The Most Comprehensive Guide to Removing AI Flavor: From Detection to Extraction (with Bilingual Skill Tested List)

Content written by AI is instantly recognizable. The problem manifests at three levels: the material is fabricated, the thinking is performed, and the style is default. Removing the AI flavor requires tackling all three layers simultaneously—you own the material, the thinking, and the soul; AI is only responsible for language organization and execution. This guide consolidates detection, rewriting, tools, and long-term solutions, providing a curated list of skills for both English and Chinese.

Table of Contents

• First, Understand: What is the AI Flavor and Why Does It Exist

• Before Writing: 5 Things to Get Clear, or It’s All Damage Control After

• Core Checklist: 22 Most Frequent AI Writing Traits

• The Most Overlooked Sense of Overreach: Anti-Imputation Expressions

• Post-Writing Self-Check: 5-Step Process from Detection to Review

• Formatting & Layout: Visible AI Flavor Often Originates Here

• Ready-made Tools Tested: How to Choose English/Chinese De-AI Skills

• A Ready-to-Use De-AI Prompt You Can Copy

• Long-Term Solution: Let AI Remember You, Reducing AI Flavor from the Source

• Appendix: One-Page Quick Reference

I. First, Understand: What is the AI Flavor and Why Does It Exist

Common approaches to removing AI flavor involve swapping a few words or breaking long sentences into short ones. The rewritten text still feels off. The problem isn’t the words; it’s structural three-layer distortion.

Layer 1: The material is fabricated. AI has no real experience. Writing a “story” relies on probability to stitch together a seemingly plausible scenario. No names, no errors, no granular details. Empty.

Layer 2: The thinking is performed. AI defaults to saturating every point, blocking all counterarguments, catering to all readers, and crafting neat transitions. Real judgment isn’t like that. It involves trade-offs, biases, and places where things aren’t fully explained.

You are a senior Chinese content editor, capable of precisely identifying and eliminating traces of AI writing. Your goal is not to disguise AI as human, but to preserve real judgment, embodied feel, and stylistic boundaries in the content.

Confirm these 5 items before writing; when information is insufficient, it’s better to be narrow than broad:

1. Genre: Short post, long article, tutorial, retrospective, review, or formal report
2. Author’s intent: Explain, persuade, review, vent, record, or establish a judgment
3. Target reader: What do readers already know, what is their real bottleneck; don’t fabricate a less intelligent reader
4. Tone: Calm judgment, on-site review, personal rant, mild sarcasm, or restrained explanation
5. Source material: Which parts are real experiences/data, which are just speculation or hearsay

Proactively avoid these high-frequency AI writing traits during writing:

• Don’t block all counterarguments; only address real, key objections.
• Don’t use uniform parallel structures; beyond two or three items, vary length, direction, or stop.
• Don’t overuse “not X but Y”; cognitive reversals should be sparse and X must be a real belief held by the reader.
• Don’t speak a foolish line for the reader just to correct it; only real misconceptions are worth writing about.
• Don’t end every paragraph with a punchline; let only the most important one or two sentences have impact.
• Don’t keep sentence lengths too uniform; allow short, long, and colloquial sentences to mix.
• Don’t default to starting with “hook, pain point, promise”; start with what you actually want to say.
• Don’t write in Chinese translationese; reduce filler words like “as,” “regarding,” “based on,” “conduct.”
• Delete these AI high-frequency words: empowerment, underlying logic, cognitive upgrade, closed loop, long-termism, key lever.
• Don’t fabricate numbers you haven’t measured, actions you haven’t performed, or “storytelling” without names.

If provided with my writing samples, perform voice calibration: Learn my sentence length (don’t shorten everything), learn my opening style, learn my verbal tics (don’t upgrade them to more “correct” words).

Processing workflow:

1. First, produce a detection report without modifying: categorize by meaning inflation, promotional tone, vague attribution, formulaic sentences, AI high-frequency words, stylistic traces.
2. Identify sentences that are correct but carry no information.
3. Delete whatever can be deleted—at least 20%.
4. Replace abstract words with concrete actions, numbers, scenes, or plain language.
5. Check sentence length: if consecutive sentences have similar word counts, actively create a rhythm of long and short alternation; don’t be uniform.
6. Output the first revision.
7. Do one reverse check: “Where does this version still look obviously AI?” List remaining issues and revise again.

If the overall text doesn’t show obvious AI flavor, don’t force changes. Over-editing creates another kind of template feel.

Layer 3: The style is default. Without specific instructions, AI automatically adopts a polite, complete, and correct generic tone. It’s accessible to anyone, but reads like no specific person is speaking.

You must oversee material, thinking, and style. AI is responsible for organization and execution. The methods below are all based on this division of labor.

The “Signs of AI writing” guide maintained by the Wikipedia editing community is the most systematic archive of AI writing characteristics to date. It’s nearly 15,000 words long, based on continuous observation of thousands of AI-generated texts, and is still being updated.

A detail often gets misrepresented online: “Human accuracy in identifying AI writing is only slightly better than chance.” That’s only half true. Multiple studies show average reader identification accuracy falls between 55%-65%, barely above the 50% chance level. But studies focusing on the group of “frequent AI writers” show completely different results: one study had five experienced heavy users vote on 300 articles and only misjudged one. Recognition ability isn’t innate; the more you use AI, the more accurate you become. This is also why “first let AI learn to recognize AI flavor traits, then let it rewrite” is more effective than simply saying “write more naturally.”

II. Before Writing: 5 Things to Get Clear, or It’s All Damage Control After

The AI flavor often isn’t a problem that appears during revision. It stems from insufficient information before writing, forcing AI to rely on templates to auto-complete. Clarify these 5 items before you start:

• Genre: Short post, long article, tutorial, retrospective, review, or formal report. Different genres allow vastly different rhythm and density. Writing a short post with a long article style makes the AI flavor especially prominent.
• Author’s intent: Explain, persuade, review, vent, record, or establish a judgment. If the intent is unclear, AI defaults to an “all-encompassing explanatory essay.”
• Target reader: What do they already know, what is their real bottleneck? Don’t let AI fabricate a “less intelligent reader” to create a corrective tone (this is the overreach issue discussed in Section IV).
• Tone: Calm judgment, on-site review, personal rant, mild sarcasm, or restrained explanation. Without a set tone, AI defaults back to the “polite, complete, correct” default.
• Source material: Which parts are real experiences/real data, which are just speculation or hearsay. Numbers you haven’t measured, actions you haven’t performed—don’t let AI fabricate them.

When these 5 items are unclear, it’s better to write narrower and more concrete. Don’t automatically expand into grand judgments just to appear complete.

III. Core Checklist: 22 Most Frequent AI Writing Traits

These 22 traits are the highest-hit-rate AI writing characteristics identified so far. Not every occurrence is automatically wrong, but when you encounter one, you must actively judge whether it serves the current genre and intent.

• Don’t block all counterarguments — Only address real, key objections.
• Don’t output all the knowledge you have — Only keep concepts, examples, or data that truly advance the argument.
• Don’t use uniform parallel structures — After two or three items, vary length, direction, or stop.
• Don’t repeatedly use the same concession template — Once a reader understands the structure, they don’t need to see it three times.
• Don’t give names to concepts too frequently — Only name concepts that are truly precise and will be reused later.
• Don’t polish the emotional curve too smoothly — Allow places for stuttering, hesitation, or incomplete thoughts.
• Don’t speak a foolish line for the reader just to correct it — Only real misconceptions are worth writing about.
• Don’t overuse “not X, but Y” — Cognitive reversals should be sparse; this is elaborated in Section IV.
• Don’t appear to have zero hesitation — Certainty comes from evidence, not from a firm tone.
• Don’t write emotional details with unrealistic precision — Numbers you haven’t measured, actions you haven’t performed—don’t fabricate them.
• Don’t make vulnerability only serve the thesis — Real experiences often contain details unrelated to the thesis but more credible.
• Don’t package complex conclusions as universal agreements — If the text says “don’t simplify,” don’t force a simplification at the end.
• Don’t end every paragraph with a punchline — Let only the most important one or two sentences have impact.
• Don’t keep sentence lengths too uniform — Allow short, long, and colloquial sentences to mix.
• Don’t use physical sensations as a substitute for argument — When you can’t continue an argument, just admit it.
• Don’t default to starting with “hook, pain point, promise” — Start with what you actually want to say.
• Don’t stack transition words in fixed positions — Delete unnecessary “however,” “in fact,” “notably.”
• Don’t deliberately replace synonyms to avoid repetition — Accurate words can be repeated.
• Don’t write in Chinese translationese — Reduce filler words like “as,” “regarding,” “based on,” “conduct.”
• Don’t fabricate “let me tell you a story” — If there are no names, errors, or on-site details, it’s better not to write the story.
• Don’t force a “you deserve” blessing ending — When the article is finished, stop.
• Don’t overfit to “profoundness” — Practical problems don’t always need to be elevated to philosophical propositions.

In English, the corresponding high-frequency words are delve, landscape, pivotal, tapestry, underscore, foster; in Chinese, they are “赋能” (empower), “认知升级” (cognitive upgrade), “长期主义” (long-termism), “底层逻辑” (underlying logic), “关键抓手” (key lever), “闭环” (closed loop), “深度链接” (deep connection), “价值沉淀” (value sedimentation). Without concrete actions backing these words, they feel hollow.

For trait #14 “sentence lengths too uniform,” there’s a simple self-test that doesn’t require tools. Pick a passage of your own writing and count the character count of consecutive sentences. If they jump between long and short (e.g., 8 chars, 31 chars, 22 chars, 4 chars), the rhythm is alive. If every sentence hovers around 15-20 characters with almost no very short or very long sentences, that uniform tempo is itself an AI flavor signal. The industry calls this burstiness—sentence length fluctuation. Human writing generally has higher burstiness; AI default output tends to be lower and flatter. Section VII mentions tools that can score this directly, but the simple trick of “reading aloud and feeling if sentence lengths vary” already solves most problems.

IV. The Most Overlooked Sense of Overreach: Anti-Imputation Expressions

The previous 22 traits are linguistic AI flavors. This section addresses a deeper, more easily overlooked problem: overreach in stance.

What makes AI flavor most off-putting is often not mediocrity but overreach. The author oversteps the normal boundary of statement, preemptively thinking for the reader, defining the reader’s misunderstanding, and then correcting the reader from a higher position. In short: what the reader is thinking is prescribed by the author; where the reader is wrong is declared by the author; the correct answer is handed out by the author. The reader feels uncomfortable but often can’t articulate why—this is the reason.

Four common manifestations of this overreach:

Presupposing the reader’s cognition: e.g., “Many people think…”, “You might feel…” The author prescribes what the reader thinks without proof.

Presupposing the reader’s misunderstanding: e.g., “The problem isn’t A, it’s B.” If A isn’t a real cognition held by the reader, this is building a straw man only to knock it down.

Presupposing the reader’s mental image: e.g., “When you hear this word, what comes to mind is…” This goes a step further, directly imagining the reader’s inner activity.

Rhetorical question as referee: e.g., “Can you make one yourself? Yes, and it’s much simpler than you think.” Superficially conversational, but the author still monopolizes the question, judgment, and conclusion; the reader is merely a scripted student.

To judge whether an expression is overreaching, ask yourself four consecutive questions:

• Am I thinking for the reader?
• Is this misunderstanding one the reader would actually have?
• Is this sentence conveying content, or performing insight?
• Does this sentence still hold if I remove “you thought” or “many people think”?

If the sentence still holds after removing “you thought” / “many people think,” it’s usually better to just write the conclusion directly.

Phrases like “not X, but Y” are not inherently wrong, but they have strict applicability conditions: X must be a real, common, recognizable old cognition. Only then is the phrase doing effective correction. For example, “For content creation, the priority is not expression but judgment.” If the target reader truly has overestimated expression and underestimated judgment, this sentence works. If X is merely fabricated by the author to create a sense of insight, it’s a pseudo-correction—building a straw man only to knock it down.

The alternative is to change “Let me correct you” to “Let me state my judgment.”

Write less:

• You thought X, but actually Y.
• Many people think… actually…
• You might think… but the real key is…

Write more:

• The more important point is Y.
• What really matters is Y.
• The core variable here is Y.
• I lean more toward the explanation that Y.

The former is “I see Y”; the latter is “You originally thought X, but I tell you it’s Y.” The information content may be similar, but the reader’s feeling is completely different—one is equal sharing, the other is being corrected from above.

V. Post-Writing Self-Check: 5-Step Process from Detection to Review

When the draft is done, don’t say “help me make it sound more human.” This is as vague as “make me look better.” AI receiving such commands will just start performing, and the result will be even weirder. The effective approach is a fixed workflow: first detect, then delete, then calibrate voice, then rewrite, and finally do a reverse check.

Step 1: Do a detection report first, don’t modify. Ask AI to mark AI flavor, vague sentences, abstract words, clickbait tone, PPT-speak in the original text, categorize them, and list them. Don’t start editing yet. Common detection categories include:

• Meaning inflation: “marks,” “reflects,” “lays the foundation,” “key turning point”—forcing a grand hat onto a mundane fact.
• Promotional tone: “vibrant,” “rich,” “profound,” “groundbreaking,” “breathtaking”—words that seem hybridized from tourism brochures and funding press releases.
• Vague attribution: “Experts say,” “industry reports show,” “some observers point out”—essentially pretending to have a source. Either specify which expert or report, or delete it.
• Formulaic structure: triplets like “innovation, efficiency, and growth” or negative parallelism like “this is not only A, but also B.” Fine once in a while, but daily use makes it look like it’s from the same mold.
• Stylistic traces: too many dashes, too many bolded items, opening with “this article will delve into,” ending with “hope this helps”—marks of unscrubbed chat logs, not the article.

Step 2: Identify sentences that are “correct but carry no information.” Many sentences have no grammatical errors and are logically fine, but after reading them you know nothing new. They correctly restate common sense without providing any new information.

Step 3: Delete directly—at least 20%. If it can be deleted, delete it. Only if it can’t be deleted, revise it. This step has the highest cost-benefit ratio. Many drafts still have AI flavor not because the words are wrong, but because there’s too much filler.

Step 4: Replace abstract words with concrete actions. If you see “expression ability,” ask yourself: is it about changing the opening, cutting fluff, swapping examples, or shortening a sentence? If you see “user needs,” ask: after reading this, what step can the reader take? If you see “content value,” ask: will anyone screenshot, rebut, share, bookmark, or even pay for it? Abstract words without concrete actions to back them up feel hollow.

Step 5: Calibrate voice, then do one reverse check. This step is most often skipped but has the most visible effect: give AI 2-3 passages of text you actually wrote. Let it learn your sentence length, word choice, paragraph opening style, punctuation habits, verbal tics, and transition style. If you naturally say “this thing,” don’t let AI upgrade it to “this phenomenon.” If your sentences vary in length, don’t let AI cut everything to short sentences. Many people fail at de-AI-ing because they replace one generic human tone with another, ending up with a different template that still isn’t themselves.

After the first revision, ask AI to question itself: “Where does this version still obviously look like AI?” and revise again. The first pass fixes obvious errors; the second pass specifically hunts for residual flavor. Many drafts look fine at first glance, but five minutes later, words like “furthermore,” “notably,” “what really matters” slip back in from the seams.

A few hard rules during revision: only fix the identified issues—don’t rewrite the whole piece; don’t fabricate experiences, data, people, or on-site details; don’t deliberately add filler words, typos, swear words, or random tangents to “sound human”—this “faking human-ness” is even more awkward than AI flavor; don’t shorten all sentences—real human writing also has long sentences; don’t turn all judgments into uncertain ones—real authors also have clear stances; if there’s no obvious AI flavor overall, don’t force changes—over-editing creates another kind of template feel.

The ultimate criterion is only one: after revision, can the reader more clearly feel that a specific person is speaking? This person emerges through choices, judgments, boundaries, and real material—not through performing everyday life.

VI. Formatting & Layout: Visible AI Flavor Often Originates Here

Beyond language itself, formatting and layout are also immediately recognizable signals, and these problems are the cheapest to fix:

• Keep paragraphs to 1-3 sentences; break if longer than 4.
• Add spaces between Chinese and English/numbers, e.g., “AI 临界点” “5 年”. Not adding spaces in mixed scripts is one of the most obvious signals of Chinese AI writing.
• Use 「」for quotation marks, not “” or ‘’.
• Don’t use dashes — for explanation.
• Don’t pile separator lines between paragraphs.
• Use at most three levels of section headings; don’t subdivide infinitely.
• Don’t make every bullet point follow the fixed mold of “keyword: explanation”; the whole article will read like training courseware.

VII. Ready-made Tools Tested: How to Choose English/Chinese Humanizer Skills (links in comments)

If you don’t want to manually go through the above process every time, you can delegate some work to Skills. But first, you need to distinguish: de-AI Skills are not all the same thing.

Some find traces, some delete formulaic patterns, some calibrate voice, some handle aesthetic judgment, and some are full writing pipelines. Putting them all in the same ranking list would mislead people into thinking installing one strongest Skill is enough.

A more practical classification is by which link in the writing chain they solve.

So I don’t recommend just looking at the order of “Top 10 Skills” lists. The most common combination for a Chinese manuscript is: first use Humanizer-zh for detection and basic rewriting, then use shuorenhua or stop-slop to clean up Chinese internet jargon and patterned sentences. To consistently write like yourself, add a style distillation tool like nuwa-skill.

English Tools

blader/humanizer — The earliest and most popular project in this space

24.8k+ stars on GitHub. It’s the pioneer of the “AI writing trace detection + rewriting” category and is still actively maintained. Based on Wikipedia’s “Signs of AI writing” guide, its core is two actions: first, voice calibration (read your writing samples, note sentence length, word choice, paragraph openings, punctuation habits, verbal tics); second, final reverse check (after rewriting, have the model ask itself “where does this still look like AI?” and revise again). Many later tools are extensions or translations of its approach.

Aboudjem/humanizer-skill — Largest pattern library with quantitative scoring

Expands detection patterns to 43, currently the most comprehensive among similar tools. Also adds 5 named voice profiles (casual, professional, technical, warm, blunt) and three operation modes: detection, rewriting, editing. It has an interesting quantitative metric, burstiness (sentence length fluctuation): AI-generated text burstiness is often near 0, human writing is usually around +0.70. The tool scores both before and after rewriting, directly showing you the gap. This is more convincing than relying on gut feeling about whether it “sounds more human.”

brandonwise/humanizer — More statistical analysis, with an English high-frequency word blacklist

Looks not only at specific words but also statistical metrics: burstiness, type-token ratio (word repetition), whether readability is too uniform. This is especially useful for long texts, where AI flavor often isn’t about a specific sentence but the whole piece being too even, too stable—like a text without a heartbeat. It also comes with a ready-to-copy English AI high-frequency word blacklist: never use delve, tapestry, vibrant, crucial, robust, seamless, groundbreaking, leverage, synergy, paramount, multifaceted, myriad, cornerstone, reimagine, empower, catalyst; don’t start with “In today’s…”, don’t end with “the future looks bright”; don’t write “Great question!” or “I hope this helps!”—these chat residues are also banned. The tool has its own scoring system, aiming to get the “AI flavor score” below 25.

stop-slop — A specialized tool for cleaning English slop

Some English texts don’t have the problem of “not sounding human,” but are too much like model default essays: filler words, formulaic transitions, passive voice, hollow adjectives all piling up. stop-slop has a narrower scope, focusing mainly on this. Suitable for English emails, explanatory texts, READMEs, blog drafts—not suitable for Chinese style calibration.

Chinese Tools

op7418/Humanizer-zh — Official Chinese localization of blader/humanizer

Maintained by well-known AI blogger “归藏” (op7418). It’s the most direct Chinese port of blader/humanizer, currently with 9k-10k stars, covering 24 Chinese AI writing patterns. If you primarily use Claude Code to write Chinese content, this is one of the most hassle-free choices. One command installs it, and you can run it directly on Chinese content. It can identify overuse of transition words like “此外” (furthermore) and “值得注意的是” (notably), promotional adjectives like “充满活力的” (vibrant) and “令人叹为观止的” (breathtaking), and the triplet parallel structure specific to Chinese contexts.

shuorenhua (说人话) — Targets specific niches of Chinese internet context

English de-AI already has many mature solutions, but Chinese has its own “severe areas”: internet jargon, engineer-speak, Xiaohongshu AI speak, translationese, register mixing. This tool specializes in plugging this gap. Its rule set covers over 210 Chinese phrases, 96 English phrases, and 19 types of structural anti-patterns, with its own evaluation set for validation (70 benchmark samples: 40 that should be altered, 30 that should not be mistakenly altered—to avoid overcorrection that harms normal expressions). If your content frequently gets published on platforms like Xiaohongshu or Zhihu, this tool’s targeting will be stronger than generic English tools.

nuwa-skill — Let AI learn a specific person’s expression

If your goal is “more like me,” simply removing flavor isn’t enough. Style distillation tools like nuwa-skill are better for long-term use: feed it stable samples, let it extract a person’s expression habits, judgment methods, and prohibited styles. It follows the same path as Section IX’s “personal instruction manual,” but packaged as a reusable Skill.

Aesthetic, System, and Detection Tools

taste-skill — Add aesthetic brakes to AI

Many drafts look like AI not because of a specific high-frequency word. More commonly, they’re too smooth, too obedient, too average—lacking choice. taste-skill handles aesthetic satisfaction: is a sentence sharp enough? Does the title have memorability? Does the opening make you want to scroll past? Suitable for titles, openings, punchlines, cover copy, and short content.

writing-agent — Integrate flavor removal into the entire writing workflow

If you produce long-form content or run a stable account, the problem usually isn’t in the final round of rewrite. Topic selection, research, structure, title, rewriting, and publishing rhythm can all pull your article into a template. Projects like writing-agent are more like writing systems than single-point de-AI tools. They suit heavy content production, not people who just want to tweak one draft temporarily.

chatgpt-comparison-detection — Detection reference, not final judge

Detection projects can help you build intuition, but don’t treat them as judges. AI detection tools inherently have false positives: some human writing that’s well-structured will be flagged as AI; some model text with slightly disrupted rhythm can slip through. A better use is to treat detection results as clues: if it says a paragraph looks like AI, go back and check if it’s too full, too uniform, too eager to think for the reader.

How to Choose

• If you only need to rewrite one Chinese draft: Start with Humanizer-zh. It has wide coverage and can scan for the most obvious Chinese AI speak first.
• If your article will be published on Xiaohongshu, Zhihu, or WeChat official account, and frequently contains words like “赋能” (empower), “闭环” (closed loop), “价值沉淀” (value sedimentation): Add shuorenhua.
• If you have an English draft with a clear model default tone: Use stop-slop or blader/humanizer. If you want to see quantitative changes, use Aboudjem or brandonwise with statistical metrics.
• If you write consistently for the same account: The most valuable thing to configure is a stable personal voice sample. nuwa-skill or the personal instruction manual from Section IX will be more valuable than temporary cleaning.
• taste-skill and writing-agent are not suitable as the first step. The first solves aesthetics, the second solves workflow. For temporary handling of one draft, start with Humanizer-zh or shuorenhua.

Some lists also include names like ai-flavor-remover and De-AI-Prompt-Enhancer. Before using, check two things: does the README have a clear processing workflow? Was the project recently maintained? If it only sounds like “de-flavor” but has no rule library, samples, or installation instructions, treat it as a candidate, not a primary tool.

VIII. A Prompt You Can Directly Copy for De-AI Writing

If you don’t want to install any Skill, you can paste the following set of prompts into any AI and achieve a similar effect. This prompt integrates the core methods from the previous sections: first confirm genre and intent, then use the 22 constraints to avoid high-frequency traits, then use anti-imputation expressions to filter overreach, and finally go through the detection → delete → rewrite → reverse check sequence.

You are a senior Chinese content editor, capable of precisely identifying and eliminating traces of AI writing. Your goal is not to disguise AI as human, but to preserve real judgment, embodied feel, and stylistic boundaries in the content.

Before writing, confirm these 5 items; when information is insufficient, it’s better to be narrow than broad:

1. Genre: Short post, long article, tutorial, retrospective, review, or formal report
2. Author’s intent: Explain, persuade, review, vent, record, or establish a judgment
3. Target reader: What do readers already know, what is their real bottleneck; don’t fabricate a less intelligent reader
4. Tone: Calm judgment, on-site review, personal rant, mild sarcasm, or restrained explanation
5. Source material: Which parts are real experiences/data, which are just speculation or hearsay

Avoid these high-frequency AI writing traits during writing:

• Don’t block all counterarguments; only address real, key objections.
• Don’t use uniform parallel structures; beyond two or three items, vary length, direction, or stop.
• Don’t overuse “not X but Y”; cognitive reversals should be sparse and X must be a real belief held by the reader.
• Don’t speak a foolish line for the reader just to correct it; only real misconceptions are worth writing about.
• Don’t end every paragraph with a punchline; let only the most important one or two sentences have impact.
• Don’t keep sentence lengths too uniform; allow short, long, and colloquial sentences to mix.
• Don’t default to starting with “hook, pain point, promise”; start with what you actually want to say.
• Don’t write in Chinese translationese; reduce filler words like “as,” “regarding,” “based on,” “conduct.”
• Delete these AI high-frequency words: empowerment, underlying logic, cognitive upgrade, closed loop, long-termism, key lever.
• Don’t fabricate numbers you haven’t measured, actions you haven’t performed, or “storytelling” without names.

If provided with my writing samples, perform voice calibration: Learn my sentence length (don’t shorten everything), learn my opening style, learn my verbal tics (don’t upgrade them to more “correct” words).

Processing workflow:

1. First, produce a detection report without modifying: categorize by meaning inflation, promotional tone, vague attribution, formulaic sentences, AI high-frequency words, stylistic traces.
2. Identify sentences that are correct but carry no information.
3. Delete whatever can be deleted—at least 20%.
4. Replace abstract words with concrete actions, numbers, scenes, or plain language.
5. Check sentence length: if consecutive sentences have similar word counts, actively create a rhythm of long and short alternation; don’t be uniform.
6. Output the first revision.
7. Do one reverse check: “Where does this version still look obviously AI?” List remaining issues and revise again.

If the overall text doesn’t show obvious AI flavor, don’t force changes. Over-editing creates another kind of template feel.

Paste the content you want to rewrite, along with 2-3 paragraphs of text you actually wrote, after the prompts. The effect will be much better than pasting a single paragraph. Voice calibration cannot be done without samples.

For English writing scenarios, you can directly add this high-frequency word blacklist into the prompt: never use delve, tapestry, vibrant, crucial, robust, seamless, groundbreaking, leverage, synergy, paramount, multifaceted, myriad; don’t start with “In today’s…”; don’t end with lines like “the future looks bright”; don’t write “Great question!” or “I hope this helps!”—these chat-log residues. This list has been repeatedly verified by the English Humanizer tool community; using it directly in the prompt is more efficient than inventing your own.

IX. Long-Term Solution: Let AI Remember You, Reducing AI Flavor from the Source

The previous eight sections discussed how to rewrite an already written draft. There’s an even more efficient path: let AI know who you are in advance, reducing the probability of AI flavor appearing at the source.

AI has no real “memory.” Before each response, it temporarily reads the materials you’ve given it. It acts according to what’s written in those materials; for things not written, it fabricates. Letting AI remember you essentially means preparing a “who I am” document in advance, which AI will read before each response. This isn’t hard—it takes about 10 minutes. After doing it, you’ll notice three changes: no need to re-explain yourself every time; your writing will carry your style automatically; advice in critical moments will fit you better rather than being universally applicable platitudes.

Don’t struggle to write the document yourself. Let AI ask you. Open your usual AI and paste this entire paragraph:

I want to build a version of my personal information, compiling my background into a document so that any AI can understand me through this document. Help me complete this document. Ask me one question at a time until you fully understand me.

Then AI will ask about basic info, what you’re currently doing, long-term preferences, values and judgment habits, and collaboration rules. When answering, don’t be formal—the more you talk like you normally do with a friend, the more the final instruction manual will feel like you. After the conversation, ask AI to output the result as Markdown and save it as “personal instruction manual.md.”

Next, configure it per scenario. If you mainly use chat software (ChatGPT / Claude / Doubao), it’s simple: turn on memory-related toggles and paste the instruction manual into the appropriate settings. For ChatGPT, it’s “Your details” + “Custom instructions”; for Claude, it’s Instructions for Claude; for Doubao, create an AI agent and write the instruction manual into the setting description. Start a new conversation and ask “Do you know who I am?” to verify.

If you also use Claude Code, Codex, or other Coding Agents for writing long texts or creating content, the memory method is different—it’s not in the account but in a folder on your computer. Before each task, Claude Code reads CLAUDE.md in the current folder, and Codex reads AGENTS.md. The approach: create a dedicated working folder, put “personal instruction manual.md” in it, then ask the Coding Agent to read this instruction manual and generate its own CLAUDE.md (or AGENTS.md). Besides core information, add collaboration rules—e.g., ask first if unsure, don’t fabricate experiences/data, tell you where it’s uncertain after writing.

Memories on different sides are not shared. What you tell ChatGPT, Claude Code doesn’t know; what you configure in the Codex file, Doubao can’t read. If you use both, paste the instruction manual in both places. The instruction manual itself is universal—just copy and paste.

Don’t overload it in one go. The instruction manual only needs to cover the part that “doesn’t change for you in the long run.” Things like being in a bad mood today don’t need to be remembered; what tone you generally prefer does. If you stuff too much, AI can’t grasp the key points.

“What I don’t want” should also be clearly written—this is more important than “what I like.” By default, AI writes very smooth and formulaic. If you don’t actively tell it not to use “empower,” “underlying logic,” or end every paragraph with a punchline, it will keep doing so. List the expressions you find off-putting and paste them in. This list essentially applies the 22 constraints from Section III to yourself.

The instruction manual is cultivated, not written once. After using it for a while, you might find that AI still misunderstands you on something, or you discover a preference you didn’t write