For so long I did struggle to get Claude to do this, but I solved for it by finding example promo videos that looked similar to what I wanted (which took some time), then asking Claude to describe what it saw in the videos (which also took some time) in order to make sure Claude was seeing what I was seeing.
A couple of rounds of this and it's 1st attempt at a promo video was actually 85% of the way there. The last 10%+ just took 2 or 3 rounds of prompts.
https://github.com/mrieck/demoday-claude-plugin#top
Claude drives the app while recording, writes a script, generates a video with an AI avatar, cuts the video using Remotion. Tutorial videos, promo, listicle, I use it for all of those.
The only thing you need is your own Fal.ai api key. 1 min demos take Claude about 30 min to make, and cost $2-$3 in API costs.