Picking the right card from 964 animated designs

Our motion library, Rushie, holds hundreds of designed card templates. When we measured a week of films, only a small part of them had ever appeared, and a handful made up a large share of all cards.

The reason was simple. The planner chose cards as a side task while writing the whole storyboard, from a short list of names, and a card could only be drawn if the scene already held every one of its fields. Rich templates, a headline with three items and a line under each, were almost never eligible, so the plain ones won every time.

The same day we brought in the full library: 964 templates in 94 packs, with new trades such as HR, legal, accounting, insurance, property and construction, each with its own description, in a proper indexed database.

Choosing a card before knowing what it must say leads to padding and generic titles. So the director works in five steps:

1. 1Content: what each drawn scene's card must show, taken from that scene alone: its shape (a sentence, a list, steps, a figure, a quote), a headline, items and figures.
2. 2Shortlist, with no model call: only templates whose fields fit that shape and count, so a list of four never lands in a card that holds three.
3. 3Choose: ranked picks per scene, the second kept as a backup, judged on layout against content, then on the film as a whole.
4. 4Set up: a writer maps the checked content onto the template's own fields and adds nothing.
5. 5Critique: one last read of meaning and wording, where any fix must quote the printed line it replaces.

Text on screen is something the film asserts, so every card passes the same checks before it ships:

• every printed line is grounded in the voice or the page;
• every figure is one this scene states, spelled-out numbers and counts in thousands or millions included;
• no list item that is really another scene's point;
• no generic title, in any of our four languages;
• no line that repeats its own label.

Numbers get one more rule. A card that shows a target, a baseline or a before-and-after needs a second figure the page rarely gives, so those cards are never offered for a page's figures, and every number the writer sets must be the scene's own.

Rules catch wrong words. They do not catch text that is clipped, tiles that stay empty or lines that overlap. So the director now looks.

For each new card, one frame is rendered through the real composition, most of the way through the scene, and shown beside the template's own sample. A vision model checks it for named defects: blank, empty tiles, clipped, overlap, unreadable. A defect earns one revision, told what the frame shows, and a second look; if it still fails, the card does not ship.

We tried a score out of ten first. It gave nearly everything a six and asked for content the voice never said. Naming defects that a text fix can mend worked far better.

A larger catalogue needs a sense of context. Some templates are made for one trade: a construction site map, a gym app, a marine electronics panel. About 300 of the 964 lead with a trade, and those are now offered only to a page that names that trade.

Cards of moving type, such as extruded 3D titles, wave letters or glitch titles, stay open to every film. Cards built around photos, and pure transitions, are left out of the menus, because the rule is motion first. And the director never touches a recording: films stay about the real site, with designed cards only between the shots.

More variety, and cards that fit. A scene that lists three benefits gets a card designed for three, with your words on it. A scene with a figure gets a counter that shows exactly that figure. A construction company gets construction cards, and a software product gets cards that suit software.

You can browse the whole library on the animations page, and see how cards and recordings work together on the features page.