How to Stop your AI Videos from Sounding like Slop: Full Prompt

Nobody talks like ChatGPT, not even the most NPC dude in the world.
A language model picks the most probable next word, over and over. Do that a thousand times and you land on the most average paragraph in the training data.
> Slop is not a bug in the output, but machine doing exactly what it was built to do.
Your visuals can be perfect and it will not save you. If the narration sounds machine written, the video is dead inside ten seconds.
The useful part is that average is measurable. It shows up as a specific vocabulary, four or five sentence shapes, and a flat rhythm. Name them and you can strip them out of a draft in ten minutes.
I have read a lot of AI scripts this year, mine and everyone else's. This is everything that gives them away, in the order that actually matters.
1. The vocabulary
A set of words that was rare five years ago is everywhere in generated text today. Readers cannot tell you why, but they hear it as machine output.
- Delete on sight: delve, utilize, leverage, tapestry, testament, paradigm, realm, myriad, plethora, underscore, embark, multifaceted.
- Allowed once, damning in pairs: robust, seamless, innovative, streamline, empower, elevate, harness, navigate, unlock, craft, transformative, compelling.
- Harmless alone, deadly by density: crucial, essential, vital, significant, remarkable, furthermore, moreover, additionally, ultimately, indeed. One per 200 words and no more.
Two words from that first list in one paragraph and the reader is gone, even if they never say so. What gives it away is the density, never one word on its own.
And do not reach for a longer synonym when you delete one. Long vocabulary reads worse, not better. When the only honest word is a banned one, describe the thing instead. "The numbers underscore the risk"* becomes "two of the six shut down inside a year".
2. The openers
"In today's fast-paced world." "In a world where." "Let's dive into." "When it comes to." "Ever wondered." If your script opens with one of those, the first three seconds are already spent.
Same with borrowed authority: "studies have shown", "research suggests", "experts agree". Name the study and the number, or cut the claim.
Same with pseudo insight: "the key is to find balance", "at the end of the day". Those fit any script about any topic, which is exactly why they are worth nothing.
3. The sentence shapes
This is where most people stop editing, and it is a louder signal than any word on any list.
- Negative parallelism. "It's not X, it's Y." "It's not just a tool, it's a workflow." Models reach for it constantly because it manufactures contrast out of nothing. Say the positive claim once.
- Fake suspense. "The result?" "The catch?" "And the best part?" Delete the question, keep the answer.
- The rule of three, in any dress. Three adjectives, three nouns, or three parallel sentences in a row. Once is a rhythm choice. Four times is a signature.
- The hedging stack. Really, genuinely, fundamentally, truly, simply put, it is worth noting. Cut every one and read the sentence again. It survived.
- Participial tails. ", highlighting the importance of..." and ", showcasing how..." The fact already ended. End there.
- Em dashes as the default connector. A person uses one every few paragraphs. A model uses one every few sentences. In a spoken script use zero, because nobody can hear a dash.
- The paragraph shape itself. Topic sentence, three supporting sentences, closing line that repeats what you just read. Page after page of identical blocks. Break the pattern on purpose.
4. Rhythm, the tell that survives a rewrite
You can paraphrase a text, keep every idea, swap every flagged word, and it will still read as machine written. The rhythm did not change.
Human sentence length swings hard. Machine text sits between 15 and 20 words per sentence, paragraph after paragraph, like a metronome.
So measure it. Count the words in your last ten sentences and write the numbers in a row. If they cluster, no synonym is going to save you.
Drop a four word sentence beside a twenty eight word one. Then a twelve. Fragments are allowed. Starting a sentence with and, but or so is allowed when the rhythm asks for it. If two sentences in a row open with the same part of speech, rewrite one of them.
5. Concreteness
Almost everything left after those four passes is abstraction, and the cure is nouns, not better adjectives.
Not "the process was inefficient" but "the same guy typed the order into one system, printed it, and typed it again into a second one".
Not "the warehouse was impressive" but "you could hear the conveyor through the office wall".
The test I run on every paragraph: could this exact paragraph appear, word for word, in a piece about a completely different subject? If it could, it is filler. That one test removes more slop than the entire ban list above.
6. If someone is going to read it out loud
A listener cannot go back a line, so spoken scripts carry their own rules.
- Contractions always. Don't, it's, we're. Spelled out, it sounds like a robot reading a contract.
- Round your numbers the way you would say them. "Almost half", not "47.3%". And never three figures in one sentence.
- Read the whole thing out loud yourself. Every stumble is a rewrite, because if your tongue trips on it the voice model will too.
- Budget by time, not by feel. Around 145 words a minute. Padding a script to reach a duration is where the slop walks back in, so drop a beat instead of stretching the rest.
7. The trap on the other side
Strip every listed word, add nothing, and you get prose that is clean and hollow. Over-correction is its own kind of slop, and it is harder to catch because it looks edited.
The fix is to start from the destination instead of the ban list. Before writing, say who is talking and to whom. Pick a real narrator: "the plumber explaining the bill to you in your own kitchen". Never ask for "a professional", because that word summons the corporate voice you are trying to escape.
Last thing, about AI detectors: ignore them. They flag human writing constantly, especially from people writing in a second language. The target is not a passing score. It is a script somebody finishes.
The prompt
Everything above, compressed into something you can paste with your draft:
> > You are my script editor. I am pasting a draft below. Rewrite it so it sounds like a person saying it out loud, not a model writing it.
> > VOICE FIRST. Before touching a word, write one line naming who is talking and to whom. If I paste a sample of my own writing, copy its rhythm and its habits, never its topic. Never aim for "professional".
> > CUT THESE WORDS: delve, utilize, leverage, facilitate, embark, multifaceted, tapestry, testament, paradigm, realm, myriad, plethora, underscore, unravel. Use at most once: robust, seamless, innovative, streamline, empower, elevate, harness, navigate, unlock, craft, transformative, compelling. Cap crucial, essential, vital, ultimately, furthermore at one per 200 words. Never swap a cut word for a longer synonym, rewrite around the thing itself.
> > CUT THESE SHAPES: "it's not X, it's Y". "The result?" and every fake question. Three item lists. Hedges (really, truly, simply put, it is worth noting). Participial tails (", highlighting how..."). Em dashes. Openers like "in today's world" or "let's dive into". Borrowed authority ("studies show") unless I gave you the source.
> > RHYTHM: after rewriting, list the word count of every sentence. If three in a row land within two words of each other, break them up. Stick a 4 word sentence beside a 28 word one.
> > CONCRETE: turn every abstraction into a name, a number, a date or something you could film. Delete any sentence that would fit an article on a different topic. Keep my facts, invent none, and mark [CHECK] where a number is missing.
> > RETURN: the rewrite in plain text, then the sentence lengths, then the 5 worst lines you cut.
Steal all of it.
Related articles

How we scaled Cal AI to $50M ARR with influencers
PLUS links to a bunch more articles and threads I've written on this subject so you can run this same playbook to scale your own consumer app.

How to Start a 6 Figure Faceless YouTube Channel (Full Breakdown)
I've helped over 400 people start profitable faceless youtube channels. i've ran over 50 channels myself. Generated over $4 MILLION by age 25. Worked with youtubers that have 10 million+ subscribers,…

$5T opportunity: AI Roll Ups
You are a sourcing analyst for a holding company buying [industry] firms in [region]. Using the attached list of licensed firms, identify businesses that likely match these criteria: 5 to 50 employee…