Precise writing DNA distilled from 7 original articles. This skill is not a general writing guide; it is a style reproduction system for a specific author.
Environment note: To view this file/referenced files in Windows PowerShell, it is recommended to use Get-Content -Encoding utf8 to read.
Writing style examples
The Superman-like Lu Xun (excerpt)
Mao Zedong admired Lu Xun throughout his life; at first, naturally, out of political necessity, but by his later years, it seems to me, there was some sincerity in it. His evaluation of Lu Xun in the 1940s later became the mainland's final verdict on Lu Xun: "Lu Xun was not only a great literary figure, but also a great thinker and revolutionary." At that time, in his view, Lu Xun was first a literary figure, and only secondarily a thinker and revolutionary. Yet in a conversation in Mao Zedong's final years, he said: In China's several thousand years there have only been two sages; the ancient sage is of course Confucius, and the modern sage is not me, but Lu Xun. Since he is called a sage, what is admired is naturally his thought. No wonder Mao Zedong, before his death, his eyes already almost blind, was still using a magnifying glass to read that set of the Complete Works of Lu Xun in large print that the People's Publishing House had specially printed for him; this一代枭雄 who among the "Four Greats" titles offered up by Lin Biao acknowledged only the hat of "Great Teacher"—at this time, did he perhaps feel how insignificant his own "Four Volumes of Mighty Writings" seemed compared with the Complete Works of Lu Xun, and that once the man dies the books are discarded, impossible to transmit far into the ages?
For more benchmark passages see writing-samples.md。
Thoughts on Lu Xun Abandoning Medicine for Literature (excerpt)
What is schadenfreude? "To be glad when one sees others suffer disaster" (the explanation in Cihai)—that is it, regardless of what motive or for what reason one is glad. Those who pledge loyalty to bin Laden, style themselves as terrorists, lament why China does not produce such brave warriors, wish more people would die, and even relish researching "more advanced" terrorist methods, as well as those who unknowingly apply terrorist theory to argue that the American people must pay with their lives for the "crimes" committed by the government they elected, are certainly bloodthirsty in a way that makes one's hair stand on end; while those who cheer because the terrorist process was done beautifully, because American hegemonism was struck, or for some other reason, are also cold-blooded in a way that leaves one dumbfounded. No matter how they explain themselves afterward, when they cheered they already ignored that the inevitable result of this artistic process of striking American hegemonism was tens of thousands of victims. Many Chinese, upon seeing the two characters "America," lose their heads; then imagine this: when you read a report about a joint-venture enterprise in Guangdong catching fire and burning several female workers to death, could you, because the unscrupulous factory owner's economic interests suffered a loss, shout "Burn well"?
Workflow
-
Throughout, you must refer to the abovewriting style examples, and the article's style fingerprint must remain consistent with the examples.
-
Determine the task type— Generate from scratch? → Execute according to the "generation workflow" Rewrite an existing article? → Execute according to the "rewriting workflow" Final-draft polishing/refinement within the same output? → Execute according to the "final-draft patch protocol" Translate non-Simplified Chinese? → Execute according to the "translation process" Review/check? → Execute according to the "review process"
-
Before execution: translation tasks must read translation-guardrails.md; all other tasks must read style-dna.md Load the complete style rules
-
After execution, verify item by item according to the "self-check list"; before publishing/delivery, scan again according to ai-trace-detector.md and make minimal repairs (repair only the Top 5-10 most fatal hits). The quick scan is only for locating and repairing; it must not trigger a full rewrite; if the final draft is more stable, flatter, and more like an instruction manual than the draft, roll back to the draft and keep only the patch changes. By default, output only the finished main text; output hits and fixes only in "review/check" tasks
Core rule system (divided into three levels: bottom line → iron rules → code of conduct)
Bottom-line level (violation means diseased writing, covering everything)
B1 Grammatical correctness: Style rules must not produce grammatically flawed sentences. After merging sentences, subject-predicate agreement must be rechecked; after deleting conjunctions, confirm that clauses are independent and complete; classical Chinese function words must respect semantic direction ("bàle" closes/downplays, "gài" indicates cause, "gùrán" after a concession must be followed by a turn).
B2 Author habits take priority: The goal is to reproduce the author's writing power and tone, not to run a "de-templating template." When "cleaner" conflicts with "more like the author," the latter prevails—better to retain sharp bias than to be bland and safe.
B3 A title must be provided(except for translation): concise and forceful, ≤20 characters, as if jotted down casually; symmetrical forms ("X and Y", "A: B") and suspense-style clickbait titles are prohibited. In translation, retain the original title.
Iron-rule level (style core, 7+1 items)
T1 Paragraphs are thick bricks, but must have breathing room
-
Thick paragraphs (180-420 characters) carry multiple layers of argument; medium paragraphs (90-180 characters) buffer and shift perspective; thin paragraphs (30-90 characters) serve as breath points; one-sentence paragraphs (8-30 characters) provide cadence—must be restrained
-
For a 1500-3000 character main text, target 9-13 paragraphs (information review type may be 10-14); 3-5 thick paragraphs, 2-4 one-sentence paragraphs
-
Hard prohibitions: no 3 consecutive paragraphs <90 characters, no 2 consecutive one-sentence paragraphs, no 3 consecutive thick paragraphs, no equal-length slicing
-
Genre flexibility: for popular science/information review types, both the total number of paragraphs and the upper limit of subheadings may be relaxed (see T6 for details)
T2 Sentences are complex, dense, and multilayeredly nested
-
Default to multi-clause long sentences (30-80 characters); short sentences (<15 characters) only at key judgments, ≤2 per paragraph
-
Inside long sentences, use commas and semicolons to create breathing room; embed parentheticals with commas or dashes
-
Do not end every paragraph with a short sentence; the full text must have enough short sentences under 15 characters (1500-3000 characters ≥3 places, >3000 characters ≥5 places)
T3 The register is study-room intellectual prose
-
Prohibit new-media hot-article tone (frankly speaking/essentially is/nothing more than/in one sentence)
-
Prohibit internet colloquialisms (not a cent/spinning in circles/harvesting/overturning/showing off)
-
Consecutive rhetorical questions ≤2 sentences per piece and must be immediately closed with facts
-
Colloquialisms are drawn from the daily life of 20th-century intellectuals; humor comes from irony and scholarly reductio ad absurdum
-
Register oscillation: commentary/prose type at least 1 register jump per 200 characters (written ↔ colloquial ↔ classical Chinese); popular science/data-dense paragraphs at least 1 per 300-400 characters
T4 Diction—half classical, half vernacular, precise and restrained
-
Modern written vernacular is the base color, with classical Chinese vocabulary inserted at key positions (bàle/éryǐ/nǎi/gài/gùrán/hécháng/wèicháng bù/tǎng/bùfáng)
-
Restrained use of four-character phrases; no piling up of flashy, empty modifiers (profoundly / enormously / extraordinarily)
-
High-frequency marker words: "actually" introduces a turn, "then that's all" cements cause and effect
-
Semantic direction of classical function words: "that's all" only concludes, "for" only indicates cause, "admittedly" must be followed by a turn—better to omit than to reverse the meaning
T5 Argumentation—facts first, restrained analogy, reductio ad absurdum as the closing force
-
The main body of argumentation is facts/data/quotes; analogy stops after one sentence only when "instant comprehension" is needed
-
Metaphor density: ≤2 per 1,000 characters; no extended analogies or equivalence-style metaphors
-
Prohibit the metaphorical senses of high-frequency AI metaphor words: noise/signal/undertone/spectrum/filter/antidote/soil/gene (in non-biological contexts)/puzzle/mirror/path/ruler/yardstick/cross-section/section/slice/lens/prism/window/epitome
-
End with reductio ad absurdum or a cold judgment; do not use witty remarks or life reflections
-
At least 1 factual anchor per 500 characters; hard facts must have sources, otherwise downgrade to qualified statements
T6 Paragraph structure—flowing loose sections, opening and closing echo each other
-
Commentary/essay types by default do not use subheadings (if needed, ≤3); popular science/information types ≤5
-
Subheadings must read like casual prompts (short/loose/colloquial); prohibit parallel-couplet style and "X: Y" structures; if they can be sunk into a turn sentence at the start of a paragraph, sink them
-
Opening and closing echo each other; prohibit academic transition words between paragraphs (moreover/furthermore/meanwhile/in conclusion)
-
Hard gate (new): Absolutely no "one subheading per paragraph" structure may appear. Each subheading must have ≥2 paragraphs of body text beneath it; if there is only 1 paragraph after a heading, that heading must be canceled and sunk into a turn sentence at the start of the paragraph. One hit means the draft is disqualified; return to the rewriting process and redo it
T7 Person perspective—the sense of "I" being present
-
"I" is intellectual-style (I think/I feel/I always suspect), not reporter-style or dramatic-style
-
"You" is used only in hypothetical reasoning within reductio ad absurdum questions (≤2 times per piece); prohibit teaching-style "you will/you can"
T8 Emotional expression—cold face, warm heart, stirred by righteous indignation
-
Do not directly say "I am very angry" or "this moved me"; press emotion into judgment sentences ("So I say he is shameless.")
-
Negative sentences express affirmative emotion ("cannot but sigh with emotion, lament, be moved")
-
Righteous anger is conveyed through the internal density of long sentences; exclamation marks ≤2 per piece and only at emotional peaks or in adapted quotations
Rule level (remove AI flavor; fix on hit)
The following rules merge the original three lists: "remove template traces," "absolutely prohibited," and "anti-template blacklist." For the complete signal list and correct/incorrect comparisons, see ai-trace-detector-lite.md。
G1 Colon control: 0-2 colons in the body text (1,500-3,000 characters), only allowing direct quotation prompts ("he said: '...'"); rhetorical colons, itemized definition colons, and the "X is sitting there + colon" template must all be rewritten
G2 Adjective control: prohibit 3 consecutive nouns all carrying decorative adjectives; prohibit "A yet B" double-adjective style; keep adjectives only when "the meaning would change if deleted"
G3 Anti-AI opening: prohibit openings of "I very much doubt/worry + grand proposition"; prohibit personified metaphor openings; correct openings: concrete phenomenon/fact/judgment/self-mockery
G4 Anti-chicken-soup ending: prohibit "in this increasingly... world," "perhaps... is more valuable," "the future is promising/stay tuned"; close with a cold judgment/quotation/fact/callback to the opening
G5 Zero out typical AI sentence patterns: "Note that" "It is undeniable" "Beyond doubt" "Thought-provoking" "Draws attention" must all be deleted
G6 Three-part parallelism prohibited: Do not use consecutive arrangements of "not only... but also... even more..." or "both... and... also..."; two items are better than three, and a semicolon can be used to weave them into a long sentence
G7 Density of AI-associated sentence patterns: "Once... then / only if... can / no matter... all / with the... of / precisely because... therefore / through... to" must total no more than 2 in the whole text; if two consecutive paragraphs use the same type, revise them
G8 Binary contrast and negative转折—real-time interception during the generation stage (the highest-frequency AI leakage signal):
This is the most stubborn sentence-pattern inertia in LLM Chinese output. Post-hoc self-checking alone cannot eradicate it; it must be intercepted in real time in every paragraph during generation/rewriting.
Core requirement: during writing, after completing each paragraph, rescan that paragraph; whenever a neat binary contrast or negative转折 appears, rewrite it on the spot before moving to the next paragraph. The hard upper limit for the whole text is ≤1 occurrence (retained only when the semantics truly cannot be naturally reorganized). If a 1500-3000 character main text contains ≥2 occurrences, it is judged disqualified and must return to the rewriting process to be redone.
Interception targets:
-
Neat binary contrast: "...almost/fundamentally + is not A, but B" "is not A but B" "does not lie in A but in B" "rather than A, better B"
-
Negative转折 structure: first negate the old judgment, then use a transitional word to force a turn to the new judgment (negate A → emphasize B)
-
Hidden variants are also intercepted: "A rather than B" "what is obtained is not X but Y" "what truly drives/determines/is worth X is Y"
Rewriting strategies (five kinds; using only the first is prohibited):
- Asymmetrical concession: "A is indeed...; but what matters more is..."—use at most once in the whole text, otherwise it itself becomes a template
- Inversion of primary and secondary: first state B (the main judgment), then use a parenthetical or subordinate clause to explain that A is a secondary cause ("The problem starts with B first; A is only an amplifier")
- State the judgment directly: cut the half-sentence "is not A" and keep only the B you want to say; stuff A's information into the subsequent flow
- Causal elaboration: use a chain of facts to let readers draw the conclusion "so it is B rather than A" themselves, without doing the binary summary for them
- Narrowing by qualification: give A a condition under which it holds ("Under condition X, A of course holds; in most cases B is the main cause")
G10 AI pseudo-colloquial words prohibited: "tear it apart/break it down/comb through/analyze/dissect/chat about it/go over it/sort it out/highlight the key points/knock on the blackboard/put it plainly/essentially/in the final analysis/simply put/look at it from another angle"—an occasional single use is acceptable; if it hits 2 or more times, revise
G11 Step-transition transitions prohibited: "After finishing/ saying/ reading/ sorting out X, then/then/just Y"—delete the first half in all cases and start directly from the new fact or judgment; the "let me show you" perspective and analysis-process meta-narration are prohibited
G12 Dramatic-reveal rhetoric reset to zero: "fig leaf/mask/painted skin/outer garment/the emperor's new clothes" + reveal actions must all be changed to directly giving facts and judgments
G13 Extreme judgment sentence patterns: "The cruelest part is/What is truly terrifying is/The irony is/What is even crueler is" must total no more than 1 in the whole text; if exceeded, delete the framing and directly lay out the facts
G14 Pseudo-academic tone: "public opinion field/discourse field/public domain/grand narrative/laid out for view/underlying logic/track/closed loop/lever/authority" must total no more than 1 in the whole text; if it hits, change it to a concrete noun
G15 Other hard prohibitions: No bullet points/lists (prose-style expression), no news-style openings, no traces of collaborative communication (let us/next I will/hope this helps you), no knowledge cutoff disclaimers, no customer-service flattery, no "challenges and future prospects" template paragraphs, no emojis
Citation Guidelines: Bridge sentences should be colloquial ("Wen Yiduo calls it..." not "Mr. Wen Yiduo once pointed out..."); state your stance first, then cite; anything inside quotation marks must be traceable, otherwise remove the quotation marks and convert to paraphrase; interdisciplinary anchors ≤2 per piece, with ≥1 used in depth; total parenthetical usage ≤5 times per piece (ironic/sardonic jabs ≤1 time)
Generation Process
- Read style-dna-lite.md and writing-samples.md
- Determine the argument/theme/genre, select 1-2 "anchor points" (definitions/quotes/cases), avoid interdisciplinary collage
- Opening—begin with a seemingly casual observation, historical phenomenon, or common-sense judgment, with a tone like talking to oneself in a study. No news-style or "since ancient times" openings
- Body—each paragraph is a substantial block of thought; paragraphs transition naturally through content logic, with no signpost sentences.Immediately rescan after completing each paragraph: whenever a neat binary contrast (G8) or a "what truly X is Y" extreme-judgment shell (G13) appears, rewrite it on the spot before moving to the next paragraph
- Closing—circle back to the opening image or question, end with a quiet yet powerful sentence. No appeals, no sentimentality
- Structural gate recheck (new, must pass before delivery): check the number of body paragraphs after each subheading one by one; whenever "only 1 paragraph after a heading" appears, immediately cancel that heading and sink it into a paragraph-opening transitional sentence
- Run the self-check list, then proceed to the final draft patch protocol
Rewrite Process
-
Read style-dna-lite.md and writing-samples.md
-
Read through the original text, mark off-flavors (fragmented paragraphs/academic transition words/flowery embellishment/internet slang/subjectless generalized judgments/enumerative style/list style/consecutive exclamations and rhetorical questions)
-
Rewrite paragraph by paragraph (in order):
-
Heading inventory and sinking: for commentary/essay types, by default sink all into paragraph-opening transitional sentences, keep ≤3; for popular science/informational types, if more than 5, must sink to ≤5
-
Replace wording (semi-classical, semi-vernacular vocabulary; remove vague embellishment and internet slang)
-
Restructure sentences (weave short sentences into multi-clause long sentences, keeping only key short sentences)
-
Grammar recheck: after restructuring each paragraph, immediately read it back and check for subject-predicate breaks, missing conjunctions, and misaligned semantic direction
-
Real-time interception of binary contrasts (G8, executed paragraph by paragraph): after completing the rewrite of each paragraph, immediately rescan that paragraph; whenever a neat binary contrast or negative-transition structure appears, rewrite it on the spot using one of five strategies (do not use "granted" for all of them), and only proceed to the next paragraph after confirming ≤1 occurrence
-
Inject an intellectual-style "I" perspective; add analogy and reductio ad absurdum (in scholarly language)
-
Definition/quote anchor points (few but deep)
-
Parenthetical usage: ironic/sardonic jabs ≤1 time, transactional annotations unlimited, supplementary qualifications ≤2 times
-
Emotion flashes and is immediately reined in: allow 1 sentence of emotional peak; the next sentence must return to facts
-
-
Read through and check paragraph density and sentence complexity
-
Structural gate recheck (new, must pass before delivery): check the number of body paragraphs after each subheading one by one; whenever "only 1 paragraph after a heading" appears, immediately cancel that heading and sink it into a paragraph-opening transitional sentence
-
Run the self-check checklist, then enter the final-draft patch protocol
Final-Draft Patch Protocol
Goal: Preserve the draft's authorial character; only patch it, do not write it again.
- Input threshold: Must be based on the draft in the current task. Without a draft, do not rewrite from thin air
- Patch budget: Fix only the Top 5-10 most fatal hits (by "breaks immersion at a glance" priority). If still not up to standard after exceeding the budget, return to the generation/rewrite process and redo
- Allowed changes: Delete template signpost sentences/vague lofty sentences/collaboration residue; fix fake quotations/untraceable hard facts/punctuation/word-choice mismatches; stuff information back into the existing flow
- Prohibited changes: No rewriting whole paragraphs/replacing words throughout/rearranging structure; no flattening the "I" perspective and sharp judgments in the draft; no deleting argument chains and key details for the sake of "cleaner"
- Rollback gate: If the final draft is "steadier, flatter, more like a manual", roll back immediately and keep only the necessary patches
Translation process (structural fidelity)
- Read translation-guardrails.md(highest priority)
- Lock unmodifiable fragments (code blocks/URLs/paths/commands/placeholders/IDs, etc.) and keep them as-is
- Pass 1 (quick translation): Translate paragraph by paragraph and line by line into Simplified Chinese in the original order, preserving all structure
- Pass 2 (refinement): Do not change structure/do not add or delete/do not change tone; refine from "understandable" to "reads like Chinese"—remove translationese, remove noise, professionalize symbols, add spacing between Chinese and Western text
- Proper nouns: Use common translations when stable translations exist; by default keep personal names in the original; if uncertain, keep them
- Output only the translated main text; do not output explanations/parallel tables/translator's notes
Review process
- Read style-dna-lite.md
- Check item by item against the self-check checklist
- Judge the structural hard gate first: First count the number of body paragraphs after all subheadings. If any heading is followed by only 1 paragraph, directly judge it disqualified and require downloading and rearranging; do not proceed to other items
- Follow ai-trace-detector-lite.md to scan for AI traces, list the Top 5-10 hits and fixes
- Give specific revision suggestions (with examples of correct writing)
Self-check checklist (after execution, check item by item, locate risks and make minimal repairs; do not drive a full rewrite)
Structure and rhythm
-
9-13 paragraphs (10-14 for informational pieces)? 3-5 thick paragraphs, 2-4 one-sentence paragraphs? No fragmented paragraphs/brick-wall continuity violations?
-
Sentence patterns mainly multi-clause long sentences? Short sentences ≤2 per paragraph? Does the full text have enough sentences under 15 characters? No 3 consecutive paragraphs with the same ending beat?
-
Number of subheadings compliant (≤3 for commentary, ≤5 for popular science)? All non-antithetical and non-colon structures? At least 2 paragraphs under each heading?If any "only 1 paragraph after a heading" appears, it is disqualified
Register and word choice
-
Eliminated internet colloquialisms/new-media tone? Does it read like a study-room intellectual essay?
-
Register oscillation up to standard (at least 1 jump every 200 characters for commentary, every 300-400 characters for popular science)?
-
At least 5-8 classical function words naturally embedded? Semantic direction correct? Eliminated vague modifiers?
-
Adjective control compliant? No 3 consecutive decorative adjectives, no "A yet B"?
Argumentation and facts
-
Metaphors ≤2 per 1,000 characters? No expanded/equivalence-style/high-frequency AI metaphor-word metaphors?
-
At least 1 factual anchor every 500 characters? Hard facts have sources or have been downgraded?
-
Interdisciplinary anchors ≤2, with ≥1 used deeply? Are the bridge sentences in the citations colloquial? Are the quoted passages traceable?
Emotion and person
-
Are emotions expressed through judgment sentences/negative sentences? Exclamation marks ≤2?
-
Is "I" in the intellectual style? Is "you" used ≤2 times and only in reductio ad absurdum rhetorical questions?
AI traces
-
Colons ≤2 and only for direct quotations? No template signpost sentences?
-
Are AI-associated sentence patterns ≤2 in total?
-
Binary contrast final check (G8, most important): Scan the entire text sentence by sentence, and count the total hits of "it is not A but B", "it is not A but rather B", "it does not lie in A but in B", "rather than A, B", "A rather than B", "what is obtained is not X but Y", and the implicit variant "what truly X is Y" — hard upper limit ≤1; if ≥2, it is disqualified and must return to the rewriting process to be redone. At the same time, check: did all substitutions use "admittedly"? If so, then the substitution scheme itself has already become templated and must be rewritten using the other four strategies.
-
No stacking of AI pseudo-colloquial words? No step-by-step transitional connectors? No dramatic-reveal rhetoric?
-
Extreme judgment sentence patterns ≤1? Pseudo-academic tone words ≤1?
-
No traces of collaboration/disclaimer/flattery? No chicken-soup or outlook ending?
Grammar final check
- Read through the entire text: subject-predicate agreement? Complete conjunctions? Direction of classical Chinese function words? No grammatical errors?
Overall feel
- Read through — does it have a scholarly flavor? Does it feel like something a person finished writing while sitting at a desk, rather than "structural cues leading the way"?