Prompting for a Full-Length Script

Type "write me a YouTube script about ancient Rome" and what comes back is technically correct and completely dead. Flat sentences. No opinion anywhere. An opening that could sit on top of any video ever uploaded. The model isn't failing you. You asked it the wrong question.

Language models know how to write English. They do not know what keeps a stranger watching at minute nine. There's no hidden store of high-converting YouTube structure in there, not even in the ones that write the most naturally, because that knowledge only exists inside videos that already worked. If you want it in your script, you have to put it there.

So the job isn't really prompting. It's choosing what you feed in.

Build yourself a master prompt first, because you make it once and use it on every video after this. Go and gather three to five competitor scripts from your niche, from videos that genuinely performed rather than ones you happened to enjoy. You'll hit Claude's length limit almost straight away, so paste in only the first five minutes or so of each one. That's plenty. The opening five minutes is where the structure lives. Have it work out what those scripts share, then hand it the parameters that are already proven on your side of the niche: a competitor title to aim at, fifteen minutes as the length. What you get back is a prompt assembled out of evidence instead of the model's average guess at what a video sounds like.

There's a quicker route for a single script. Take one competitor transcript from a video that did real numbers, or just a link to that creator's most popular upload. Hand it over along with your topic, your title and your word count, and tell it to rewrite that exact format around your subject. Their intro becomes the template for your intro. Same beats, same order, same pacing, your material in it.

Notice what both routes have in common. You are feeding it actual winners. Four proven scripts out of your exact niche, or a link to the video itself. Never your description of what good looks like.

Then be explicit about the settings every single time. The style you want. The audience, including roughly what age they are. The length. Leave any of that unsaid and it quietly defaults to bland.

Spell out the shape too. A twenty-minute script with a clear beginning, middle and end. Sections ordered so each one earns the next. Real transition sentences between them rather than hard cuts. A conclusion that comes back round to the premise you opened on. And speaker text only. No camera notes, no bracketed b-roll ideas, no visual directions, because anything that doesn't get spoken aloud is just something you'll delete by hand later.

When a section lands thin, don't regenerate the whole thing. Feed that section back and ask for it expanded by about thirty percent, with more detail, more emotional tension, more storytelling. That instruction pulls out texture. A vague "make it longer" pulls out padding.

If you're working from an existing transcript instead, boil it down to ten bullet points in ChatGPT, then rewrite each point from a news-reporter perspective at around fifteen hundred characters apiece. Ten of those and you're sitting at ten to fifteen minutes without ever asking for a long script directly.

Last thing, and it catches everyone. Models miss word counts badly, in both directions, and then misreport what they produced. A script that announces itself as five thousand words will measure about eighteen hundred. It isn't lying on purpose. It just can't count.

So paste it into a word counter. Every time.