Films that land on the length you asked for

A film's script is the cheapest place to improve it. Once a scene is recorded, voiced and rendered, every change costs time; before that, it costs a few words. So every storyboard is now reviewed before filming.

A quick automatic check scores the draft out of 100. Does it open with a hook? Does the website get its share of the screen? Is there filler, an empty graphic, an odd duration, a repeated motion, a vague call to action? When something is off, a script-supervisor pass rewrites the draft, and we keep whichever version scores better, scene by scene, because a rewrite can improve some scenes and not others.

Behind it, every prompt now speaks through a persona with one job: analyst, format director, narrator, reviewer, producer, revision editor. They share one set of craft rules, so the whole film is written by one consistent crew.

One of those shared rules matters more than the rest: a film only states what your page states. Language models learn from examples, and a model can take a number from an example and treat it as material.

So examples in our prompts are now placeholders, with no real-looking prices or percentages. Every persona carries the reminder that examples are not facts. And the review treats any price, percentage or large figure that does not appear in the page, the brief or our analysis of the site as an error to fix before filming.

The same care goes into where the camera points. Every element the planner asks to show is matched to a real element we found on the page, and the recorder tries a simpler version of the target before giving up on it.

Then we measured finished films against their briefs, and learned something simple. Each scene stretches to fit its own narration, so the number of words is the only real control over a film's length. Nothing later in the pipeline can shorten a script that says too much.

Our planner was budgeting for a faster speaking pace than our voices actually use, about 112 words a minute. One small module now holds the measured pace for each language and tone, plus the measured time of the opening and closing cards, and every step that counts words reads from it.

The results are locked in by tests:

• a 15-second brief lands at 14.9 seconds,
• 25 seconds at 24.8,
• 35 seconds at 35.1,
• and 60 seconds at 60.3.

The delivered length is also checked after every render now, so a film's length is measured, not assumed.

Every motion in our library has an energy, from a slow float to a bold pop. Now the film uses it. Motions are cast to the beat of each scene:

• the energy follows the scene's mood or the film's format, with a stronger entrance for the opening hook and the close;
• no two scenes in a row enter from the same side or with the same kind of gesture;
• a scene shorter than 3.5 seconds never gets a text motion that is still animating after its first second;
• a motion you picked in the editor is kept on every re-render and revision.

The effect is a film where a problem lands with weight, a result lands with lift, and nothing feels dealt at random.

The review counts graphics the way you will see them. A planner may name a rich graphic and leave it without content; such a card is shown as a simple statement. So the variety rules now count only graphics that will truly render with something in them. An empty title or statement card is filled from the scene's own caption, and richer forms that need real material go back to the supervisor.

That is the thread through the whole release: measure the film the viewer will watch, not the plan we hoped for. It came with something for you as well. You can now leave timestamped notes on a finished film, right on the player, and apply them as a revision in one step. See what else a film can do on the features page, or try it on /new.