Read first, write second: a planner in seven steps

Our classic planner asked a single Gemini call for a whole storyboard. In that one answer it had to understand the page, shape a story, choose a visual for every scene, fill each visual and write the narration.

With hundreds of designed cards now on offer, that was simply too much in one go. The signs were subtle but real: a card whose words repeated the caption, a figure that did not quite match the page, and few designed cards at all. The model was choosing a card before it knew what it would put inside.

The new planner, which we call the cast planner, works the way a careful editor does, one decision at a time:

1. 1Read the page into checked material: figures that appear in a real sentence, quotes with a named author, product moments worth filming.
2. 2Structure the film around its goal, setting aside room for real recordings of the site from the start.
3. 3Cast each drawn moment to a card by the shape of its content: a list, a figure, a quote, a set of steps.
4. 4Fill each card's fields, trimmed to that card's own limits.
5. 5Write the narration and the card words together, now that the cards are fixed.
6. 6Time the script to the length you asked for.
7. 7Check it: the review may reword, but never rebuild.

The heart of it is step one. Every later step draws on material checked against your page, so nothing has to be invented to fill a slot.

We ran both planners on the same 11 test sites. Designed cards per film went from 0.5 to 2.3. Scenes where a card echoed its own caption fell from four to one, and every film kept its share of real recordings.

It is also lighter. Because each step only sees what it needs, the text we send the model per film fell from about 20,900 tokens to 5,500, roughly a quarter.

We are rolling it out carefully: from today it plans one page film in five, chosen at random, and each film notes which planner made it. That lets us compare the two on real films before deciding anything.

Building the new planner sharpened our sense of what a good shot is, so a few rules now apply to every film, whichever planner makes it:

• the site's logo or social preview image is never used as the hero picture;
• the footer, menu, cookie banner, language picker and shopping cart are never filmed as the subject;
• a typing shot always types a real word, never just a lone number;
• a quote needs a named person and is never put in the brand's own mouth.

A good plan deserves a clear frame. On a shoe shop's site, a "Where are we shipping to?" box sat over the whole page; elsewhere it was cookie walls, newsletter offers and language pickers.

The recorder now clears them first. Its pop-up pass, popupGuard, runs after every page load, before each shot and after any click that changes the page, and closes each blocking layer the gentlest way available:

• the pop-up's own close button, recognised in eight languages;
• "Reject" or "Necessary only" on a cookie wall, before anything else;
• the film's own country on a region picker, staying on the same site;
• then Escape, and only last, hiding the layer.

It has firm limits. It never submits a form, never types, never leaves the site and never claims an age. An age check is respected, and that shot becomes a designed card instead.

More of your film is now built from what your page actually says: real figures, real quotes, real moments of your product, each in a card shaped for it. And the shots of your site show your site, not a cookie banner.

Splitting one call into seven did not make the model smarter. It made every decision smaller and gave us a place to check the material before anything was built on it. You can see films made this way on the videos page.