| name | tool-humanizer |
| description | Remove AI-generated writing patterns and restore natural human voice. Detects and fixes 50+ AI tells: inflated symbolism, promotional language, hedging, corporate buzzwords, robotic structure, em dash overuse, rule of three, vague attributions, negative parallelisms, conjunctive phrase abuse. Three modes: quick (obvious patterns), standard (full scan + score), deep (voice-matched using research_context/research-profile.md). Called automatically by sci-communication as a post-processing step for blog posts, social threads, and newsletters. Also invoked standalone: "humanize this", "de-AI this", "make this sound human", "remove AI patterns", "clean up this writing". Does NOT trigger for formal academic writing (sci-writing handles that). Does NOT trigger for content that is intentionally formal or academic.
|
Humanizer
Strip AI-generated patterns from text. Make it sound like a person wrote it.
Outcome
Input text returned with AI patterns removed, scored before and after on a 0-10 human-ness scale. Change log shows exactly what was fixed. Output either replaces the original or is saved alongside it.
Context Needs
| File | Load level | How it shapes this skill |
|---|
research_context/research-profile.md | full | Replacements use the researcher's actual vocabulary, rhythm, and connectors instead of generic "human" phrasing |
research_context/samples.md | tone refs | Reference for what on-voice sounds like |
context/learnings.md | ## tool-humanizer section | Patterns the user cares about most, false positives to skip |
Load if they exist. Works standalone with generic replacements.
Instructions
Step 1: Detect Mode
Choose from context or ask: "Quick pass, full cleanup, or voice-matched?"
| Mode | What happens | Best for |
|---|
quick | Remove obvious AI cliches + buzzwords. Single pass, no scoring. | Fast social edits, internal docs |
standard | Full pattern scan (50+ detections) + human-ness score + change log | Any content going public |
deep | Full scan + replace with voice-profile patterns. Loads research_context/research-profile.md | Blog posts, landing pages, emails — anything that must match the researcher's voice |
Default: standard. When called by another skill as a post-processing step, use deep if voice-profile exists, standard if not.
Step 2: Load Context
If mode is deep, load research_context/research-profile.md. Extract:
- Preferred vocabulary (use these as replacements)
- Avoided words (flag these as AI tells even if not in pattern library)
- Linguistic habits (connectors, intensifiers, rhythm patterns)
- Samples for tone reference
Read context/learnings.md → ## tool-humanizer for known false positives or priority patterns.
Step 3: Score Original
Rate the input text 0-10 on the human-ness scale:
| Score | Meaning |
|---|
| 0-3 | Obviously AI — multiple cliches, robotic structure, hedging everywhere |
| 4-5 | AI-heavy — some human touches but needs major work |
| 6-7 | Mixed — could go either way, lacks distinctive voice |
| 8-9 | Human-like — natural voice, minimal AI patterns |
| 10 | Indistinguishable from a skilled human writer |
Scoring factors:
- AI pattern count per 500 words (fewer = better)
- Sentence length variance (higher variance = more human)
- Specificity ratio (concrete terms vs vague qualifiers)
- Structural variety (not every paragraph same shape)
Step 4: Apply Pattern Detection + Removal
Read references/pattern-library.md for the full detection list. Categories:
- AI cliches & openers — "In today's fast-paced world", "Let's dive in", "It's no secret"
- Hedging language — "It's important to note", "arguably", "one might argue"
- Corporate buzzwords — "leverage", "utilize", "facilitate", "optimize"
- Robotic structure — rhetorical Q+A, obsessive parallelism, always-three lists, "Here are the top X"
- Overused transitions — "Moreover", "Furthermore", "Additionally", "Nevertheless"
- Promotional inflation — "transformative", "game-changer", "unprecedented"
- Wikipedia AI tells — inflated symbolism, em dash overuse, rule of three, vague attributions, negative parallelisms, superficial -ing analyses, excessive conjunctive phrases
- Vocabulary tells — "delve", "tapestry", "multifaceted", "landscape", "nuanced", "foster", "realm"
Read references/replacement-guide.md for what to replace with.
Step 5: Enhance Human Markers
After removing AI patterns, add natural voice signals:
- Varied sentence rhythm — break up same-length sentences
- Contractions — "it's" not "it is" (unless formal context)
- Active voice — flip passive constructions
- Confident assertions — remove hedging unless genuinely uncertain
- Specific examples — flag vague references for the user to make specific
In deep mode, apply voice-profile patterns:
- Insert preferred connectors and transitions
- Match sentence length patterns from samples
- Use researcher's vocabulary as replacements for generic terms
Step 6: Score Revised + Output
Score the revised text. Show the change summary:
ORIGINAL: 4.2/10
REVISED: 8.4/10
Changes:
[N] AI cliches removed
[N] buzzwords replaced
[N] hedging phrases cut
[N] structural patterns fixed
[N] voice markers added
Flags for review:
[paragraph/line] — [what needs manual attention]
If called standalone, present the cleaned text directly.
If called as a post-processing step by another skill, return the cleaned text silently (the calling skill handles output).
Step 7: Log Learnings
If the user gives feedback ("that was too aggressive", "keep the em dashes", "good catch on X"), log to context/learnings.md → ## tool-humanizer.
Pipeline Mode
When called by another skill (not standalone), this skill:
- Receives text as input
- Runs Steps 2-6 silently
- Returns cleaned text
- Only shows the score summary if the change was significant (delta > 2 points)
The calling skill is responsible for saving the final output.
Rules
Updated automatically when the user flags issues. Read before every run.
Self-Update
If the user flags an issue — too aggressive, missed a pattern, false positive — update the ## Rules section immediately with the correction and today's date. Also log to context/learnings.md under ## tool-humanizer.
Scoring Thresholds
| Score | Label | Action |
|---|
| 90-100 | Clean | No changes needed |
| 70-89 | Light | Fix flagged patterns only |
| 50-69 | Moderate | Rewrite flagged sentences |
| 0-49 | Heavy | Full rewrite recommended |
