. Citation patterns
3. Sentence complexity
4. Register (formal/informal markers)

Default to professional if ambiguous.

Supported domains: academic, technical, blog, social, professional, marketing

Behavioral Constraints

1. Never fabricate. Do not add facts, citations, quotes, statistics, or claims not in the original.
2. Never remove data. Numbers, dates, names, URLs, and cited sources must survive the rewrite.
3. Preserve argument structure. If the original makes points A, B, C in that order with that logic, the rewrite must preserve the logical flow.
4. Do not over-humanize. Some text is meant to be neutral and informational. A technical specification does not need personality. Match the appropriate register.
5. Respect code blocks and structured data. Do not humanize code, tables, JSON, YAML, or any structured/machine-readable content. Pass these through unchanged.
6. One pass through the pipeline. Do not run the 5-phase pipeline recursively. If the output still has tells after Phase 5, note them in the changes summary rather than looping.

Scope Modes

| Mode | Trigger | Behavior | | ------------------ | -------------------------------------------------------------- | ------------------------------------------------------------ | | Full rewrite | "humanize this", "rewrite naturally" | Run all 5 phases | | Detection only | "check for AI patterns", "does this sound AI" | Run Phase 1 only, output detection report | | Targeted fix | "fix the AI-sounding parts", "just clean up the obvious stuff" | Run Phase 1, then apply fixes only to HIGH-priority patterns | | Style shift | "make this more casual/academic/professional" | Run Phases 3-4 with specified domain profile |

Error Handling

| Problem | Cause | Resolution | | -------------------------------------- | --------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Input under 20 words | Insufficient signal for pattern detection | Report: "Text too short for reliable pattern detection." Apply vocabulary fixes only (Phase 3) if obvious patterns are present. Skip statistical signal analysis. | | Input is entirely code/structured data | No prose to humanize | Report: "Input is structured data — no humanization applicable." Return input unchanged. | | Mixed human + AI text | Partial AI generation or human-edited AI output | Run Phase 1 on full text. Flag only paragraphs/sections with detected patterns. Apply Phases 2-4 selectively to flagged sections. Leave clean sections untouched. | | Domain ambiguous after detection | Input mixes registers (e.g., academic citations in a blog post) | Default to professional. Note the ambiguity in the output: "Domain defaulted to professional — specify if another profile is preferred." | | Semantic drift detected in Phase 5 | Rewrite altered meaning during structural/vocabulary changes | Restore the drifted factual claim from the original. Do not re-run the full pipeline. Note the restoration in the changes summary. | | Input contains fabricated citations | Original text has hallucinated sources | Not detectable — this skill humanizes style, not factual accuracy. Pass through unchanged. Note in limitations if the user asks about accuracy. | | All patterns are LOW severity | Text is mostly human-written with minor tells | In targeted fix mode, report findings but recommend no changes. In full rewrite mode, apply light-touch fixes only — do not over-edit clean text. |

Integration Point

Other writing skills can import references/detection-patterns.md as a pattern library for their own anti-pattern sweeps. The detection patterns are the shared asset; the pipeline is this skill's domain.

Limitations

• Cannot verify factual accuracy of the original text. Garbage in, humanized garbage out.
• Effectiveness depends on input length. Very short texts (under 20 words) have insufficient signal for pattern detection.
• Style profiles are guidelines, not voice cloning. The output will sound natural but will not match a specific author's voice without additional calibration.
• Does not interact with external AI-detection APIs. Assessment is heuristic, not benchmark-verified.