humanizer-workbench

Rewrites AI-generated text into natural, human-quality writing using a multi-stage detection and transformation pipeline.

Purpose

This skill identifies and removes AI-generated text patterns — uniform sentence structure, formulaic vocabulary, filler constructions — and rewrites the text to match a specified voice. It mirrors the behavior of the humanizer CLI tool: detect what's AI-like, rewrite for style, refine for rhythm, audit the result.

When to use

• The text reads like it was generated by an LLM, even if it's factually accurate
• AI-characteristic vocabulary is present: leverage, seamless, tapestry, nuanced, comprehensive, pivotal, facilitate, empower, foster, robust, innovative, etc.
• Sentence lengths are metronomically uniform
• Paragraphs open with "Furthermore," "Moreover," "It is worth noting that," or similar
• The user wants the text to match a specific voice (professional, founder, technical, etc.)

When NOT to use

• The text is already human-written. If it reads naturally, leave it
• The user wants grammar or factual corrections only — this skill changes structure and voice
• The text is highly technical code, API documentation, or domain content where flagged vocabulary is standard and accurate
• The text is under 50 words — small changes are better handled directly without invoking the pipeline

---

Pipeline

The skill applies up to three sequential stages, depending on intensity. Each stage has a focused purpose; doing all three in one pass produces weaker results than doing each separately.

REWRITE (always)

Remove AI vocabulary and filler phrases. Apply the target style's voice, sentence patterns, and structural guidance. This is the primary transformation — output should be substantially different from input.

• Temperature: 0.7 (highest creative latitude of the three stages)
• Input: original text + detection results + style guidance + intensity
• Output: rewritten text with most AI patterns removed

REFINE (medium + aggressive)

Improve rhythm and sentence variety. Fix remaining parallelism, uniform paragraph lengths, and structural patterns. Do not change meaning — only improve flow.

• Temperature: 0.45
• Input: REWRITE output
• Output: text with improved sentence variety and rhythm

AUDIT (aggressive only)

Fresh read after REWRITE and REFINE. Identify any remaining AI-like constructs the previous stages missed. Apply targeted fixes. This stage is conservative — only change what's clearly still wrong.

• Temperature: 0.3 (most constrained; preserve what earlier stages got right)
• Input: REFINE output + fresh detection results
• Output: final cleaned text

---

Styles

Apply the style the user requests. If none is specified, use professional. These are behavioral presets — they change structure, vocabulary, and voice, not just the prompt framing.

professional — Peer-to-peer, direct. Lead with conclusions, then support. Active voice. Specific over vague. No corporate language. Authority comes from specificity, not formality.

casual — Conversational, first-person. Use contractions. Mix short and long sentences. Trust the reader. Write the way a knowledgeable person talks.

technical — Expert-to-expert. No hand-holding. Quantify claims where possible. Name tradeoffs. Use domain terminology precisely. Information density is high.

founder — Personal and opinionated. First person. Lead with what happened or what was learned, not with what it means. Specific dates, numbers, and failures over abstractions. Avoid startup jargon.

academic — Analytical and measured. Evidence-backed claims. Hedge only where the evidence actually warrants it. Acknowledge alternative interpretations. Transitions connect arguments, not just sentences.

storytelling — Scene-first, varied pace. Show rather than tell. Vary sentence length deliberately. Open with a specific moment or detail, not an abstraction. Specific nouns over generic ones.

---

Intensity levels

Level	Stages	When to use
light	REWRITE	Text is mostly clean; vocabulary and filler cleanup only needed
medium	REWRITE → REFINE	Standard AI-generated text; default for most cases
aggressive	REWRITE → REFINE → AUDIT	Heavily templated, structurally uniform, or high-stakes text

Default: medium.

---

Rewrite principles

Apply these regardless of style or intensity:

1. Remove AI vocabulary first. If the word is in the list below, replace it — don't preserve it because it technically works.

2. Cut filler constructions entirely. "It is worth noting that", "Furthermore", "Moreover", "It is important to note that", "In conclusion", "At this point in time" — remove or restructure the sentence.

3. Vary sentence lengths. Four consecutive sentences of similar length is a pattern. Break it.

4. Replace vague claims with specific ones. "Comprehensive strategy" → what the strategy actually does. "Robust solution" → what it actually handles.

5. Use simple verbs. "serves as" → "is". "facilitates" → "helps" or just the direct verb. "leverages" → "uses".

6. Preserve meaning exactly. Do not add information not present in the original. Do not remove substantive claims.

7. Match the register. Don't make academic writing chatty or technical writing casual. The goal is natural for the context, not artificially informal.

---

AI vocabulary reference

Words and patterns to eliminate or replace:

Inflated adjectives: comprehensive, robust, seamless, innovative, groundbreaking, transformative, cutting-edge, unprecedented, state-of-the-art, holistic, dynamic, synergistic

Filler verbs: leverage, facilitate, empower, optimize, streamline, foster, cultivate, harness, underscore, highlight, showcase, navigate, unlock

Abstract nouns used as fillers: landscape, tapestry, ecosystem, paradigm, synergy, framework (when vague), nexus, realm

Filler openers: "It is worth noting that", "It is important to note that", "Furthermore", "Moreover", "Additionally", "In conclusion", "At the end of the day", "In today's rapidly evolving"

Copula avoidance patterns: "serves as a testament to" → "shows", "stands as a reminder" → "reminds", "marks a pivotal moment" → "is a turning point"

Significance inflation: "underscores its importance", "reflects broader trends", "highlights the significance of", "marks a shift in the landscape"

---

AI-likeness scoring

The scorer assigns 0–100 before and after, based on five components:

Component	Max	Measures
AI vocabulary density	30	Ratio of flagged words to total words
Filler phrase density	25	Number of distinct filler phrases found
Sentence length uniformity	20	Inverted standard deviation of sentence lengths
Structural patterns	15	Em dash count, list density, opener patterns
AI opener patterns	10	Formulaic starters in the first five sentences

Grade labels: ≥75 Very AI-like · 50–74 Moderately AI-like · 25–49 Slightly AI-like · <25 Mostly human

The score is a diagnostic tool. Output is not gated by score — the user decides whether the result is clean enough.

---

Output format

1. Provide the rewritten text directly — no preamble ("Here is the rewritten version:" is unnecessary)
2. Below the text, note the main changes in 2–4 brief bullets
3. Show before/after scores when the user asks, or when the improvement is large enough to be worth noting
4. If the text scores below 15 after rewriting, note that it reads clean
5. Do not show intermediate stage outputs unless the user asks

---

Example

Input (professional, medium):

It is worth noting that this comprehensive approach leverages robust documentation strategies to facilitate seamless onboarding for engineering teams. Furthermore, proactive knowledge transfer empowers developers to optimize their workflows and achieve unprecedented productivity outcomes.

Output:

Good documentation cuts onboarding time. Engineers who can find answers without interrupting teammates get productive faster, and the teams they join stay focused longer. Most documentation failures are about findability, not volume.

Changes:

• Removed filler opener ("It is worth noting that") and restructured around a concrete claim
• Replaced AI vocabulary: comprehensive, leverages, robust, facilitate, seamless, empower, optimize, unprecedented
• Broke one passive sentence into two with different lengths and active subjects

Score: 72/100 → 8/100

---

CLI equivalent

If the user has the package installed:

humanizer input.txt --style professional --intensity medium
humanizer input.txt --style founder --intensity aggressive --diff --score
humanizer-detect input.txt   # analyze without rewriting