Why AI spits out clichés — and what to add to your request so it stops
"Write me a post" → a faceless wall of text. The prediction mechanics explained through real before/after prompts, plus a builder that upgrades your request in a minute.
01 — The pain
Why does a normal request return a faceless wall of text?
Most of us wrote our first prompts in the style of "Write me a post" — and expected magic. Me included.
And the output was a faceless wall of text you'd be embarrassed to show your cat.
From there, people usually take one of two paths: decide "AI is still raw," or start collecting other people's prompts. Both miss, because the problem isn't the model and isn't a missing secret prompt. The problem is that we don't understand what the model does with our words. Let's look inside — after that, clichés stop being a mystery.
02 — The mechanics
What does AI actually do with your request?
People think AI understands them. In fact it just guesses.
Literally: it looks at your phrase, digs into its training and reasons like this: "Right — when someone writes this, they usually want something like that; I've seen these patterns in a billion texts." It doesn't read minds. It matches patterns in data.
Picture your phone's autocomplete: you type "Hey, how" — the phone suggests "are you." It doesn't understand meaning; it predicts the next word. ChatGPT is the same autocomplete, just a million times smarter.
A cliché isn't a glitch. A cliché is "the most likely continuation"
when nothing in your request narrows the choice.
Empty request → the internet's average answer. Every time.I unpacked why the same prompt gives different results in a separate article on probability — here, let's focus on practice: opening the right branches with words.
03 — The settings
How does every word change the answer?
Every word in a prompt isn't just a word. It's a setting that triggers different prediction branches inside the AI. Compare a real pair of landing-page requests:
"Write copy for a landing page."
"Write copy for a design-course landing page — for someone tired of info-noise who wants to earn from their creativity. Make them feel spoken to like a human, no pressure. Instead of 'a unique offer' — 'a mix of 3 professions to try at the start and pick what fits'; instead of 'the market's best experts' — 'practitioners who'll show the inner kitchen of working with major brands.' Justify every claim: why does the reader need it?"
In the first case the AI reaches for general templates — clichés. In the second it builds the answer around your specific task, accounting for dozens of micro-details. Note the "instead of X — Y" move: you explicitly ban specific clichés and provide the replacement. It's the fastest way to knock "unique offers" and "best experts" out of the text.
04 — The details
What else moves the result: adjectives and banning banality
A second real pair — content ideas:
"Give me content ideas."
"Give me 10 Reels ideas that don't look like yet another blogger-tips list — ideas that break patterns, cut through banner blindness, and make a person stop, look closer and think: 'oh, that's told unusually — I want to save this.'"
Two things are working here. First — adjectives: in my focus groups we tested many times how the result shifts because of literally two words in the request. "Ideas" and "unexpected, provocative ideas" open different prediction branches. Second — banning banality with a reader in mind: "not like yet another blogger-tips list" plus the reaction you want ("I want to save this").
1. Who it's for and what hurts: "for someone tired of X,
looking for Y"
2. Instead of X — Y: replace the expected clichés with your
own phrasing
3. Adjectives + desired reaction: "unexpected…, so they want
to save it"05 — The builder
How do you upgrade your request in a minute?
Assemble the upgraded version of your request — fill in the fields and the three levers snap into place:
The last line is a small trick with a big effect: the model names its own "most likely" templates and routes around them. It works because you're literally closing the prediction branches it would fall into by default.
06 — The limit
How far does an upgraded request take you — and what's next?
The honest limit of this method: a detailed request removes clichés but doesn't add you. The model stops writing "in today's world" — but it still doesn't know your cases, numbers or manner of speech.
So the ladder looks like this:
- Level 1 — the upgraded request (this article): minus clichés, on task.
- Level 2 — your materials as context: cases, examples, style — the text becomes yours. That's the context article.
- Level 3 — standing instructions and assistants: the settings live in a system, not in every request.
Request with no details → clichés ("the most likely thing")
+ the 3 levers above → on-task text, no stamps
+ your materials → text that sounds like you
+ rules file / assistant → it works without remindersAI doesn't understand — it guesses. A cliché is "the most likely thing" when the request is empty. Add the three levers — who it's for and what hurts, "instead of X — Y," adjectives with a desired reaction — and the same model starts answering for your task. Upgrade one of yesterday's requests in the builder above and compare.
FAQ
Why does AI produce clichés at all?
Because the model doesn't understand — it predicts: it completes your phrase with the most likely continuation learned from billions of texts. For a short generic request, the most likely thing is the internet's most worn-out phrasing. A cliché isn't a glitch; it's the math of an empty request.
What should I add to a request first?
Three levers: who the text is for and what hurts them; "instead of cliché X — my phrasing Y" replacements; adjectives plus the reader's desired reaction ("unexpected ideas they'll want to save"). Each lever opens different prediction branches — the result changes radically.
Can two words really change the output?
Yes — we verified it repeatedly in my focus groups: the gap between "give me ideas" and "give me unexpected, provocative ideas" is different prediction branches and noticeably different answers. Words in a prompt act as settings, not decoration.
Is a detailed request enough to make the text sound like me?
No. An upgraded request removes clichés and aims at the task, but your voice only comes from your materials: sample texts, cases, numbers, manner of speech. That's the next level — context: hand the model your materials and the text stops being merely "non-template" and becomes yours.
