AI Script Coverage

AI script coverage is exploding and I hope people take it with a grain of salt.

It’s genuinely useful to get free and instant feedback, but you have to know how to use it.

For one thing, there is no reason to use a paid service when you can just use Claude (the smartest of the LLMs) for free or for a basic subscription.

I hope people understand that Greenlight Coverage did not actually build an LLM themselves? They just run your script through ChatGPT, with their own “wrapper” of instructions, and pass on the cost of the tokens (at a huge multiple) to you.

The models used to “hallucinate” to the point where they were worthless, but they seem to have surpassed that problem. But their new problems:

They tend to love/overscore everything to make you feel good. When you ask them to score on a 1–10 range, basically everything will be between 6.5 and, at peak, 8.5 (and “consider”). Please don’t take it seriously.

They run on “pattern matching”—so they’re not actually discerning if a script is good, but rather whether it fits existing patterns from scripts known to be good.

Consequently, their prescriptions for improvement don’t really make a script better, but more conventional.

And they fail to take into account how all of those prescriptions will combine into a whole that makes sense. Try to follow all their advice, you’ll have mush.

In fact, they write coverage “backwards”—they write the coverage and then plug in unique information from the script. They don’t actually have a reaction to the script, and then write the coverage, as a human would.

That said, they do point out weaknesses that are useful to know. In fact, the weaknesses are the only real takeaways. Sometimes they’re serious, and on a really good script, they become trivial—but it’s hard to tell from the way they are presented.

Be careful, writers.