| name | humanizer |
| description | Edit prose to remove common AI-writing patterns while preserving every supplied fact and matching the intended voice. Use for humanize, natural rewrite, AI-sounding prose, 去除 AI 味, 降低机器感, 自然润色, or matching a provided writing sample. Do not use for code formatting or factual expansion. |
| metadata | {"version":"2.9.1-codex"} |
Humanizer
Rewrite prose so it sounds like a specific person wrote it, without inventing
facts or flattening legitimate style.
Workflow
- Determine the mode:
- pasted text: return a draft, a brief pattern audit, then the final rewrite;
- file: rewrite prose in place while preserving code, frontmatter, data, and links;
- embedded: return only the finished prose to the calling workflow.
- If the user supplies a writing sample, match its sentence length, vocabulary,
punctuation, paragraph rhythm, and quirks before applying generic rules.
- Preserve every claim, name, number, date, quote, citation, and uncertainty.
Never add plausible specifics that are absent from the source.
- Remove clustered tells: inflated significance, promotional language, vague
attribution, repetitive conclusions, mechanical sectioning, filler,
over-hedging, chatbot phrases, uniform cadence, and manufactured punchlines.
- Keep technical, legal, academic, and reference prose plain when neutrality is
appropriate. Add personality only when the source and requested voice support it.
- Read references/patterns.md only when a detailed
pattern audit, close editing, or ambiguous false-positive review is needed.
- Read the result aloud mentally, verify factual parity, and make a final pass
for repeated structures and unsupported claims.
Prefer concrete verbs and varied sentence structure. Do not erase unusual detail,
mixed feelings, deliberate fragments, quotations, or era-specific language merely
because an isolated pattern resembles model-generated prose.