The goal is to make the writing feel more like a specific person expressing themselves in a specific situation, not to mechanically replace a set of "AI high-frequency words."
First determine what the user wants
- Review only: provide the candidate problems that most affect readability and a small number of representative examples, without changing the original text.
- Direct rewrite: complete the rewrite without stopping to ask whether to make changes.
- Review and rewrite: briefly state the main judgments first, then provide the rewritten result.
- When the user specifies tone, length, platform, or content that must not be changed, those requirements take precedence.
When the input is text in a chat, return it directly in the conversation. When the input is a file, follow the user's specified saving method; if no overwrite is specified, by default keep the original file and create the next _V2/_V3 version.
Working method
First read through and confirm four things: what the author wants to say, who it is written for, where it will be published, and which facts and expressions cannot be changed. Ask only for missing information that would substantially change the result; infer the rest reasonably from context.
When rewriting, prioritize handling:
- Vague judgments: replace them with facts, actions, scenes, or perceptible details.
- Templated sentence patterns: reduce the consecutive use of the same transitions, parallel structures, rhetorical questions, and summary structures.
- False identity: do not invent experiences, friends, interviews, or firsthand observations on the author's behalf.
- Monotonous rhythm: adjust sentence length, paragraphs, and pauses according to genre, without deliberately manufacturing rough edges.
- Terminology and clichés: when a simple, accurate way of saying it works, do not use show-off phrasing, report phrasing, or advertising phrasing.
- Over-explanation: when the point or story already stands, delete repeated summaries and forced elevation.
- Idle commentary: delete all sentences that only make emotional comments on other people's words (like "that was a pretty harsh thing to say"), and directly connect to facts, inferences, or actions.
- Feigned slowness and back-reference: when the author is an expert, do not write slow discoveries like "after thinking about it for a while, I realized"; for transitions, do not use "back to X"; you can use a question as a transition (asking the question readers would really ask) or directly continue with the content.
- Pretended shock: delete all emotional reactions fabricated on the author's behalf (like "I got a little restless" or "I was shocked"), and directly give the judgment or fact; write emotions only after they actually happened, and keep them restrained.
- Colloquial speech and tone: replace plain words with how real people say it ("got the hang of it," "felt sure in my heart"), and at key points the tone can be sharper ("what the hell is this"); for vague numbers, use colloquial analogies to prevent nitpicking (in the style of "among any three people, there must be a teacher for me").
- Bring abstractions down to the subject: "there is a possibility of being redesigned" → "there is a possibility of being redesigned by AI and doing the work for you."
- Make back-reference explicit: when using "these two things" or "that passage" to refer back, readers have to count backward; repeat the specific noun and do not save those few characters.
- Land the result in one sentence: after explaining the mechanism, do not stop at describing the process; nail the result into one sentence ("soon, the list of 200 people was picked out"), so readers have a landing point.
- Self-questioning and self-answering: do not ask questions on the reader's behalf (like "what exactly does this statement mean?"), just keep talking; if readers really have questions, they will ask them themselves.
These are all contextual signals, not global forbidden words. Dashes, parallelism, triadic structures, technical terms, and contrastive sentences can be retained in appropriate genres. Do not deliberately add grammatical errors, digressions, unverified details, or emotion just to "sound human."
Boundaries of factual accuracy
- Do not add unverifiable data, quotations, cases, or personal experiences.
- When the original already contains questionable facts, flag them or rewrite them conservatively; do not fabricate sources for them during polishing.
- Fictional settings must retain their fictional nature and must not be disguised as facts.
- Preserve the author's core position, information density, and key qualifications, unless the user explicitly requests a rewrite of the viewpoint.
Tools and references
When processing local files with many mechanical signals, you may run:
python3 scripts/hanyihan_qc.py <文章路径>
Script results are only candidate signals and must be judged in light of genre and context; quality must not be decided by the number of hits.
When item-by-item review or substantial rewriting is needed, read references/editing-guide.md. Simple polishing does not require loading the full reference.
Output
By default, deliver only the result the user needs. Review reports should be ordered by impact and explain "what is unnatural, why, and how it is suggested to be changed"; rewritten results should not include lengthy self-praise or sentence-by-sentence explanations. Provide a change list only when the user requests a comparison.
Before finishing, check: the viewpoint has not been subtly swapped, facts have not been fabricated, genre and length meet requirements, obvious template repetition has been reduced, and it reads smoothly while still sounding like the original author.