A connected LLM is fast at the typing and unreliable at taste. It does its best work when it is given your decisions to carry out: your shot count, your camera moves, your look. It proposes anything that costs money and waits for you.
What we learned
Your shot list is the spine.
WhyGiven a director's list of shots, an LLM merged some, dropped others and added its own. The count changed and so did the scene.
What we doTell it the list is fixed. It fills in what you left out: the camera on each shot, the blocking, the references. It does not merge, drop or add.
Make it tell the scene before it cuts.
WhyAsked for a shot list cold, it produced coverage that justified each shot with invented business.
What we doAsk for the scene as prose first, then the cut. Read the prose. If the story is wrong there, the shots will be wrong too.
It should not set a length on each shot.
WhyA length on every shot is a guess, and it fights the length of the render.
What we doShot lengths are left empty. The render's total length is yours to set.
Let it read, not look.
WhyLooking at pictures uses up your allowance with your LLM far faster than reading does.
What we doWe ask your LLM to use computer vision only when it is absolutely necessary. It reads each picture's prompt and settings instead, which costs you nothing.
What it cannot do
Spend. Anything that costs money arrives in the project's chat as a line with its price, and waits for you to approve or reject it. It sees only the projects you tick, for as long as you set.
The app teaches a connected LLM these rules itself, so you do not need to paste them into the chat.