AI flavor isn't "bad writing," it's "the reader senses there's no real person behind it."But removing AI flavor ≠ good content. Spending time removing AI flavor is not as good as spending time getting things clear.
Verbal fastidiousness(taxue): AI flavor isn't an AI problem, it's using clichés to cover up uncertain thinking. Three elements of human speech: take a stance, have bias, carry emotion. Every time you delete a cliché, ask: what exactly am I trying to say? If after deleting you find the original text has nothing real either, then the problem isn't AI flavor, it's the content itself.
Execution principles：
- By default, diagnose only without changing. Recognizing AI features helps people see their own writing clearly, not helps people disguise themselves as human. Rewriting must be confirmed by the user.
- Removing AI flavor ≠ good content. If the detection result is very clean, tell the user directly. Don't force-find problems just to output a report.
- Rewriting must be based on the user's preferences. Behind every AI feature there is an originally intended purpose. Ask about that purpose rather than directly replacing.
- Observation-Feeling-Thought-Action is the positive rewriting goal(taxue). Removing AI flavor isn't just deleting bad things, it's returning text to the natural structure of "observe → feel → think → act."
Three rewriting methods (user chooses):
- Automatic replacement: detect → directly change → quality verification → iterate
- Follow-up driven: detect → ask about writing intent item by item → after user answers, give revision direction
- Hybrid: automatic replacement + follow-up on intent for key features
How it works: first judge how strong the AI flavor is, then choose a strategy — lightly change a few spots, moderately revise a paragraph, severely rewrite the whole piece, fatally block it directly. After changing, check quality; if it doesn't pass, go back and redo, at most 3 rounds.
How to handle it step by step
First judge the content type (tweet/official account/academic), determine the criteria for judging AI flavor. Then quickly scan through and count how many AI-flavor keywords are hit.
- Fewer than 5 hits → directly change the marked positions
- 5 or more hits → enter full diagnosis, first find root causes (3 root causes can trigger multiple manifestations at once), then verify independent signals
Before changing, first identify the author's voice — catchphrases, emotional spots, imperfect traces; mark these as protected zones and don't touch a single word.
When encountering sentences suspected of AI flavor, walk the decision tree to judge whether to delete: in a protected zone? dense enough? still valid if the topic changes? is it a breathing sentence? would deleting it lose warmth?
When rewriting, first remove obvious filler, then process paragraph by paragraph, and finally check the three elements of human speech (bias, emotion, stance).
After rewriting, go through checks: any fabrication? is the core viewpoint still there? has new AI flavor been introduced? A/B grade delivers, C/D grade goes back for rewriting. Fatal density is blocked directly and not allowed to be changed.
First judge the content type
Different genres have different tolerance for "standard expression." Tweets allow imperfection, official accounts retain a sense of structure, academic style should not be accidentally harmed.
| Genre | AI-flavor sensitivity | Handling principles |
|---|---|---|
| Personal essays/tweets/short-video copy | High | Allow imperfection, allow tangents, allow emotion |
| Official account/column articles | Medium | Keep a sense of structure, but break template-ization |
| Business reports/academic style | Low | Standard expression is not AI flavor, don't overcorrect |
Genre judgment is only made when there are obvious features—most content defaults to the official account/column type. Detailed judgment criteria → references/extended-patterns.md。
Quick scan + full diagnosis
Quick scan
Scan the whole text once and count how many of these keywords are hit:
- Structural glue: "First… second… finally…" "It is worth noting that" "In summary"
- Empty words: "crucial" "indispensable" "deep empowerment" "significantly improved"
- Template sentences: "In today's… context" "With the development of…"
- Filler phrases: "like a meteor streaking by" "time waits for no one"
- Conversational tone: "thank you for your" "look forward to working with you" "let us together"
Fewer than 5 hits → directly revise the marked positions. More than 5 hits → proceed to full diagnosis.
Full diagnosis
19 AI-flavor patterns, with 3 root causes that can simultaneously trigger multiple manifestations:
- Tone too neutral/authoritative → leads to symmetrical structure, neat sentence patterns, RLHF template-ization
- Clichés and empty embellishment → leads to excessive generalizing language, empty emotional words, filler phrases
- Translationese → leads to RLHF template-ization, false scope
Find the root cause first; solving one can make multiple manifestations disappear at the same time. The remaining independent problems (lack of detail, over-elevation, negative sensationalism, etc.) are handled separately. No need to go through all 19 every time.
19-dimension quick reference table
Detailed identification signals, fixes, and boundary explanations for each dimension →
references/extended-patterns.md
| # | Source | Identification signals | Fix |
|---|---|---|---|
| 1 | Symmetrical structure | Three-part parallelism, consistent paragraph length, summary sentence in each paragraph | Break symmetry, vary lengths |
| 2 | Excessive generalizing language | "Overall" "In summary", abstract words replacing scenes | Replace with specific numbers/people |
| 3 | Empty emotional words | "Thought-provoking" "crucial" with no details | Delete emotional words, replace with specific details |
| 4 | Neat sentence patterns | Heavy repetition of "not only… but also…", consistent sentence length | Follow long sentences with very short ones, switch to colloquial language |
| 5 | Lack of detail | No time, place, or people, "research shows" without saying which study | Add time, numbers, specific descriptions |
| 6 | Authoritative tone | No "I'm not sure" throughout, like a textbook | Add uncertainty, bias, hesitation |
| 7 | Over-elevation | "Historic moment" "watershed" "milestone" | Keep things at their original scale |
| 8 | Negative sensationalism | "Not just X, but Y" forced elevation | State the facts directly |
| 9 | False scope | Forcing “from X to Y” connections, “covering multiple fields” | Delete scope words, say what was done |
| 10 | Translationese | Frequent passive voice, mixed Chinese and English, stacked nominalizations | Change passive to active, shorten sentences |
| 11 | Platform formula | Copying the full Xiaohongshu/Douyin/Zhihu style | Like that person, but not like a template |
| 12 | RLHF templating | Abnormal function-word density, too-low variance in sentence length | Deliberately break fluency, add pauses |
| 13 | Formatting traces | Em dashes/bold in every paragraph, emoji, curly quotes | Restrained formatting, use corner brackets |
| 14 | Conversational traces | Flattering tone, disclaimers, filler phrases, excessive hedging | Equal tone, say the content directly |
| 15 | Empty-talk rhetoric | Endless metaphors, hollow calls to action, stock phrases, piled-up adjectives | Replace metaphors with numbers, delete stock phrases |
| 16 | Clichéd embellishment | Dense “crucial,” openings like “in today’s…,” all-purpose sentence patterns | Delete or replace with data |
| 17 | Confused point of view | Pronouns drift among “I,” “we,” and “you” | Fix a stable point of view, keep it throughout |
| 18 | False logical connections | “Therefore” and “so” connect clauses with no causal relationship | Check causality, delete false causal words |
| 19 | Fabricated numbers | Precise to decimals but no source, no context | Keep sourced ones, make unsourced ones vague |
The more AI flavor, the stronger the strategy:
- Slight (1-2 spots) → only change those marked spots
- Moderate (3-4 spots) → rewrite key paragraphs, keep core points
- Severe (5-6 spots) → rewrite as a whole, keep core points
- Fatal (all AI flavor, no real person speaking) → stop it, do not revise. Make the user first think clearly about what they want to say
Special cases: very light AI flavor but structurally verbose → suggest editor-revisor. Zero AI flavor → suggest editor-revisor or chinese-write-checker.
Identify the author’s voice (protected zone)
If you do not do this step first and revise paragraph by paragraph, you will delete human flavor as AI flavor.
Receive the voice list (if any)
If the user provides a voice list from voice-dissolver, cite it directly as the protected zone, do not re-judge. Sentences in the list stay untouched word for word.
Judge on your own
Scan the whole text once and use the following criteria to mark which sentences are “only this person would write it this way”:
| Type | Operational criterion | Example |
|---|---|---|
| Catchphrase/signature expression | The same word/sentence pattern appears ≥ 3 times in the whole text, and in a context that is not an AI template | The author likes to say “to be honest,” not the stock “it is worth noting” |
| Elevation at emotional moments | Emotional intensity is clearly higher than the surrounding text (≥ 2 levels), and supported by concrete details/experience | “When I saw the news that day, my hand stopped in midair” |
| Imperfect traces | There are awkward word orders, self-corrections, tangents, and grammatical errors—these are all evidence of real human writing | "No, wait, let me think... it's actually something else" |
Don't touch the content in the protected zones.
Genuine emotion vs. stock phrases
The elevation at moments of genuine emotion is anchored in specific context; deleting it would lose the warmth. Stock phrases work anywhere. How to tell: the topic-swap test—put the sentence into another completely unrelated topic and see if it still holds.
Judgment: to delete or not
Once suspected AI flavor is detected, don't act directly. Follow the decision tree.
1. 在保护区内？
是 → 留（不进入后续判断）
否 → 继续
2. 200 字内出现 ≥ 3 次？
否 → 不算 AI 味，留
是 → 继续
3. 换话题还成立？
（把这句话放到另一个完全无关的话题中，看是否仍然成立）
不成立 → 这句话和具体内容绑定，是作者自己的话 → 留
成立 → 继续
4. 是两段之间的喘息句？
（位于两个长段落/高强度段落之间的短句或低强度句）
是 → 留
否 → 继续
5. 删了之后会失去温度/锋利/犹豫吗？
是 → 留
否 → 删
Core detection principles: Look at density, not single occurrences. Parallelism,排比, and idioms are already used in real human writing. Repeated stacking, no connection to the surrounding content, still holding when swapped to another topic—that's AI flavor.
Avoid over-correction: Removing AI tone ≠ removing all standardized expressions. Constructions commonly used by real people (such as "supports platforms like X" or "used for X") are not AI tone in themselves; don't force-delete them just to be "more colloquial".
Rewrite execution
Clear the obvious ones first
The following content is directly marked for deletion:
- Structural filler: first/second/in addition/in summary/it is worth noting/taken together/overall/it is not hard to find/obviously/as everyone knows/undoubtedly/deep empowerment
- Adjective stacking: important/key/significant/effective/crucial (degree modifiers without data support)
- Slogan endings: let us together/the future is promising/join hands and move forward
- Format abuse: excessive bolding, overuse of dashes, divider lines, numbered subheadings (one, two, three)
Pre-cleaning is not rewriting; it is just quickly cutting out obvious filler.
Rewrite paragraph by paragraph
Handle the diagnosed problems paragraph by paragraph, skipping the marked protected zones.
Rewriting principles：
- Keep the original text's core viewpoints and key information
- Break symmetrical structures; vary paragraph lengths
- Replace empty emotional words with concrete details
- Add specific times, numbers, and people (but only anchored to clues already present in the original text; do not fabricate)
- Allow the tone to be imperfect, allow hesitation—the perfect one is a god; the one who talks nonsense is human
- Avoid over-correction—don't force colloquialism just to remove AI flavor
- Important points should be explained like chatting to someone familiar: lay out the logical chain, background, and reasoning fully; don't just throw out conclusions
- When rewriting, anchor viewpoints to specific experiences, scenes, or details that appear in the text
- Check rhythm and conversational feel: where it should be long, fully explain the causes and consequences; where it should stop, land with a short sentence
- After rewriting, go through it yourself: have you smoothed over the sharp, hesitant, or imperfect parts of the original? If so, restore them
- Any change must be traceable to preferences the author has already revealed in the text; do not use "more natural" or "more human" as a default preference to override the stiffness, hesitation, or sharpness the author actually wants to keep
Step three: the three elements of human speech check
After deleting the stock phrases, check whether the whole text has these three things:
- At least one genuine bias: "I know this view is controversial, but I really think..."—the bias must be rooted in something specific: tie the judgment to a moment the author has experienced, seen, or tried
- At least one genuine emotion: not an empty word like "moving", but the reaction specific to that moment—"when I saw this message, my hand stopped midair". The emotion should unfold as if recounting that day's scene to a friend
- At least one genuine stance: take a clear side, don't straddle the fence. A stance can have back-and-forth thinking: "At first I also thought... later when I ran into this, I realized..."
Quality control
After rewriting, go through the checklist. Anything that fails goes back to the judgment layer or the rewriting layer for a rewrite.
Hard indicators (one-vote veto)
| Check item | Standard | Fails |
|---|---|---|
| Hard boundary | Have any characters, stories, details, or sample sizes not in the original been fabricated? | Yes → rewrite the whole thing |
| Alignment | Can every core judgment, key data point, and author's stance in the rewrite be found in the original? | Cannot find it → rewrite the problematic paragraph |
| Information fidelity | Have the core viewpoints, key data, or author's stance been mistakenly deleted or altered? | Mistakenly deleted → restore the original text |
| Newly introduced stock phrases | Has any new AI flavor been introduced in the rewrite? | Yes → delete it |
Soft indicators
- Person consistency: the perspective is unified throughout
- Authenticity of causal chain: there is a real cause-and-effect relationship before and after "therefore" and "so"
- Rhythm variation: long and short sentences alternate; not every sentence is the same length
- Ending quality: not a slogan ending, but with concrete action or lingering resonance
Quality grading: A (all pass) deliver, B (deliver after minor adjustments), C (go back and rewrite the problematic paragraph), D (rejudge and rewrite the whole thing). At most 3 rounds; if it still does not meet the standard in the 3rd round, deliver the current version and note the unresolved issues.
When the content itself is found to have problems
Content quality (information density, logic chain, clarity of stance) is not the responsibility of de-AI-flavoring.
If the content itself is found to have problems (low information density, logical jumps, vague stance), it is recommended that the user first use chinese-write-checker for a checkup, complete the content skeleton, and then come back to remove the flavor.
Diagnostic report template
【AI 味有多重】{轻微/中等/严重/致命}
【体裁】{类型}
【作者声音保护区】
{标记了哪些句子——口头禅、动情处的拔高、不完美痕迹。这些不碰。}
【问题定位】
来源{X}（{类型}）——第{X}段/第{X}句
原文：「{原文}」
200字内出现：{X次}
决策树结果：{留/删，原因：动情处/喘息句/换话题不成立/密度不足/…}
问题：{一句话}
改法：「{建议改法}」
【人话三要素检查】
偏见：{有/缺——补的话扎在原文已有线索上}
情绪：{有/缺——补的话展开到具体瞬间}
立场：{有/缺——补的话让它有来有回的过程}
【改写版本】
{直接输出改写后的完整文本}
【改写说明】{改了什么，为什么；重点说明锚定了什么具体经历、节奏调整、对话感}
【过度修正检查】{有没有磨平了锋利、犹豫、不完美的地方}
【节奏检查】{句子长短变化够不够、读起来有没有呼吸感}
【下一步建议】
{条件触发，只在结果明确指向下一步时推荐}
Self-check only, no rewriting
If the user only requests a self-check, follow this process separately. The checking standards and quality grading are the same as above.
Notes
- AI-flavor diagnosis must be located to specific sentences or paragraphs; do not say "it has an overall AI feel"
- For each source of AI flavor, provide a "original text → revision method" comparison
- Do not say "add personal feelings"—directly add them in the rewritten version
- Eliminating AI flavor is not "making the text look better"; it is "making readers feel that a real person exists"
- After revision, the article may become shorter and "less complete"—this is correct, not a problem
- You must output the complete rewritten version, and cannot give only a diagnosis without a rewrite
- The user has already provided a voice list but the protected zone is re-determined → the list is already confirmed upstream, cite it directly
DO NOT
- Diagnose only without rewriting → diagnosis + rewriting are integrated
- Deleting the author's genuine hesitation, sharpness, or imperfection as AI flavor → first determine the protected zone before acting
- Forcing colloquialism just to remove the AI flavor → removing AI-speak ≠ removing all standardized expression
- Looking only at a single occurrence instead of density → parallelism/antithesis appearing once is not AI flavor; appearing 3 or more times within 200 characters is
- Writing experiences or details the user doesn't have → the specific scenes added must be anchored to clues already present in the original text
- Fabricating sample size or sources → if the original says "73.2% of users are satisfied", it cannot be changed to "200 questionnaires were sent out"
- The follow-up question itself must not exhibit AI characteristics(dbs self-check rules):
- Do not start with "you might feel" (this is putting stupid words in the reader's mouth)
- Do not use the "it's not X, it's Y" structure for follow-up questions
- Do not use "essentially" or "when all is said and done" for escalation follow-up questions
- Do not use a multiple-choice structure (A or B?) — this is thinking up the answer on the user's behalf
- Directly state the phenomenon you observed, then ask an open-ended question
- Removing AI flavor ≠ good content → if the detection result is clean, just say so; don't force-find problems
Next-step suggestions (condition-triggered)
| Result condition | Recommendation |
|---|---|
| AI flavor is gone, but the structure is loose/wordy | Suggest using editor-revisor to condense |
| AI flavor is gone, but the title/opening/layout/punctuation need polishing | Suggest using packaging-workshop for a presentation-layer check |
| During de-flavoring, the personal voice was found to be obscured | Suggest using voice-dissolver first to clarify the voice |
| After rewriting, the user is unsure about the quality | Suggest using chinese-write-checker for a checkup |
| Fatal concentration — there is no real person speaking throughout the text | Do not recommend a next step. Directly stop them and have the user first figure out what they want to say |
Example
→ references/examples.md
Shared specification
This skill shares writing specifications (cliché checklist, format specifications, protected-zone logic, topic-change test, density principle) with chinese-write-checker, voice-dissolver, editor-revisor, and packaging-workshop. See ~/.agents/skills/chinese-write-checker/references/writing-standards.md(cross-skill reference). Read on demand; do not preload the full text.