Setting Up Your AI Assistants

Most people get thin, generic output from AI and conclude the model isn't good enough. It's almost never the model. It's that they gave it a sentence and expected it to know what they wanted.

The quality of what comes back moves almost exactly in proportion to the context you put in. That's the whole rule, and it changes what setup even means. You're not writing a clever prompt. You're building a pile of reference material and handing it over.

So build the pile. Transcripts, and not three of them. Hundreds, pulled from the channels that are already winning in your niche. The channels themselves, so it can see what the titles look like, what the thumbnails do, how long the videos run. Examples of the exact thing you want made, and examples of the thing you don't want, because a bad example is often more instructive than a good one. Your own best script, once you've got one. Every piece of that is context, and every piece of it narrows the gap between what's in your head and what lands on the page.

Feed it more than feels reasonable. People stop at two examples because two feels like enough to demonstrate the idea. Two is a hint. Fifty is a specification.

Then the second half, which is where most of the remaining quality comes from.

Treat the first answer as a draft. The second one too. Not the result, not a verdict on the tool, just the opening move in something that's supposed to take a few passes. Read it, tell it precisely what's wrong, hand it another example of what right looks like, ask again. The people getting good output aren't prompting better. They're iterating and everyone else is stopping at one.

And when you're trying to match a style rather than just get something usable, stop judging it yourself. You've read the reference forty times and you'll see a resemblance that isn't there.

Do it objectively instead. Take several candidate scripts, strip the labels off them so nothing signals which is which, and put them in front of a few different frontier models alongside the reference transcript. Then ask the plain question: which of these were written by the same person as the reference? If the models can't pick your version out, the match is real. If every one of them points at the genuine article straight away, your style transfer failed and you now know it before you published it.

That's a test, not an opinion. Use it.