Research → Script → Humanize Workflow

Most people script with AI the same way. Paste a title into a chatbot, ask for a fifteen-minute video, copy out whatever comes back. That method had a good run. It's finished now, and not because the models got worse. YouTube is patching that exact hole, and a script produced by one instant prompt is the easiest thing in the world to spot.

So stop doing it in one pass. Do it in three.

Pass one is research, and only research. You're not asking for a script here, you're asking for raw material. Use Claude 4.5 with web research turned on and extended thinking enabled, because without those it reaches into what it already knows and hands you a tidy summary of old information. Useless for anything trending this week.

Tell it you want verifiable information only, sorted by category. Sourced statistics. Exact dates. Names. Quotes with attribution. Lesser-known anecdotes. A timeline of how the thing unfolded, the controversies attached to it, and where the story stands today. Then give it an escape hatch: anything it can't confirm gets marked unverified in brackets, or left out entirely. Those brackets are what protect you two steps later.

It'll come back with clarifying questions before it starts. Answer them, at least three of them, as accurately as you can. Vague answers produce a vague generic script. Precise answers produce one nobody else has.

Now go past the single source everyone else is using. That's the difference between a rerun of a topic and something that feels new. Ask the model for ten lesser-known points on your exact title, not the general subject. Then go where the other channels won't: the well-known authors on the topic, the academic papers underneath the news articles. Obscure detail is the whole edge.

Then read the comments. Pick five to ten videos already covering your topic and go through what people said underneath them, or paste the comments into one prompt and ask for the most common questions, the frustrations, the gaps nobody covered, plus title ideas. Then answer those directly in your script. You're writing to an audience that has already told you what it wants and got ignored.

Add what the model was never going to find, too. A Reddit thread with a rumour buzzing in it, the one paragraph in an article that's the real hook of the story. Paste them in. That's how you end up with angles the AI wouldn't have surfaced on its own.

Working in a niche you know nothing about is fine, by the way. Run the same search step to pull current news in that space, with a rule written into the prompt keeping the stories on topic so it doesn't wander into adjacent nonsense. You don't need subject-matter expertise. The research surfaces the facts and the script gets written off the facts.

Pass two is the script. Turn web research off for this one. You want it building purely from the sources already gathered, not slipping back onto the internet halfway through paragraph nine and dragging in something unchecked.

And here's the rule that carries through everything after pass one: the writing pass may not add a single fact, statistic, date or claim that isn't already sitting in your research. Not one. Put it in the prompt in plain words, and repeat it in the next prompt. That's the entire defence against hallucination, and it's the reason research was a separate step in the first place.

Pass three is humanization. One final prompt over the finished script that rewrites it to sound like a person talking and tunes it for text-to-speech, so ElevenLabs delivers the lines instead of flattening them. Same restriction applies. It rewrites. It doesn't invent.

Three prompts. Never one.