How I use AI to write scripts that still sound like me
# How I use AI to write scripts that still sound like me
Every script I record starts as a draft from an AI model. By the time I read it into a microphone, almost none of the model's wording is left. Both halves of that sentence matter, and the distance between them is the whole craft of writing with these tools.
Plenty of people already write with AI and quietly deny it. I would rather describe the actual workflow, including the places where the model quietly makes writing worse if you let it.
Where the model earns its keep
Three jobs, and it does all three well.
The first is beating the blank page. An empty document is the hardest part of writing anything, and a model will fill it in seconds with something. That first pass is bad, and bad words are far easier to fix than no words. There is a psychological trick buried in this, too. Once text exists, my brain switches from inventing, which I find slow and painful, to reacting, which I am fast at. I can look at a clumsy sentence and know instantly that it is wrong, even when I could never have told you in advance what right looked like.
The second job is structure. I dump a messy pile of notes in and ask for a rough order, and what comes back is usually a sensible spine. I rarely keep it as delivered. What it gives me is something to argue with, and arguing moves faster than staring at white space ever did.
The third job is tightening. Once a paragraph says the right thing in too many words, the model proposes a version half the length, and some days the short version beats what I would have produced on my own. I paste in a bloated sentence, ask for it at half the word count, and often get back something sharper than my own second attempt would have been.
Drafting, ordering, tightening. I use it for all three without guilt.
What it cannot do, whatever the prompt
The breathless threads always skip this part. A model has no point of view. Ask it for an opinion and you get a balanced, reasonable, on the one hand sort of answer, which is exactly the answer nobody remembers. A real take picks a side and accepts being a little wrong in exchange for saying something true. That has to come from me.
It also has no taste. Choosing the right sentence out of three decent ones, sensing when a joke is trying too hard, knowing what to cut: taste is a thousand small judgments, and the model has no stake in any of them.
The biggest gap is hierarchy. To a model, the throwaway aside and the one line the whole video exists for carry equal weight, because both are just likely sequences of tokens. Only I know which sentence is the reason the piece exists and which sentences merely walk the viewer toward it. Ranking words by how much they matter, rather than by how likely they are to follow each other, is the most human judgment in all of writing, and the model has no access to it.
The specifics it can never invent
The thing I guard hardest is lived detail. A model can write a smooth paragraph about running renders on an old graphics card. What it cannot supply is the night the driver died at two in the morning, or the specific dumb mistake that cost me a weekend, because it was never there. It was never anywhere. It can only average over what other people have already written, and a good story is the opposite of an average.
Here is the test I run on my own drafts. If a detail could appear word for word in anyone else's video on the same topic, it is wallpaper. Real specifics are unrepeatable. They belong to one person and one moment, which is exactly why they land, and exactly why a machine built to predict the most likely next word can never reach them.
Naming the smell
Generated text has a smell, and once you can name it you can hunt it.
The first tell is cliche. Phrases like "in today's fast paced world" and the classic "it is important to note". Filler dressed as content. A person who cared would just say the thing.
The second is hedging. Models soften every claim into mush: it could be argued, some might say, results may vary. Real writing commits.
Then come the filler sentences that restate the previous sentence in new clothing, adding words and no meaning. Then sameness, where every paragraph has the same shape and length and rhythm, a spreadsheet wearing a prose costume. And finally the stiff register where nothing is ever plain bad, only suboptimal. Once you start seeing these you cannot stop, which is the point of learning them.
The process, which is discipline and nothing clever
Step one happens before the model sees anything: I outline by hand. The bones, the order, the one real point the piece exists to make. Outlining is where the thinking happens, and handing it away means editing the model's idea of my video instead of building my own. Every time I have skipped this step, I have regretted it.
Step two, the model drafts into my outline. It gets my messy notes and my structure and produces the first ugly pass in minutes. I treat that output as raw clay. The mindset matters more than it sounds, because treating a draft as nearly finished is how the smell survives into the final cut.
Step three is the one nobody wants to hear. I rewrite every single line. I do not skim and approve. Each sentence gets said in my head, in my own voice, until it sounds like me and never like a press release, and by the end very little of the model's wording survives. What survives is the speed. The model wrote a draft. I wrote the script. Those are different jobs, and the gap between them is the reason this piece exists.
The read aloud test
The strongest tool in the whole workflow is embarrassingly simple. I read the draft aloud, or say it in my head as if telling a friend across a table, and apply one brutal rule: any sentence I would never actually say to a friend gets cut or rewritten until I would say it. Nobody tells a friend that in today's fast paced world content strategies must adapt. The spoken version of me is more honest, more specific and a lot less stiff than the written version a model defaults to, so I let my talking voice be the editor.
Reading aloud catches more than tone. It catches every flat stretch of rhythm, every sentence too long to breathe through, every phrase I trip over because the words were never really mine. Your mouth knows things your eyes skip past. On tired days, when approving the draft and moving on feels tempting, this is the one pass I refuse to skip, because it is the cheapest quality check in existence and it has never lied to me.
Two edits, so this stays concrete
The model once handed me a line shaped like this: using artificial intelligence can significantly improve your content creation workflow. Technically a sentence. It says nearly nothing, and no human has ever talked that way. My version: AI makes the first draft faster, and that is most of what it is good for. Shorter, plainer, and it commits to a claim.
Another draft line: there are many factors to consider when deciding how to approach this. Pure stalling. A sentence that only promises a point makes no point, so I deleted it and lost nothing.
Those two moves, swapping an inflated abstraction for a plain concrete claim and deleting the sentence that only stalls, cover maybe eighty percent of every edit I make.
Small habits that hold it together
I never accept a whole paragraph from the model, only ideas and rough shapes, because whole paragraphs are how the smell sneaks back in. I keep sentences short and uneven on purpose, since the model defaults to a steady medium length drone, and breaking that drone instantly reads more human. I cut qualifiers hard: words like basically or quite or somewhat almost always weaken the line they sit in. And I aim to land one concrete specific in every section, a real number or a real moment, because specifics are the fingerprint a model cannot fake.
One habit took me longest to learn. I never edit in the window I drafted in. Text sitting inside a chat still feels like the model's writing, and I am strangely polite to it there. Moved into my own document, it is just words on a page, and I am ruthless with words on a page.
None of these are clever. They are small repeated motions, and added up they are the difference between a script and a generated artifact.
Where this leaves me
Faster, mostly. A script that used to eat most of a day now takes a few hours, and the slow remaining part is the rewriting, which was always the real work anyway. The model did not write this piece. It helped me write it, and those are different sentences. The judgment about what to say, what to cut and what stays true to what actually happened remains entirely mine, and as far as I can tell it will stay mine for a good while.
Some people frame this workflow as cheating. A draft is no more a script than a lump of clay is a pot. The work was never the typing. The work was always the choosing, the cutting and the deciding what is true, and the model simply hands me more clay, faster than I could dig it myself. What I make of it is still on me.
Get new guides and videos first — join the Telegram channel.