| name | humanizer |
| description | Strip AI writing patterns from text. Standalone access to the economics-paper-writer agent's humanizer pass. Checks 24 patterns across 4 categories (structural, lexical, rhetorical, formatting) with academic adaptation. Use on any text that reads too "AI-generated". |
| disable-model-invocation | true |
| argument-hint | [file path to .tex, .md, or .txt file] |
| allowed-tools | ["Read","Write","Edit","Grep","Glob"] |
Humanizer
Strip AI writing patterns from academic text. This provides standalone access to the humanizer pass that's built into the economics-paper-writer agent.
Input: $ARGUMENTS — path to file to humanize.
Workflow
Step 1: Read the File
Read the target file from $ARGUMENTS. Support .tex, .md, .txt, and .qmd files.
Step 2: Scan for AI Patterns
Check all 24 patterns across 4 categories:
Category 1: Structural Tics (6 patterns)
- Triplet lists — "X, Y, and Z" appearing 3+ times in a section
- Formulaic transitions — "Moreover", "Furthermore", "Additionally" as sentence starters
- Echo conclusions — final paragraph restates every point
- Uniform paragraph length — all paragraphs suspiciously similar in length
- Topic sentence + support — every paragraph follows identical structure
- Numbered/bulleted reasoning — "First... Second... Third..." in prose
Category 2: Lexical Tells (6 patterns)
- "Delve" — almost never used by humans in academic writing
- "Landscape" — as metaphor ("the policy landscape")
- "Crucial/pivotal/vital" — overused intensifiers
- "Multifaceted" — AI favorite
- "Underscores" — as verb ("this underscores the importance")
- "Navigate" — metaphorical ("navigate the challenges")
Category 3: Rhetorical Patterns (6 patterns)
- Excessive hedging — "it is worth noting", "interestingly", "arguably"
- Performative enthusiasm — "fascinating", "remarkable", "exciting"
- False balance — "while X, it is also true that Y" for every claim
- Hollow acknowledgment — "this raises important questions" without answering them
- Premature synthesis — summarizing before enough evidence is presented
- Universal agreement — "scholars agree", "it is widely recognized"
Category 4: Formatting Tells (6 patterns)
- Over-signposting — "In this section, we will discuss..."
- Excessive parallelism — every sentence in a list has identical structure
- Definitional opening — starting sections with dictionary-style definitions
- Disclaimer stacking — multiple caveats before making a point
- Summary before content — "This section covers X, Y, and Z" at the start
- Colon-list pattern — "There are three key factors: (1)... (2)... (3)..."
Step 3: Apply Fixes
For each detected pattern:
- Identify the specific instance
- Rewrite to sound more natural/human
- Preserve the academic content and meaning
Academic Adaptation Rules
- Preserve formal structure where it's genuinely needed (equations, theorems, proofs)
- Keep technical terms — don't simplify domain vocabulary
- Maintain citation density — don't remove scholarly references
- Vary sentence structure — mix short and long, simple and complex
- Allow imperfection — real writing has slight asymmetries and personality
Step 4: Report Changes
Present a summary of changes:
## Humanizer Report: [filename]
**Patterns found:** N / 24
**Changes made:** N
| # | Pattern | Category | Location | Change |
|---|---------|----------|----------|--------|
| 1 | Triplet lists | Structural | Section 3, para 2 | Varied list lengths |
| 2 | "Delve" | Lexical | Line 47 | Replaced with "examine" |
...
Step 5: Save
Overwrite the original file with the humanized version (the user can review via git diff).
Principles
- Preserve meaning. The content must be identical — only the expression changes.
- Don't over-correct. Some formal academic patterns are normal, not AI tells.
- Context-aware. "Moreover" in a proof is fine. "Moreover" as every paragraph opener is not.
- Reversible. User can always
git checkout to undo.
- Light touch for good writing. If only 2-3 patterns found, the text is probably fine. Focus effort on heavily AI-patterned text.
