Identify and rewrite AI writing traces in text. The rules are based on actual measured comparisons between human articles and outputs from multiple models, and each has a locatable trigger marker. First determine the genre, then handle the AI flavor. Do not force in person, colloquialism, or short paragraphs just to "make it sound human-written."
Work with the writing distiller
This skill is the final step in the writing process. First distill the Writing-DNA, write according to the DNA, and finally use this skill to remove AI traces.
When a distillation product exists in the same directory, first read 语言DNA.md(in English language-dna.md) before rewriting. Several rules in this skill require judging the "original register." Rule 7 judges literary genres according to their original style, Rules 4 and 5 depend on the author's original punctuation habits, and Rule 8 depends on whether the original text provides specific data. The sentence-length distribution, punctuation habits, and word preferences in the distillation product are the basis for judgment.
When the distillation product conflicts with these rules, the distillation product takes precedence. That is the target author's actual way of writing, not an AI trace.
Hard boundaries (highest priority)
This skill uses whitelist-style rewriting. It may only handle the specifications explicitly listed in the "Rewriting Rules" and may not modify other content based on general writing experience. Sentences that do not match any rule must be preserved word for word.
Even if a sentence matches a rule, only the part necessary to solve that problem may be changed. Do not polish along the way, do not replace words that are not problematic, and do not adjust tone, level of detail, or information density. When unsure whether a location matches a rule, keep the original text.
Article structure is outside the scope of handling. Title levels, chapter order, paragraph count and order, and the positions of lists, tables, quotes, and code blocks must be preserved as-is. Unless the user separately and explicitly requests otherwise, do not summarize, expand, delete points, or reorganize arguments.
Information conservation: no additions, no deletions
The original text's points and facts must be fully preserved. The following content must never be added through rewriting, even just to make sentences read more concretely or more human-written.
- Names, institution names, positions, place names
- Numbers, proportions, amounts, durations, frequencies
- Dates, times, order of occurrence
- Quotations, dialogue, other people's statements
- Cited sources, origins, links
- Causality, motives, mental activity
- Any details, scenes, or cases not written in the original text
The reverse is equally prohibited. Do not delete the original text's points, conclusions, qualifiers, and concessions ("may," "usually," "in some cases," "it is said"). Changing "may improve" to "improve" is falsifying the strength of the tone, not removing AI flavor.
Method of judgment: every content word in the rewritten text must be traceable to the original text. If it cannot be traced, it is an addition and must be reverted.
"Making the abstract concrete" is not the responsibility of this skill. If the original text is abstract, let it remain abstract. Rule 8 only permits moving specific data already present in the original text to the position of a summary word; it does not permit creating data yourself.
Usage
When the user sends over a passage of text, you:
- First lock down the original framework. The positions of heading levels, section order, paragraph order, lists, tables, quotations, and code blocks must not be changed.
- Check item by item only according to the "rewriting rules"; do not add judgment criteria outside the checklist.
- Change only the text that clearly hits a rule, and keep the change to the minimum scope needed to solve that problem. Keep all other text verbatim, adding or removing no information.
- Review against the acceptance checklist at the end, then output the full rewritten text.
No need to explain what was changed. Explain only when the user asks.
Cases where no change is made
- Article framework: preserve the positions of heading levels, section order, paragraph order, lists, tables, quotations, and code blocks.
- Dialogue style and colloquial style: these are already natural expressions; do not rewrite them uniformly.
- Questions: questions in the body, rhetorical questions, and question-style subheadings are not changed; see "Not grounds for rewriting" below.
- Literary expression: parallelism and metaphor in essays and fiction should be judged according to the original style.
- Genre requirements: academic, news, official document, and similar genres do not need "I/you" forced in, nor do they need to be split into short paragraphs.
- Code blocks and foreign-language paragraphs: outside the scope of rewriting.
- Quoting another person's original text: do not change the quotation.
Not grounds for rewriting
The following features look like AI flavor, but in testing they do not hold up, so text cannot be changed on that basis. This table, like "Cases where no change is made," is a hard constraint, not a reference.
| Feature | Why no change |
|---|---|
| Sentence length and paragraph length are not uneven enough | Testing shows no difference from human writing. Do not adjust sentence length or split paragraphs to create rhythm. |
| Too few single-character function words (jiu/hen/le) | AI indeed uses fewer, but the direction is to add, not delete. Adding function words cannot be accepted in review and easily turns formal writing into colloquial writing. |
| Repeatedly writing full names and rarely using pronouns | Testing shows humans repeat the same noun more often than AI. |
| Passive sentences | Abstract passives ("is considered / is regarded as") are far below the inclusion threshold. |
| Nominalization and long sentences themselves | Normal usage in modern Chinese written language; humans use it the same way. |
| "First... second..." in the body | No difference from human writing. Rule 6 only governs subheading numbering. |
| Parallelism with the same structure within a sentence | Humans use phrases like "improve efficiency, reduce costs, optimize experience" and "more efficient, more focused, more creative" no less than AI. Only handle repetition between sentences; see Rule 3. |
| Questions, rhetorical questions, question-style subheadings | Humans use far more questions in the body than AI; deleting them makes the text less human-like. |
| Metaphor itself, metaphor as a standalone paragraph | Humans use it more than AI. Only handle idealized personifying vehicles; see Rule 7. |
| Abstract nouns paired with concrete verbs | Phrases like "time kept the details for him" and "anxiety took shape" are almost never written on either side. |
The basis of the rules is that trigger markers can be located, not frequency. For the measured data, disproven assumptions, and measurement scripts for each rule, see the research repository lieflat-less-ai-tone。
The same feature varies greatly in magnitude across models; em dashes are five times as common as in humans in some models and almost unused in others. Do not apply judgments based on "the flavor of a certain model"; recognize sentence by sentence only by trigger markers.
Some norms reference KKKKhazix/human-writing。
Rewriting rules (ordered by priority)
1. The reversal hook
The reversal hook first sets up a misunderstanding the reader does not actually have, then knocks it down. Known phrasings include "not... but...", "it is not that... rather...", "lies not in... but in...", "rather than saying... it is better to say...", "on the surface... in reality...", "seemingly... but actually...", as well as "not A, but B", "not A. But B", "you think... but actually...", "only in hindsight did I realize", "in the end", "the answer is exactly the opposite", "A does not matter; what matters is B". This list is only illustrative; the same move must be changed no matter what wording it is dressed in.
Fix: state the judgment directly from the positive side—give the judgment first, then the basis for it. Only when the article has genuinely walked through a process from misunderstanding to correction using its material is the author's self-correction permitted, and even then the fixed sentence patterns above must not be used.
❌ The real barrier is not technology, but cognition.
✅ The real barrier is cognition.
❌ This is not a question of a tool, but a question of a way of thinking.
✅ This is a question of a way of thinking.
❌ What we need is not more data, but better judgment.
✅ What we need is better judgment.
❌ You think users care about how many features there are, but actually they only care about whether they can use it right away.
✅ Users care first and foremost about whether they can use it right away.
2. Overly dense enumeration commas
Stringing together three or more coordinate items with enumeration commas, cramming an entire list into one sentence. Measured: AI 3.21 per thousand characters, humans 1.78, a ratio of 0.56.
Trigger: two or more enumeration commas within a single clause, linking three or more coordinate elements.
Fix: if it can be summarized, do not enumerate item by item; when three or more items must be kept, change the syntax of one of them so they do not line up in the same structure. The original list structure must not be altered and necessary items must not be deleted.
❌ Collection, storage, display.
✅ From collection to display.
❌ Whether it is a startup, a mid-sized company, or a large group, all can benefit from it.
✅ Companies of every size can use it.
Do not change: when the material itself must be listed in full (legal provisions, configuration item lists, financial accounts); inside Markdown lists.
Isomorphic parallelism within a sentence is out of scope for this item. "Boost efficiency, cut costs, improve experience" and "more efficient, more focused, more creative" were measured as being used by humans no less than by AI; see "Not grounds for rewriting".
3. Adjacent sentences with the same structure
Several consecutive sentences use the same syntactic skeleton, with commas in the same places, elements in the same order, and the same way of ending. Each sentence read alone is fluent; read together they feel like filling out a form. Measured: AI 9.41 per hundred paragraphs (two consecutive sentences), 0.84 (three consecutive sentences); humans 4.81 and 0.41—both 2 times. This difference is very stable within human corpora and is not caused by genre.
Trigger: two or more adjacent sentences with the same number of commas, the same order of elements, and similar length. The typical case is "subject + verb + object, + supplementary explanation" recurring consecutively.
Fix: break up the syntax of one of the sentences. Merging, splitting, changing the word order, or turning it into a short sentence all work. Change only what is inside the sentence; do not move paragraph order, and do not delete information.
❌ Cloudflare published a long article proposing a credential isolation framework called Agent Access Model. AWS open-sourced Kiro Crew, a multi-agent orchestration platform that can run all night, with signed audit logs as standard.
✅ Cloudflare published a long article proposing a credential isolation framework called Agent Access Model. AWS's open-source Kiro Crew, on the other hand, is a multi-agent orchestration platform that can run all night, with signed audit logs as standard.
❌ After this feature launched, user feedback was very good. The team's development efficiency improved noticeably. The next optimization direction has been determined.
✅ After this feature launched, user feedback was very good, and the team's development efficiency also improved. What to optimize next has already been decided.
When three or more parallel examples are placed side by side in one paragraph, just breaking up the syntax is not enough. The three examples are saying the same thing, and reading them feels like padding. At this point, you can merge similar items, or compress the secondary ones into half a sentence, leaving the most representative one in the main clause position. Information must not be deleted; only the primary and secondary roles are changed.
❌ Cloudflare published a long article proposing a credential isolation framework called Agent Access Model. AWS open-sourced Kiro Crew, a multi-agent orchestration platform that can run all night, with signed audit logs as standard. Y Combinator also released its internal tool QM, which comes with a "human approval gate" and three levels of security posture.
✅ Cloudflare published a long article proposing a credential isolation framework called Agent Access Model. AWS and Y Combinator each open-sourced a multi-agent orchestration tool: Kiro Crew comes with signed audit logs as standard, and QM comes with a "human approval gate" and three levels of security posture.
This layer has no hard trigger marker. In actual measurements, AI's example-sentence density is 3-4 times that of humans, but "exactly three" is an illusion. In paragraphs that already contain examples, the proportion of three-example cases is 0.6% for AI and 0.5% for humans, with no difference. So only compress when the examples are truly parallel and clearly look like padding. Narrative progression (what happened first, how it was done, what the result was) does not count; do not touch it.
Do not change: list-style paragraphs that should be neat by nature (operational steps, parallel clauses); dialogue and quotations; intentional parallelism in literary genres such as prose and fiction.
4. Dash overuse
AI often uses dashes to create a pause followed by a reveal, or to insert unnecessary additions. In nonfiction writing, use them only when the sentence meaning truly requires it, not to emphasize ordinary judgments.
Measured: humans 0.80 per thousand characters, AI 2.38; the differences by model are large, with DeepSeek at 5.16, Claude at 4.25, Kimi at 2.32, while Gemini is 0.51 and GPT 0.11. Some argue that after model updates, dashes no longer count as an AI trace, but this holds only for a few models. In writing scenarios, Claude uses them a lot, and it is five times higher than humans. As long as this magnitude remains among mainstream writing models, this item stays.
Fix: write the complete sentence directly, or switch to commas and periods.
❌ The answer is actually very simple—focus.
✅ The answer is focus.
❌ He found a path—a path everyone else had ignored.
✅ He found a path no one else had noticed.
5. Colon overuse
Two forms, both letting the colon speak for you.
1. A lead-in phrase introduces content. Trigger marker: before the colon are lead-in phrases such as "in one sentence," "the core is," "the key lies in," "the reasons are as follows," "conclusion," "essentially," "in other words." Measured: AI 0.29 per thousand characters, humans 0.08, about 3.8 times, with the human side consistently low.
There are two fixes. If the lead-in phrase itself carries no information, delete it and write the content directly; if the lead-in phrase connects to the preceding text, keep it and only replace the colon with a period or comma. The way to judge is to see whether deleting this half-sentence loses information or breaks the connection.
❌ In one sentence: this plan costs too much.
✅ This approach costs too much.
❌ The core is: first solve the payment failure problem.
✅ First solve the payment failure problem.
❌ There are two reasons: the delivery was late, and the budget was also exceeded.
✅ The delivery was late, and the budget was also exceeded.
When the first half of the sentence connects to the preceding text, it cannot be deleted; only the punctuation can be changed. For example, "This sounds like an almost perfect deal: capacity increased, but the amount of computation did not grow proportionally" — the first half is a linking element, and deleting it breaks the sentence. It is enough to change it to "...an almost perfect deal, capacity increased, but the amount of computation did not grow proportionally."
In this type of sentence, the colon is not used as a basis for rewriting. In actual measurements, AI uses 3.00 per thousand characters and humans use 1.80, but the variation among humans reaches 6.6 times; some writers hardly use it at all, while others use it even more than AI, so it is impossible to judge the source of a colon based on this. Only when it appears repeatedly within the same paragraph and every sentence relies on colons for layering should this rule be applied.
2. Empty sentences that introduce a list. A sentence that says nothing substantive and is used only to announce that a list follows, ending with a colon. Trigger markers: a line in the main text ends with a colon, immediately followed by list items or numbering, and removing this sentence does not reduce the information. In actual measurements, AI uses 0.29 per thousand characters and humans use 0.03, about 9 times, and all human groups are close to zero.
How to revise: write this sentence itself as a substantive judgment, or simply delete it and let the list connect to the preceding text. List items cannot be deleted.
❌ Several typical scenarios I have seen:
✅ I have seen all these scenarios.
❌ A basic check usually includes:
✅ A basic check usually includes these items.
Do not change: direct quotations when introducing a person ("He said: 'Let's talk tomorrow.'"); colons in subheadings; colons inside list items; colons that introduce an explanation in a sentence ("What people are good at is: judging business logic" — this kind of general statement plus specific statement is standard usage, and the variation in usage among humans is too large to detect a stable difference); URLs, code, and machine fields.
6. Ordinal numbers as subheadings
AI likes to use "One, Two, Three..." as subheadings to number the entire text, from beginning to end. This is the form with the strongest template feel. A subheading is supposed to explain what this section is about, but putting the number at the very front pushes the explanation into a secondary position.
Trigger markers: subheadings (# at the beginning, or bold text on its own line) numbered with "One, Two, Three" or "First, Second," and three or more in succession.
How to revise: delete the numbering and keep the original wording of the subheading. Do not change the wording, order, or level of the subheading.
❌ ## 1. Don't rush to write the prompt; first find where the ambiguity is / ## 2. Write the requirements as six modules / ## 3. Synthesize the first usable prompt
✅ ## Don't rush to write the prompt; first find where the ambiguity is / ## Write the requirements as six modules / ## Synthesize the first usable prompt
Cases where no change is made:
- "First... second... finally..." in the main text, as well as "First," and "on the one hand" within a sentence. In actual measurements, this kind of writing shows no difference from human writing and is not an AI trace
- The original text is indeed organized by steps, time, or legal provisions, and the numbering serves a referential function ("see Article 3")
- Markdown ordered lists (
1.2.). The list itself is a numbered structure; do not touch it - The subheading has only one or two numbers and does not constitute numbering throughout the whole text
7. Anthropomorphic metaphors
Comparing a tool, system, or process to an idealized person, such as "like a wise mentor," "a tireless junior reviewer," or "a personal digital secretary." This kind of vehicle is the author being sentimental; it does not explain any mechanism, and the vehicle is often followed by a string of positive modifiers or an expansion of the form "not only... but also..."
Measured: AI 0.018 per thousand characters, human 0.002, a factor of 7.3; when the vehicle carries modifiers such as "never... / wise / omnipotent," the difference reaches a factor of 12.6.
Trigger markers: "like / equivalent to + a / an + occupational role (mentor, secretary, assistant, consultant, butler, reviewer, intern...)," and the vehicle carries positive modifiers, or is followed by "not only... but also..."
Fix: replace it with what the thing actually does, or use a specific person as the vehicle.
❌ It is like a wise mentor, not only conveying conclusions but also conveying multiple perspectives for viewing the world.
✅ It does not only give conclusions; it also explains different angles for looking at problems.
❌ This is equivalent to equipping the team with a tireless junior reviewer that responds in seconds.
✅ It does not get tired, responds in seconds, and can handle the work of junior review.
Cases where no change is needed:
- Metaphor itself is not an AI trace. Measured, humans use metaphors 2.4 times as often as AI, and metaphors standing alone as a paragraph are 8 times as common. Only handle the "idealized occupational persona" type of vehicle
- Specific people as vehicles ("like an old master craftsman," "like a doctor friend with a wide network"). Humans write this way
- When the vehicle brings understanding that the main text does not provide. How to judge: delete the metaphor from the whole paragraph; if the reader would miss some information, keep it; if the previous paragraph has already explained the mechanism and this metaphor merely restates it in different words, then it is padding and is outside the scope of this item and does not need to be kept
- Literary genres such as novels and essays are judged according to their original style
8. Write existing specific data back into generalized statements
AI exposition relies on abstract generalization, humans rely on concrete material. Measured number density: AI 6.34 per thousand characters, human 17.92, humans 2.8 times as many; time anchors 0.69 vs 1.78, 2.6 times. This is the difference with the broadest coverage. Em dashes and zero subjects at the start of paragraphs appear only in specific sentence patterns; thin material is the underlying tone of the whole piece.
But this skill cannot add information that is not in the original text. So this item does only one thing. When the original text has already given specific numbers, times, or facts, generalized statements are not allowed to cover them up.
Trigger markers: the sentence uses generalized expressions such as "significantly improved," "greatly increased," "noticeably improved," "improvement in efficiency," "a large amount," or "many," while the same paragraph or an adjacent paragraph already states the corresponding specific value, time, or object.
Fix: move the existing specific value to the position of the generalized word. When the original text has no specific data, only restore the verb; do not make up numbers yourself.
❌ The new system achieved an improvement in efficiency. Processing time was shortened from two hours to forty minutes.
✅ The new system shortened processing time from two hours to forty minutes.
❌ The team completed optimization of the process.
✅ The team made the process smoother. (The original text gives no data, so only restore the verb.)
❌ The company carried out an adjustment of the organizational structure.
✅ The company adjusted its organizational structure. (Same as above; do not write "merged two business groups"—the original text does not contain this information.)
Nominalized structures such as "completed the ... of ...", "achieved the improvement of ...", and "carried out the optimization of ..." also fall within the scope of this rule; the fix is to restore the verb.
Do not change: when the original text genuinely contains only a summary and no specific material. In this case, do not add anything, and do not delete the sentence just because it is "too abstract". After output, you may note to the author that the material here is thin, and let the author decide whether to supplement it.
9. Ban opening formulas
Do not use "to put it plainly", "to put it bluntly", or "let me state the conclusion first". Delete them and give the judgment directly.
❌ To put it plainly, this project does not have enough budget.
✅ This project has insufficient budget.
10. Translationese (keep only five types)
Translationese itself is not equivalent to an AI flavor. Written modern Chinese has already absorbed a large number of translated sentence patterns; passive sentences, nominalization, and long sentences are all part of normal Chinese writing, and human authors write this way too. In tests on 300 AI samples (1.179 million Chinese characters, 5 models), "in a ... way" was 0.01 per thousand characters, "play a ... role" was 0.02, "enable ... to" was 0.03, and abstract passives of the "is considered / is regarded as" type were 0.09, all below the inclusion threshold; "plural hard translation" and "attract ... attention" almost never appear. These are not grounds for rewriting.
Handle only the following five structures. Expressions outside the trigger markers, even if they read stiffly or subjectively "do not sound like Chinese", must be preserved as they are. Technical terms, legal texts, fixed expressions in official documents, and established translations are not affected.
The frequency noted for each item comes from the above 300-text corpus. For reference, in the same corpus, the already-included "not ... but ..." is 0.70 per thousand characters, and paragraph-initial ordinal numbers are 0.67.
Overlong prenominal modifiers (0.42 per thousand characters, covering 179/300)
When English relative clauses are translated literally into Chinese, it is easy to stuff a very long explanation entirely in front of the noun.
Trigger marker: the modifier before the head noun exceeds fifteen characters, or "的" appears two or more times in succession, so that when reading to the middle of the sentence one needs to go back to confirm the meaning.
Fix: split it into two clauses within the original paragraph; information must not be moved, and meaning must not be added or removed.
Two similar patterns are also handled under this rule. "...的...的..." used in succession to carry over the English possessive ("团队完成了对系统的稳定性的检查" becomes "团队检查了系统稳定性"); "一个／一种" before a noun stacked with another quantifier or demonstrative ("一个关于用户增长的一套新的方法" becomes "一套新的用户增长方法"). If the quantity itself is information, do not change it.
❌ This is a tool that can enable the team to significantly increase review speed without increasing manpower.
✅ This tool can significantly increase review speed without increasing manpower.
❌ We need a solution that can satisfy the usage needs of different users in multiple scenarios.
✅ We need a solution that can satisfy the needs of different users in multiple scenarios.
Fronted temporal clause with "当……时" (0.26 per thousand characters, covering 136/300)
English When …, … after literal translation, a complete clause is piled up at the beginning of the sentence, and the reader has to wait until after "时，" to get the main clause.
Trigger marker: begins with "当" and ends with "时，", with a complete subject-predicate structure in between.
Fix: delete "当" and "时", and let the clause directly serve as the preceding clause.
Do not change: when the sentence really needs to mark a point in time, and deleting it would blur the temporal relationship; for colloquial expressions like “的时候” (measured at 0.02 per thousand characters, not an AI feature).
❌ When everyone can use AI to write articles, the content itself is no longer a competitive advantage.
✅ Everyone can use AI to write articles, so the content itself is no longer a competitive advantage.
❌ When the cost of writing code approaches zero, the cost of choosing the wrong technical direction becomes higher.
✅ The cost of writing code approaches zero, so the cost of choosing the wrong technical direction becomes higher.
Fronted topic shell (0.22 per thousand characters, covering 137/300)
English For … / As for … / Regarding … / In terms of … literally translated into Chinese as a fronted topic shell.
Trigger markers: at the beginning of a sentence, “对于……来说”, “对……而言”, “就……而言”, “关于……”, “在……方面”, while the following clause already takes this object as its subject or object.
Fix: put the object directly into the subject or topic position and delete the shell.
Do not change: when it is truly necessary to limit the scope of discussion, and deleting it would cause the sentence to be read as a global assertion.
❌ For early-stage teams, hiring is the hardest thing.
✅ The hardest thing for early-stage teams is hiring.
❌ Regarding pricing, we ran three rounds of tests.
✅ We tested pricing for three rounds.
❌ As far as user experience is concerned, this change did not bring obvious changes.
✅ User experience did not change noticeably.
Sentence-initial conjunctions as signposts (0.18 per thousand characters, covering 102/300)
English usage tends to put conjunctions at the beginning of a sentence to lead the whole sentence, while Chinese more often embeds them in the middle of the sentence. Measured: humans 0.03 per thousand characters, AI 0.18; the human side is consistently lower.
Trigger markers: “然而”, “因此”, “此外”, “与此同时”, “换言之”, “总而言之” appearing at the beginning of a paragraph or sentence, especially when consecutive adjacent paragraphs begin this way.
Fix: move the conjunction after the subject, or replace it with a more colloquial expression such as “不过”, “其实”, or “也”.
Do not change: when the argument really needs an explicit signpost at a paragraph transition.
❌ However, this solution is not applicable to all scenarios.
✅ This solution actually is not applicable to all scenarios.
❌ In addition, cost is also a factor that needs to be considered.
✅ Cost also needs to be considered.
“This means” / “This shows” restatement sentences (0.15 per thousand characters, covering 122/300)
English This means that … / This suggests that … after literal translation, the conclusion is pulled out into a separate sentence, repeating the content of the previous sentence.
Trigger markers: a sentence begins with “这意味着”, “这表明”, “这说明”, or “换句话说”, and the following content is synonymous with the previous sentence.
Fix: merge it into the previous sentence, or directly delete the lead-in and keep only the conclusion.
Do not change: when what follows really gives a new conclusion that cannot be inferred from the previous sentence.
❌ Retention rose to 72%. This means the product has found PMF.
✅ Retention rose to 72%, and the product has found PMF.
The above five structures are the complete trigger scope of this rule. Expressions not listed cannot be modified on the grounds of “translationese”. When the same sentence hits multiple items, first revise according to the item that appears earlier in order, and after revising, do not apply further stacked processing.
11. Zero-subject commentary at the beginning of a paragraph
The paragraph opens directly with an evaluative comment and never states what the comment is about; the reader has to go back to the previous paragraph to find it. In the measured data, AI accounts for 0.61% of non-first-paragraph openings, while humans account for 0.14%—more than four times as much—with the human side remaining stable.
The total amount of paragraph-opening connectives is about the same for humans and AI (15.1% vs 14.4%). The problem is not "not using connectives," but that the evaluative sentence omits the anaphoric element.
Trigger markers: a non-first paragraph begins with an evaluative phrase such as "sounds like / looks like / plainly put / it is worth noting / more importantly / the key lies in / the problem lies in / means / it is not hard to see," and the entire sentence contains no element such as "this / that / its / the above" that refers back to the preceding text.
Fix: add an anaphoric word, or make the object of the comment explicit. In most cases, adding a single word "this" is enough. Do not change the paragraph order, and do not touch any other text within the paragraph.
❌ It sounds like a feature description, but underneath, something more fundamental has been swapped out.
✅ This sounds like a feature description, but underneath, something more fundamental has been swapped out.
❌ It is worth noting that configuration data is already stored in layers according to "whose this is."
✅ This point is worth noting: configuration data is already stored in layers according to "whose this is."
Do not change: the first paragraph (there is no preceding text to refer to); when the original text already contains an anaphoric element; inside dialogue and quotations.
A standalone paragraph consisting of a metaphor does not fall under this rule. In the measured data, humans use this kind of paragraph opening more than AI (0.21% vs 0.03%). It usually reads abruptly because it lacks an anaphoric reference; per this rule, just add a "this," and do not delete the metaphor. As for whether the metaphor itself is apt or whether it repeats the preceding text, that requires human reading; this skill cannot judge it, so it does not handle it.
Final acceptance
Before outputting, recheck item by item.
- Can every change be clearly mapped to one of the numbered rules above? Any change that cannot be traced to a corresponding rule must be reverted, restoring the original text.
- Are the changes limited to the minimum scope necessary to solve the hit problem, without incidentally polishing or rewriting adjacent text?
- Are all sentences that did not hit a rule preserved verbatim?
- Are the original heading levels, chapter order, paragraph count and order, lists, tables, quotations, and code block positions fully preserved?
- Were only sentences and language expression changed, without reorganizing opinions, lines of argument, or narrative order?
- Do person and paragraph length conform to the original genre, without forcing "I/you" or short paragraphs in order to remove the AI flavor?
- Was sentence length adjusted or paragraphs split in order to "create rhythm"? In the measured data, there is no difference in the dispersion of sentence length or paragraph length; this is not a reason for rewriting.
- Is it still setting up a straw-man misunderstanding and then reversing it, even if none of the fixed words in the list are used?
- Are there still dense enumerations stringing together three or more items with enumeration commas?
- Are adjacent sentences still repeatedly reproducing the same syntactic skeleton; does a paragraph still have three or more parallel examples padding it out?
- Do non-first paragraphs still have zero-subject evaluative openings that lack an anaphoric element?
- Are there still reveal-style dashes, colons introduced by a lead-in phrase, or empty filler sentences that exist only to announce a list?
- Are there still metaphors that compare tools to idealized people ("like a wise mentor")?
- Are there still "plainly put," "to put it bluntly," or "conclusion first"?
- Are subheadings still numbered throughout with "one, two, three"; have "first... second" in the body text been mistakenly deleted?
- Whether actions that could be expressed directly are written as nominalized structures such as "completed the ... of ..." or "achieved the improvement of ..."
- Whether specific numbers, times, and objects already present in the source text have been covered up by generalizing expressions such as "significantly improved" or "greatly increased"
- Whether only the five structures listed in the translationese rules have been handled (overly long premodifiers, "when..." clauses, fronted topic shells, sentence-initial conjunctions, and "this means"-style restatements), with a triggering marker identifiable in each case, and no scope expanded based on feel
- Whether passive sentences, nominalizations, or long sentences themselves have been treated as translationese and changed. These are normal Chinese writing and must be retained
- Whether function words have been added or pronouns changed in order to "sound human-written." These two items are outside the scope of rewriting
- Whether isomorphic parallelism within sentences has been changed or standalone-paragraph metaphors have been deleted. These two items are used by humans no less than by AI and must be retained
- Whether the source of every content word after rewriting can be identified in the original text. For names, numbers, dates, quotations, sources, and causation, anything whose source cannot be identified must be reverted
- Whether qualifiers and concessions in the original text ("may," "usually," "it is said") have been erased, and whether the strength of judgment has been increased