| name | writing-humanizer |
| description | Humanize Chinese or English drafts. Use when the user asks to make writing less AI-like, more human, more natural, less generic, more authored, or when a draft feels over-polished, too symmetrical, generic, translated, or detached from the author's real judgment. Do not use as an AI-word blacklist, synonym-replacement pass, or fact-expansion pass, or to make technical writing casually chatty. Near miss: for structural revision or a blog-specific rewrite, use writing-blog. |
Writing Humanizer
Goal
Make the draft sound authored and situated while preserving the original artifact, meaning, evidence, and boundary. A good pass leaves it sounding like the same author, with mechanical smoothness gone, each edited sentence's job clearer, and nothing turned generic, promotional, or chatty.
Human writing is defined by fit, not synonym swaps or word blacklists: a real speaker, talking to a real reader, under a real purpose, with selective evidence and lived judgment.
Constraints
Preserve the requested artifact type, genre, length, density, outline, factual claims, citations, links, code, diagrams, frontmatter, file paths, and technical terms unless the user asks to change them.
Do not add new claims, examples, sources, sections, metrics, publication context, or a more promotional/casual tone just to sound human.
Add grounding only from the provided text, author-known context, or explicitly cited sources. If grounding is missing, keep the claim more general, ask for the smallest missing fact, or mark the assumption.
Evidence And Retrieval
Default to the user-provided draft and nearby provided context.
Read external sources only when the user requests fact checking or citations, provides a specific source to use, or the rewrite would otherwise require an unsupported factual claim.
Do not search to improve phrasing, add decorative examples, or make the draft sound more authoritative.
Diagnosis Signals
Use these dimensions to diagnose sentences or passages that feel generated. Do not force a five-part analysis for every sentence unless the user explicitly asks for sentence-level review.
- Speaker: who is saying this, and from what position?
- Reader: what does the intended reader know, need, doubt, or resist here?
- Purpose: what job does this sentence do in the paragraph?
- Grounding: what observation, example, consequence, or choice makes it true?
- Voice: does it sound like this author, or like a generic explainer?
Common AI-smell patterns:
- No real speaker: general rule or balanced conclusion with no authorial responsibility. Repair by restoring what the author chooses, refuses, tried, believes, or doubts.
- No real reader: explanation is too much, too little, or in the wrong order. Repair by adding the definition, reason, example, transition, contrast, or consequence needed at that point.
- No selection: prose lists all reasonable points instead of showing what the author decided to include or exclude. Repair by making priority visible.
- No ground: plausible conclusion floats above experience. Repair with what happened, failed, changed, cost something, or can be verified.
- Too much symmetry: balanced pairs, rule-of-three rhythm, repeated openings, identical category blocks, or mechanical contrast. Repair by varying rhythm and letting priority lead.
- Fake intimacy: cute, vivid, emotional, or conversational phrases that do not match the author's normal density. Repair by making the phrase quieter.
Decision Rules
If the user asks for editing, edit directly rather than returning a review.
If missing audience, author stance, source context, or publication context would materially change the rewrite, ask one narrow question. Otherwise make the smallest reasonable assumption and continue.
Use sentence-level diagnosis for reviews, short excerpts, or recurring local problems. Use document-level passes for whole drafts.
Fix the dominant one to three smell patterns first. Stop before polishing the draft into a different voice.
Editing Heuristics
- Convert abstract claims into situated claims.
- Convert general rules into author choices.
- Convert balanced summaries into prioritized judgments.
- Convert vague improvement language into visible consequences.
- Convert mechanical transitions into real logical turns.
- Convert decorative personality into plain authorial stance.
- Convert repeated structures into varied paragraph rhythm.
Keep a sentence even if it contains abstract language when the abstraction is the topic, the term is used consistently, surrounding sentences ground it, or removing it would weaken precision.
Delete a sentence only when it repeats, smooths, praises, summarizes, balances, or announces importance without adding logic or evidence.
Output
When editing a file directly:
- make the edits
- briefly name the main smell patterns removed
- mention core terms, structure, or protected elements intentionally preserved
- mention concrete validation performed
When reviewing without editing, use a compact table and list only the highest impact passages:
| Location | Cause | Why It Feels Generated | Suggested Rewrite |
|---|
Keep comments tied to exact sentences or passages. Do not use "AI flavor" as a substitute for diagnosis.
Validation
For file edits, inspect the diff to confirm each changed claim keeps its original truth value and hedge level (no new certainty, no added or dropped claim), then check protected elements when present: frontmatter, headings, links, citations, code blocks, diagrams, file paths, commands, schema fields, and technical terms.
For Markdown/docs edits, run git diff --check -- <file> when available and relevant. If validation cannot run, say why and name the next best check.
Stop Rules
Stop after the requested pass. Do not broaden into structural rewrite, fact checking, citation work, translation, SEO, publishing advice, or style-system redesign unless the user asks.
If the text is already authorial enough, say what was preserved and make only minimal edits or no edit.
