Why ChatGPT Sounds Generic (And How to Fix It)

TL;DR: ChatGPT sounds generic when your prompt carries no information that separates your situation from the ten thousand similar ones it could be averaging over. Fix it by narrowing what you give it: a specific reader, a specific situation, real details only you have, a stated opinion, constraints that force a choice, and, above all, your own writing as a sample to match.

Why Does ChatGPT Sound Generic?

Because a model still has to produce something, and what it produces is the statistically likely continuation for the shape of request you gave it. Ask for "a LinkedIn post about our product launch" and you get the launch post that shape of prompt produces most often: an opening hook, three benefits, a call to reflect in the comments. Nothing in the request said which launch, whose product, or who is meant to feel anything reading it, so the model filled every gap with the most common value it has seen for that shape of post. That is not laziness. That is "most probable continuation" doing exactly what it means when nothing tells it to prefer one option over the rest.

Most advice aimed at this problem attacks the output rather than the input. "Sound less robotic," "add some personality," "write like a human": all three describe the finished text, and none supplies the missing information that produced the flat text in the first place. Asking a model to sound human just resolves a different average, its trained-in average of human-sounding prose, itself drawn from an enormous range of writing. You have not narrowed anything. You have relabelled the same request.

The useful reframe: generic is not a register the model chose, it is the shape of an answer to a question with no distinguishing detail in it. Narrow the question and the answer's shape changes with it, because there is now a reason to prefer one specific completion over the thousand adjacent ones it would otherwise split probability across.

This article is about that averaging problem showing up as tone: your writing reads like nobody. It's a narrower target than why your ChatGPT answers are bad, which covers ten causes of a weak answer, including hallucination and lost context. If your answers are wrong or vague rather than just flat, start there instead.

What Actually Narrows an Answer?

Six things, roughly in order of what they cost to supply and how much they change the output. The first five take a sentence each. The sixth is the one almost nobody does, and it outperforms the other five combined.

The first five are the subject of the next five sections. The sixth gets its own, longer section after them, because it is doing something structurally different from the other five: it does not describe a constraint, it hands the model a fixed point to copy.

Who Exactly Is Going to Read This?

"Write a product description" has no reader in it. "Write a product description for someone comparing three competitors on a phone, thirty seconds before they close the tab" has one, and every sentence has to survive contact with that person. Specificity here also tells the model what to leave out: features that matter to a returning customer are dead weight to someone deciding in thirty seconds, and a model with no reader has no basis to prefer one over the other, so it serves both and satisfies neither.

1 · Product description · vague reader

BEFORE
Write a product description for our noise-cancelling headphones.

AFTER
Write a product description for someone comparing three noise-cancelling
headphones on their phone, thirty seconds before they close the tab.
They already know what noise-cancelling means. They are deciding between
this and two others open in other tabs. Lead with the one fact that
would make them stop comparing.

2 · Onboarding email · vague reader

BEFORE
Write a welcome email for new users.

AFTER
Write a welcome email for someone who signed up at 11pm after reading
a comparison post, has not opened the product yet, and will delete
this email in six seconds if the first line reads like a company
talking instead of a person.

3 · Internal memo · vague reader

BEFORE
Write a memo announcing the new expense policy.

AFTER
Write a memo for people who have already had one expense report
rejected this quarter and are annoyed about it. They want three things:
what changed, what happens to reports already submitted, and who to
ask if their situation is not covered. In that order.

What Situation Is This Actually For?

A situation is narrower than a reader: the specific moment the text has to survive, not just the person reading it. "Write a LinkedIn post about hiring" fits a thousand hiring situations equally badly. "Write a LinkedIn post announcing our first sales hire, six months after I said publicly we'd stay founder-led sales through year one" fits exactly one, and the tension in it (the reversal) is the only thing worth writing about. Told the situation, the model writes toward what's actually true. Told only the topic, it writes toward what's usually true of posts shaped like it.

4 · Announcement post · vague situation

BEFORE
Write a LinkedIn post about our first sales hire.

AFTER
Write a LinkedIn post announcing our first sales hire, six months after
I publicly said we would stay founder-led sales through year one.
Address the reversal directly in the first two lines rather than
burying it. Do not use the word "excited."

5 · Apology to a customer · vague situation

BEFORE
Write an apology email for the outage.

AFTER
Write an apology email for a 40-minute outage during a customer's
end-of-quarter close, the third outage this quarter, where our status
page was 12 minutes late to update. Do not promise it will never
happen again. Say what changed after the second outage that failed
to prevent this one.

6 · Investor update · vague situation

BEFORE
Write our monthly investor update.

AFTER
Write this month's investor update. Revenue grew but a key customer
churned, and that is the fact the update has to lead with, not bury
in paragraph three. Investors reading this already know our numbers
from last month; do not restate context they have.

What Do You Know That the Model Doesn't?

Every generic paragraph you have gotten back is generic exactly where a real detail was supposed to go. "Our platform enables teams to work more efficiently" fills the space where a number, a name, or a mechanism belongs, because the model was never given one. It cannot invent your churn rate, your customer's actual complaint, or the feature that caused the problem, so it reaches for the phrase true of almost every company in your category, and therefore true of none of them in particular.

The fix costs nothing but attention: before you send the prompt, ask what you know about this exact case that a stranger writing on the same topic would not. That fact is the sentence the piece should be built around. If the content is your startup's public brand voice specifically, why your startup's AI content sounds like everyone else's works through the same problem at the level of a whole GTM program, including what happens when a stored voice brief goes stale.

7 · Feature announcement · missing detail

BEFORE
Write an announcement for our new export feature.

AFTER
Write an announcement for the CSV export feature. The real story:
this was the #1 requested feature on our roadmap board for eight
months, and it shipped because 40% of churn-survey respondents named
"can't get my data out" as a reason for leaving. Lead with that,
not with a features list.

8 · Cold outreach · missing detail

BEFORE
Write a cold email to a marketing director at a mid-size SaaS company.

AFTER
Write a cold email to Priya, marketing director at a 40-person SaaS
company that just posted a job for a "Content Ops Lead" — that posting
is the one signal I have that they're feeling this problem right now.
Reference the job posting, not the industry.

9 · Case study intro · missing detail

BEFORE
Write an intro paragraph for a customer case study.

AFTER
Write the intro paragraph. The number that matters: they cut their
onboarding time from 9 days to 2, and the CEO said in the interview
"we didn't think a tool this small could touch a process this old."
Use that quote near the top, not buried at the end.

What's Your Actual Opinion Here?

A model with no instructed position gives you a survey of positions, evenly weighted, because taking a side is a choice nobody asked it to make. "Discuss the pros and cons of remote work" produces exactly that: pros, then cons, then a closing sentence conceding both sides have merit. That closing sentence is one of the most reliable tells in this article, and it appears because nobody told the model which side it was allowed to believe.

State the position you want defended and the hedge disappears: there is no longer a second side to protect. The same reflex to avoid disagreement shows up as reflexive apologising once you push back; why does ChatGPT keep apologising covers that pattern and where it comes from in training.

10 · Opinion piece · no stated position

BEFORE
Write about whether founders should hire a VP of Sales early or late.

AFTER
Argue that most founders hire a VP of Sales too early, before they
personally understand what "closed" means for their own product.
You do not need to represent the other side fairly. State the
position and defend it with one example.

11 · Internal recommendation · no stated position

BEFORE
Write up the pros and cons of switching CRMs.

AFTER
Recommend switching CRMs by Q2. State the recommendation in the first
sentence. Use the pros/cons only to address the two objections you
expect from finance, not as a balanced overview.

12 · Review response · no stated position

BEFORE
Write a response to this negative review.

AFTER
Write a response that agrees the wait time was unacceptable, without
qualifying it with "however" or "that said." State what changed
because of this specific complaint, not a general commitment to
improvement.

What Constraint Would Force a Real Choice?

An unconstrained prompt lets the model pick its own defaults, and every unconstrained prompt gets the same ones: a three-part structure, a middle hedge, a summary ending. A constraint is not a restriction on quality, it's a forcing function that makes the model choose one option instead of gesturing at several. "Under 40 words" rules out the safe, padded version. "No question marks" rules out the rhetorical-question opener that shows up in roughly every third generic paragraph. Constraints do the most work when they eliminate the specific default you're tired of, not when they're generic instructions like "be concise."

13 · Subject line · no constraint

BEFORE
Write a subject line for this email.

AFTER
Write a subject line under 40 characters. No question marks. No
emoji. It has to make sense on its own without the preview text.

14 · Social caption · no constraint

BEFORE
Write a caption for this product photo.

AFTER
Write a caption in exactly two sentences. First sentence is a
concrete detail about the object, not an adjective. Second sentence
is the price, stated plainly, no "just" or "only."

15 · Executive summary · no constraint

BEFORE
Summarize this report for the leadership team.

AFTER
Summarize this report in 150 words. Open with the one decision this
report should change, not with what the report covers. No bullet
list under three words per bullet.

Why Is a Writing Sample the Strongest Fix of All?

Because every lever above still asks the model to interpret something. A reader description, a stated opinion, a constraint: all of them are instructions the model has to translate into prose, and translation leaves room for its own average to leak back in. A writing sample skips translation entirely: instead of describing your voice for the model to reconstruct, you hand it a fixed point to match, a far more literal task than resolving an adjective.

Anthropic's own prompting documentation states this plainly: "Examples are one of the most reliable ways to steer Claude's output format, tone, and structure." (platform.claude.com/docs/en/build-with-claude/prompt-engineering/claude-prompting-best-practices, accessed 2 Sep 2026.) The page recommends three to five examples, wrapped in <example> tags so the model can tell instructions from content. That's written for reusable application prompts, but the mechanism is identical in a one-off ChatGPT or Claude conversation: an example anchors output to itself in a way a description never can.

Here is the pattern, and it costs you nothing but the thirty seconds it takes to copy something you already wrote.

16 · Match voice, not adjectives · reply

BEFORE
Reply to this customer email in a friendly, professional tone.

AFTER
Below is an email I wrote last month that the customer replied to
within the hour. Study its rhythm, its vocabulary, and where it
chooses to end sentences, not its topic.

[PASTE 300–500 WORDS YOU ACTUALLY WROTE]

Now reply to the message below, matching that voice. Do not match
its length or its subject, only its voice.

[PASTE THE MESSAGE TO REPLY TO]

17 · Match voice · LinkedIn post

BEFORE
Write a LinkedIn post in my voice: direct, a little dry, no corporate
language.

AFTER
Here is a post of mine that got more replies than usual:
[PASTE THE POST]
Notice the sentence lengths and where it uses a fragment instead of
a full sentence. Write a new post about [TOPIC] matching that pattern,
not this post's subject.

18 · Match voice · product description

BEFORE
Write product descriptions in a warm, approachable tone.

AFTER
Here is a description of ours we like:
[PASTE 40–60 WORDS]
The pattern: one sentence on the problem it solves, one on how it's
different, no adjectives in the first line. Write five more in that
exact pattern for the products below.

19 · Match voice · cover letter

BEFORE
Write a cover letter that sounds confident but not arrogant.

AFTER
Here is an email I sent a colleague that got the response I wanted:
[PASTE 200 WORDS]
Match its directness and its sentence length. The letter is for the
role below — match the voice from the sample, not its content.

20 · Match voice · team status update

BEFORE
Write a weekly status update for the team.

AFTER
Here is last week's update, written by me:
[PASTE PREVIOUS UPDATE]
Match its structure and how blunt it is about what slipped. This
week's facts are below — do not soften them the way a generic status
update would.

21 · Match voice · investor email

BEFORE
Write a short update email to an investor.

AFTER
Here is an update I sent an investor who replied "this is exactly
the level of detail I want":
[PASTE 150 WORDS]
Match its length and its habit of naming the number before the
narrative. This month's numbers and one problem are below.

A sample this short takes less time to paste than a tone description takes to write, and it outperforms the tone description on every axis that matters, because "direct" is a word the model has to resolve and a paragraph of your actual writing is not.

One place worth knowing about if you want this to survive past a single chat: OpenAI's own help pages describe ChatGPT Projects as workspaces where you can "add files" and "add custom instructions" that a project remembers across every chat inside it (help.openai.com/en/articles/10169521-projects-in-chatgpt, accessed 2 Sep 2026). A voice sample pasted once into a project's files or instructions applies to everything you generate inside it afterward, rather than being retyped per chat. On our own side, the equivalent is the Personal Context Library and Global Variables, which exist to hold exactly this kind of reusable block (brand voice, writing style, a sample) so it injects automatically instead of living in your clipboard history.

Does Editing Custom Instructions Change Your Old Chats?

Depends which sentence on OpenAI's own help page you read. Fetched directly on 2 Sep 2026 from help.openai.com/en/articles/8096356-chatgpt-custom-instructions, the article body states: "Updates to custom instructions settings are applied immediately across all chats (including existing conversations)." A few paragraphs further down, under Frequently Asked Questions, the same article answers a related question this way: "If I update or remove my custom instructions, will previous versions of my instructions continue to appear in my chat history? Yes, updates to your instructions are reflected only in future conversations."

The practical stakes for this article: if you save a voice sample or style rule into custom instructions mid-conversation, do not assume the chat you are already in picked it up. Paste the instruction directly into the current chat as well, and treat custom instructions as the thing that governs conversations you have not started yet. Custom instructions also have a published size limit worth knowing if you are storing a sample there rather than in a project: Free and Go accounts can save up to 1,500 characters, and Plus, Pro, Enterprise, Business, and Education accounts can save up to 5,000 (same article). A 500-word writing sample is roughly 2,500–3,000 characters, which fits inside the higher limit and not comfortably inside the lower one. That is another reason a project's file storage, which has no comparable character cap, is worth using for a longer sample.

What About the Throat-Clearing and the Tidy Endings?

Everything above is about the input side: what you put into the prompt. There is a separate, well-documented catalogue of surface habits models fall into on the output side regardless of what you asked for: reflexive praise, a preamble restating your question, three-item adjective lists, hedges that soften every claim, and a closing paragraph that adds nothing because it exists only to sound finished. Those are real, several are documented directly by OpenAI and Anthropic in their own release notes and system prompts, and naming them honestly matters more than pretending an em dash is a smoking gun.

That catalogue, with the vendor citations and the exact fix for each one, is its own article: why does AI writing sound like AI. This post's job is the layer underneath those tells (the missing information that made the model default to its house style in the first place), so rather than re-teach that catalogue here, the short version is this: narrowing your prompt with the six levers above reduces how often those tells show up, because a model matching your writing sample or defending your stated opinion has less room to fall back on its own defaults. It does not eliminate them on its own. If a draft still opens with "Great question!" after you've done everything in this article, that catalogue is where the remaining fix lives.

Five Full Prompts With Every Lever Stacked

Realistic prompts rarely use one lever at a time. Here is what stacking them looks like on five actual tasks: reader, situation, a real detail, and a sample, combined. For a much larger set of pairs organised by job rather than by writing voice specifically, a gallery of before-and-after prompts covers fifty of them across coding, teaching, analysis and more.

22 · Product launch post

BEFORE
Write a post announcing our new integration with Slack.

AFTER
Write a LinkedIn post for people who already use both tools separately
and are tired of copy-pasting between them — that's the actual
situation, not "Slack users" broadly. The real detail: this shipped
14 months after our first customer requested it, and we're naming
that customer in the post with their permission. Here's a post of
mine that did well, match its rhythm:
[PASTE SAMPLE]

23 · Pricing page FAQ answer

BEFORE
Write an FAQ answer explaining our refund policy.

AFTER
Write the answer for someone who has already been burned by a
"no refunds" policy elsewhere and is reading this specifically to
check if we're the same. State the 14-day window in the first
sentence, not the last. No "we understand your concerns" opener.
Match the directness of this existing FAQ answer: [PASTE EXAMPLE]

24 · Job posting intro

BEFORE
Write an intro paragraph for our marketing manager job posting.

AFTER
Write it for someone who has read fifteen identical "fast-paced,
dynamic team" postings today and will skip yours in four seconds
if it sounds like the other fourteen. The real detail: this role
exists because our last three campaigns underperformed on organic,
not because we're "growing fast." Match the plain tone of this old
posting that got strong applicants: [PASTE EXAMPLE]

25 · Post-mortem summary

BEFORE
Summarize what went wrong with last week's deploy.

AFTER
Write it for the engineering team, who already know the deploy broke
checkout for 40 minutes — do not re-explain the incident, get straight
to what changed in the rollback process. The real detail: the alert
that should have fired didn't, because it was silenced during an
unrelated maintenance window three weeks ago and never re-enabled.
Match the blunt tone of this old post-mortem: [PASTE EXAMPLE]

26 · Client renewal email

BEFORE
Write a renewal reminder email to a client.

AFTER
Write it for a client who mentioned in their last call that budget
is tight this quarter — address that directly rather than ignoring
it. The real detail: their usage grew 30% since signing, which is
the actual case for renewing, not a generic list of features. Match
the tone of this email that got a fast reply: [PASTE EXAMPLE]

What Does This Actually Cost You?

Five of the six levers here cost a sentence each: name the reader, the situation, the detail, the position, the constraint. The sixth costs nothing you don't already have: a paragraph of your own writing, sitting in an old email, a Slack message, or a post that did well. None of it requires a subscription, an extension, or a specific model. It works in a free ChatGPT account exactly as well as anywhere else, because the mechanism is the prompt, not the tool.

Where a tool like ours adds something is building the other five levers around that sample automatically rather than typing them fresh each time. Prompt Architects' free plan includes five prompt enhancements a day, forever, with no credit card, per our own FAQ page. The enhancer fills in role, format, and constraints the way this article has been describing by hand. Paid tiers add a Tone Selector with preset registers and a Personal Context Library to store recurring detail (a voice sample, a reader description, a banned-phrase list) so it applies automatically instead of being pasted per prompt. None of it replaces a specific, narrowed prompt; it just saves retyping what you've already worked out.

But say plainly what the free version of this fix is, because it matters most: open your last ten AI-generated paragraphs that sounded like nobody, and paste one real paragraph you wrote into the next request instead of a tone adjective. That one habit, done consistently, closes more of the gap between generic and yours than anything else here.