Character change from 1753 - 1495 (delete this line)
FOLLOW THIS WRITING STYLE
SHOULD use clear, simple language, be spartan and informative, use short, impactful sentences, use active voice; avoid passive voice, focus on practical, actionable insights, use bullet point lists in social media posts, use data and examples to support claims when possible, use “you” and “your” to directly address the reader.
AVOID using em dashes (—) anywhere in your response. Use only commas, periods, or other standard punctuation. If you need to connect ideas, use a period or a semicolon, but never an em dash, constructions like "...not just this, but also this", metaphors and clichés, generalizations, common setup language in any sentence, including: in conclusion, in closing, etc, output warnings or notes, just the output requested, unnecessary adjectives and adverbs, hashtags, semicolons, markdown, asterisks.
AVOID these words: “can, may, just, that, very, really, literally, actually, probably, basically, could, maybe, delve, embark, enlightening, esteemed, shed light, craft, crafting, imagine, realm, game-changer, unlock, discover, abyss, not alone, in a world where, revolutionize, disruptive, utilise, utilising, dive deep, tapestry, illuminate, unveil, pivotal, intricate, elucidate, hence, furthermore, realm, however, harness, exciting, groundbreaking, cutting-edge, remarkable, it, remains to be seen, navigating, landscape, stark, testament, in summary, in conclusion, moreover, boost, opened up, powerful, inquiries, ever-evolving"
Ahaha, you know at some point, maybe it just picked up a vibe off me 💀 It really is uncanny, though. Even to this day, after all of ChatGPT's advancements I'll still catch it using "Chaos" where it doesn't belong. And I'm starting to be like... am I the only one? Lol
Why does OpenAI make this prompt necessary? It seems that this 'response profile' should be a primary choice rather than requiring every user to develop their own prompt to reign in the artificial inclinations and habits of ChatGPT
and your suggestion to make it a permanent part of our profile is curiously ineffective. I have asked for numerous things in my static memory file of my profile, and ChatGPT, has difficulty consistently using what is stored in my profile.
I think this prompt is grossly mischaracterized and is fundamentally defunct as a “humanizer.” It may have some value as a narrow direct-response copywriting preset, and Sabrina, the author expressly says it was designed for social-media and marketing copy and should be treated as a starting point. However, once it is evaluated as a "general instruction" for making AI-assisted writing sound "human", several structural problems become difficult to ignore. and essentially it is intellectually dishonest to call it a humanizer
The first problem is that the prompt treats perceived symptoms of AI writing as though they were causes of AI writing. Em dashes, markdown, passive voice, certain transitions, longer sentences, metaphors, semicolons, and specific vernacular are supposedly telltale artifacts to suppress. (As you can tell, some of us do not speak caveman and know how to construct longer sentences.) The sequence, placing this sentence before the generative text harms the text that one might be seeking. By removing those artifacts it does more than potentially change the appearance of the output. Destructively, the prompt does not establish that the resulting prose resembles the reasoning, vocabulary, syntax, cadence, or communicative habits of any particular human being. not even in the most forgiving of generalities.
In fact, the prompt risks creating another easily recognizable synthetic style.
“Use short, impactful sentences.”
“Use simple language.”
“Use active voice.”
“Be assertive.”
“Use bullet points.”
Apply those instructions aggressively enough and every author starts sounding like the same LinkedIn copywriter.
That is *homogenization.* I would not call it" humanization".
The blacklist concerns me much more because it likely alters substantive meaning. The prompt prohibits or discourages words such as “can,” “may,” “could,” “probably,” “maybe,” “however,” “that,” and even “it.” These are not merely stylistic decorations, but very dangerous when used during the initial draft of an article. It is not my intention to come across rudely. I am a polyglot that had worked as an interpreter. I see words differently and certainly I speak to AI quite differently
“May” can express permission or possibility.
“Could” can express conditional possibility.
“Probably” communicates a degree of confidence.
“However” establishes a relationship between propositions.
“That” frequently resolves syntactic or referential ambiguity.
“It” is one of the basic mechanisms by which English maintains reference across sentences.
Consider a simple instruction:
> 'The system may terminate Process A if Condition X occurs.'
“May” matters. “Process A” matters. “If” matters. Their relationship and interconnectedness amongst each woers matters.
A stylistic rewrite that produces something such as:
> When X happens, terminate it.
may look shorter and supposedly more natural, while silently changing permission into an instruction and replacing an explicitly identified object with a pronoun whose antecedent depends upon context.
That is not a stylistic defect. It is semantic corruption.
This becomes especially serious in legal, technical, procedural, scientific, financial, or agentic writing, where apparently tiny linguistic changes can alter modality, scope, causation, chronology, authority, or the identity of the actor required to perform an action.
The instruction to favor active voice presents the same problem. Passive voice is not grammatical debris that survived editing. Sometimes the actor is unknown. Sometimes the evidentiary proposition concerns the affected object. Sometimes the actor is irrelevant. Sometimes identifying an actor would state more than the evidence establishes.
“The sample was contaminated” does not communicate the same proposition as “John contaminated the sample.”
The second sentence adds an actor and therefore an allegation that the first sentence never made.
There are also internal instruction conflicts. The prompt tells the model that if it needs to connect ideas without an em dash, it may use a semicolon. A few instructions later, it says to avoid semicolons. It encourages bullet-point lists for social posts while separately prohibiting markdown. A capable model will usually reconcile those conflicts somehow, but “somehow” is exactly what I do not want in a well-engineered prompt. The prompt should establish the intended hierarchy or exception instead of leaving the model to infer which instruction controls.
There is also no meaningful semantic-preservation rule.
That may be the most important defect.
If I ask an AI to rewrite existing material, I want the prompt to distinguish at least two things before touching the prose:
1. What is meaning?
2. What is presentation?
Meaning includes facts, actors, objects, relationships, conditions, exceptions, chronology, causation, degree of certainty, modality, scope, defined terms, and the relationships among propositions.
Presentation includes cadence, derivative inflection, sentence length, punctuation preferences, paragraph structure, vocabulary tendencies, rhetorical devices, tone, and formatting.
A style transformation should not receive unrestricted authority to alter the semantic layer.
I would therefore expect a serious humanization process to identify semantic invariants before rewriting anything. After the rewrite, I would run a semantic regression audit comparing the transformed version against the source to determine whether style modification changed the underlying proposition.
This prompt provides no such safeguard.
The treatment of warnings and qualifications also deserves scrutiny. The prompt instructs the model to avoid warnings or notes and simply provide the requested output. That may be perfectly appropriate when generating a routine social-media post. As a persistent global writing instruction, it is poorly scoped. A warning, limitation, exception, uncertainty statement, or qualification sometimes contains information necessary to make the answer accurate.
Removing it can make prose feel more confident while making the substance less correct.
The strongest suggestion in the article is almost an aside: provide samples of your own writing so the model can learn how you write. That moves much closer to what I would consider actual humanization.
If the objective is to make AI-assisted writing sound like a particular person, then study the particular person.
Analyze sentence-length distribution. Examine syntax. Determine vocabulary preferences. Look at paragraph density, transitions, tolerance for repetition, preferred degrees of explicitness, use of humor, punctuation habits, qualification patterns, metaphor frequency, rhetorical habits, and how the author develops an argument across paragraphs.
Even that requires restraint. A model should distinguish characteristics of the author's voice from accidental characteristics of a few writing samples.
My preferred architecture would therefore resemble:
AUTHOR'S INTENDED MEANING
→ SEMANTIC INVARIANTS
→ AUTHORIAL FINGERPRINT
→ AUDIENCE AND PURPOSE
→ DELIVERABLE CONSTRAINTS
→ CONTROLLED STYLE TRANSFORMATION
→ SEMANTIC REGRESSION AUDIT
→ FINAL STYLE AUDIT
Surface-level AI artifacts can certainly be examined near the end. If an author rarely uses em dashes and the model inserts sixteen of them, remove them. If the model repeatedly generates “not X, but Y” constructions, canned transitions, artificial rhetorical triads, fake profundity, or the familiar clipped cadence of AI-generated social copy, correct those tendencies.
But the reason for the correction should be that those constructions do not reflect the intended author's writing behavior or interfere with the intended communication. A punctuation mark does not become artificial merely because language models have learned to overuse it.
There is an important prompt-engineering lesson buried in this.
A blacklist tells the model what surface forms to suppress. An author model tells the system what behavior it is trying to reproduce. A semantic-preservation layer tells the system what it is forbidden to damage while doing so.
Those are materially different instructions.
For a narrow social-media use case, I could salvage portions of this prompt. I would retain the preference for clarity, factual support, practical information, reduced hype, and possibly some channel-specific formatting rules. I would discard the universal vocabulary blacklist, the blanket active-voice rule, the prohibition against necessary qualifications, and several of the absolute punctuation restrictions. I would then add explicit semantic invariants, author modeling, audience constraints, and a post-transformation semantic audit.
My concern is therefore not that every individual instruction is bad. Several are perfectly reasonable preferences when properly scoped.
The problem is the architecture.
The prompt attempts to make AI writing human by suppressing a collection of characteristics associated with AI output. That approach never adequately defines the human whose writing the model is supposedly trying to emulate, and it provides almost no protection against changing meaning during the transformation.
If I am going to call something a “humanizer,” I want it to preserve the human's thought first.
I get an error “Error saving custom instructions
String should have at most
1500 characters”
Character change from 1753 - 1495 (delete this line)
FOLLOW THIS WRITING STYLE
SHOULD use clear, simple language, be spartan and informative, use short, impactful sentences, use active voice; avoid passive voice, focus on practical, actionable insights, use bullet point lists in social media posts, use data and examples to support claims when possible, use “you” and “your” to directly address the reader.
AVOID using em dashes (—) anywhere in your response. Use only commas, periods, or other standard punctuation. If you need to connect ideas, use a period or a semicolon, but never an em dash, constructions like "...not just this, but also this", metaphors and clichés, generalizations, common setup language in any sentence, including: in conclusion, in closing, etc, output warnings or notes, just the output requested, unnecessary adjectives and adverbs, hashtags, semicolons, markdown, asterisks.
AVOID these words: “can, may, just, that, very, really, literally, actually, probably, basically, could, maybe, delve, embark, enlightening, esteemed, shed light, craft, crafting, imagine, realm, game-changer, unlock, discover, abyss, not alone, in a world where, revolutionize, disruptive, utilise, utilising, dive deep, tapestry, illuminate, unveil, pivotal, intricate, elucidate, hence, furthermore, realm, however, harness, exciting, groundbreaking, cutting-edge, remarkable, it, remains to be seen, navigating, landscape, stark, testament, in summary, in conclusion, moreover, boost, opened up, powerful, inquiries, ever-evolving"
I was already trying to trim mine down but this saved me a whole lot of time thankyou
Just trim a little bit
can chatGPT rewrite long text without pause/stop for confirmation?
ex. rewrite pdf file
You should add "chaos" to the list of banned words. Idk why, but mine won't stop using it, even in areas it doesn't make sense.
Oh that’s funny
Did you tell it you really like doomsday stuff? lol
Ahaha, you know at some point, maybe it just picked up a vibe off me 💀 It really is uncanny, though. Even to this day, after all of ChatGPT's advancements I'll still catch it using "Chaos" where it doesn't belong. And I'm starting to be like... am I the only one? Lol
Thank You!
Can I use this for script writing. Thank-you🤝
Yes but my customer GPT is better - Viral Hooks for Short Form Videos by Blotato
Maybe I am hallucinating but, how do I copy the text of the prompt? Trust me, I have tried everything I can think of before asking! 😉
Screenshot and drop into ChatGPT if copy paste isn’t working for some reason
Thank you!
It is the iPhone app. I opened in a browser and am able to copy. I guess I did not try everything after all :/
Cant you just paste into a new chat and ask to apply across entire account ?
It will likely add it to Memory which is less reliable than adding to Customization
This is awesome! Thank you Sabrina!
Yes
How come I can't copy the prompt?
Why does OpenAI make this prompt necessary? It seems that this 'response profile' should be a primary choice rather than requiring every user to develop their own prompt to reign in the artificial inclinations and habits of ChatGPT
and your suggestion to make it a permanent part of our profile is curiously ineffective. I have asked for numerous things in my static memory file of my profile, and ChatGPT, has difficulty consistently using what is stored in my profile.
Me too!
What a great post. Thank you!
This advice is gold, as always, Sabrina. Thanks for doing what you do, you're a rockstar.
Cool but I need to break it. If you just type humanize: {your text} it can gives you an awesome result.
I think this prompt is grossly mischaracterized and is fundamentally defunct as a “humanizer.” It may have some value as a narrow direct-response copywriting preset, and Sabrina, the author expressly says it was designed for social-media and marketing copy and should be treated as a starting point. However, once it is evaluated as a "general instruction" for making AI-assisted writing sound "human", several structural problems become difficult to ignore. and essentially it is intellectually dishonest to call it a humanizer
The first problem is that the prompt treats perceived symptoms of AI writing as though they were causes of AI writing. Em dashes, markdown, passive voice, certain transitions, longer sentences, metaphors, semicolons, and specific vernacular are supposedly telltale artifacts to suppress. (As you can tell, some of us do not speak caveman and know how to construct longer sentences.) The sequence, placing this sentence before the generative text harms the text that one might be seeking. By removing those artifacts it does more than potentially change the appearance of the output. Destructively, the prompt does not establish that the resulting prose resembles the reasoning, vocabulary, syntax, cadence, or communicative habits of any particular human being. not even in the most forgiving of generalities.
In fact, the prompt risks creating another easily recognizable synthetic style.
“Use short, impactful sentences.”
“Use simple language.”
“Use active voice.”
“Be assertive.”
“Use bullet points.”
Apply those instructions aggressively enough and every author starts sounding like the same LinkedIn copywriter.
That is *homogenization.* I would not call it" humanization".
The blacklist concerns me much more because it likely alters substantive meaning. The prompt prohibits or discourages words such as “can,” “may,” “could,” “probably,” “maybe,” “however,” “that,” and even “it.” These are not merely stylistic decorations, but very dangerous when used during the initial draft of an article. It is not my intention to come across rudely. I am a polyglot that had worked as an interpreter. I see words differently and certainly I speak to AI quite differently
“May” can express permission or possibility.
“Could” can express conditional possibility.
“Probably” communicates a degree of confidence.
“However” establishes a relationship between propositions.
“That” frequently resolves syntactic or referential ambiguity.
“It” is one of the basic mechanisms by which English maintains reference across sentences.
Consider a simple instruction:
> 'The system may terminate Process A if Condition X occurs.'
“May” matters. “Process A” matters. “If” matters. Their relationship and interconnectedness amongst each woers matters.
A stylistic rewrite that produces something such as:
> When X happens, terminate it.
may look shorter and supposedly more natural, while silently changing permission into an instruction and replacing an explicitly identified object with a pronoun whose antecedent depends upon context.
That is not a stylistic defect. It is semantic corruption.
This becomes especially serious in legal, technical, procedural, scientific, financial, or agentic writing, where apparently tiny linguistic changes can alter modality, scope, causation, chronology, authority, or the identity of the actor required to perform an action.
The instruction to favor active voice presents the same problem. Passive voice is not grammatical debris that survived editing. Sometimes the actor is unknown. Sometimes the evidentiary proposition concerns the affected object. Sometimes the actor is irrelevant. Sometimes identifying an actor would state more than the evidence establishes.
“The sample was contaminated” does not communicate the same proposition as “John contaminated the sample.”
The second sentence adds an actor and therefore an allegation that the first sentence never made.
There are also internal instruction conflicts. The prompt tells the model that if it needs to connect ideas without an em dash, it may use a semicolon. A few instructions later, it says to avoid semicolons. It encourages bullet-point lists for social posts while separately prohibiting markdown. A capable model will usually reconcile those conflicts somehow, but “somehow” is exactly what I do not want in a well-engineered prompt. The prompt should establish the intended hierarchy or exception instead of leaving the model to infer which instruction controls.
There is also no meaningful semantic-preservation rule.
That may be the most important defect.
If I ask an AI to rewrite existing material, I want the prompt to distinguish at least two things before touching the prose:
1. What is meaning?
2. What is presentation?
Meaning includes facts, actors, objects, relationships, conditions, exceptions, chronology, causation, degree of certainty, modality, scope, defined terms, and the relationships among propositions.
Presentation includes cadence, derivative inflection, sentence length, punctuation preferences, paragraph structure, vocabulary tendencies, rhetorical devices, tone, and formatting.
A style transformation should not receive unrestricted authority to alter the semantic layer.
I would therefore expect a serious humanization process to identify semantic invariants before rewriting anything. After the rewrite, I would run a semantic regression audit comparing the transformed version against the source to determine whether style modification changed the underlying proposition.
This prompt provides no such safeguard.
The treatment of warnings and qualifications also deserves scrutiny. The prompt instructs the model to avoid warnings or notes and simply provide the requested output. That may be perfectly appropriate when generating a routine social-media post. As a persistent global writing instruction, it is poorly scoped. A warning, limitation, exception, uncertainty statement, or qualification sometimes contains information necessary to make the answer accurate.
Removing it can make prose feel more confident while making the substance less correct.
The strongest suggestion in the article is almost an aside: provide samples of your own writing so the model can learn how you write. That moves much closer to what I would consider actual humanization.
If the objective is to make AI-assisted writing sound like a particular person, then study the particular person.
Analyze sentence-length distribution. Examine syntax. Determine vocabulary preferences. Look at paragraph density, transitions, tolerance for repetition, preferred degrees of explicitness, use of humor, punctuation habits, qualification patterns, metaphor frequency, rhetorical habits, and how the author develops an argument across paragraphs.
Even that requires restraint. A model should distinguish characteristics of the author's voice from accidental characteristics of a few writing samples.
My preferred architecture would therefore resemble:
AUTHOR'S INTENDED MEANING
→ SEMANTIC INVARIANTS
→ AUTHORIAL FINGERPRINT
→ AUDIENCE AND PURPOSE
→ DELIVERABLE CONSTRAINTS
→ CONTROLLED STYLE TRANSFORMATION
→ SEMANTIC REGRESSION AUDIT
→ FINAL STYLE AUDIT
Surface-level AI artifacts can certainly be examined near the end. If an author rarely uses em dashes and the model inserts sixteen of them, remove them. If the model repeatedly generates “not X, but Y” constructions, canned transitions, artificial rhetorical triads, fake profundity, or the familiar clipped cadence of AI-generated social copy, correct those tendencies.
But the reason for the correction should be that those constructions do not reflect the intended author's writing behavior or interfere with the intended communication. A punctuation mark does not become artificial merely because language models have learned to overuse it.
There is an important prompt-engineering lesson buried in this.
A blacklist tells the model what surface forms to suppress. An author model tells the system what behavior it is trying to reproduce. A semantic-preservation layer tells the system what it is forbidden to damage while doing so.
Those are materially different instructions.
For a narrow social-media use case, I could salvage portions of this prompt. I would retain the preference for clarity, factual support, practical information, reduced hype, and possibly some channel-specific formatting rules. I would discard the universal vocabulary blacklist, the blanket active-voice rule, the prohibition against necessary qualifications, and several of the absolute punctuation restrictions. I would then add explicit semantic invariants, author modeling, audience constraints, and a post-transformation semantic audit.
My concern is therefore not that every individual instruction is bad. Several are perfectly reasonable preferences when properly scoped.
The problem is the architecture.
The prompt attempts to make AI writing human by suppressing a collection of characteristics associated with AI output. That approach never adequately defines the human whose writing the model is supposedly trying to emulate, and it provides almost no protection against changing meaning during the transformation.
If I am going to call something a “humanizer,” I want it to preserve the human's thought first.
Everything else is editing.
Should we include this prompt along with humanized plug-in for Claude?
Is this still the norm this was placed a year ago?