Prompting well

01Select first, then talk about "this"

The selected layer is the assistant's context. Selecting saves you a description and removes the guess.

Make the card glass and lift it on hover — with nothing selected, which card? Select the card, then: Make this glass and lift it 6px on hover.

Layer names work the same way when several layers are involved: "fade in Caption when Card is clicked" is unambiguous if those are the names in the Layers tab. Rename layers before a long session; it pays off in every prompt.

02Say the outcome, not the mechanism

The assistant knows how the tool works. Describe what you want to see; let it choose keyframes versus recipes versus streams.

Add two keyframes on opacity and one on Y with an out-cubic ease on the first… Better: Bring the card in from 40px below, fading in, over about 800ms — confident, not bouncy.

Mechanism words still help when you mean them: "on a sequence called Intro", "as a scroll section", "as a stream from Mouse X" pin the structure without dictating every value.

03Give numbers where numbers matter

Vague words get sensible defaults; numbers get exactly what you said.

Make the grid a bit wider. Better: Three columns, 24px gap, 32px padding.

Durations in milliseconds, sizes in pixels, scales as factors (1.05×), colors as hex. Ranges are fine — "200–300ms" — when you want taste applied inside a limit.

04Motion: staging and timing

Good motion prompts answer four questions: what moves, from where, how long, and in what order.

Animate the hero. Better: Intro on a sequence called Intro: the background fades in first, then the headline rises 24px into place, then the button. About 1.2s in total, each element 100ms after the last. End exactly on the resting layout.

If you want the intro separate from the rest of the timeline, say "on its own sequence" — the default sequence stays clean and rules can play the intro by name.

05Interactions: name both halves

A hover is two rules — in and out. Say it, or say the word that implies it.

Grow the button on hover. Better: Grow the button to 1.04× on hover and back on leave, 150ms in, 220ms out.

For chains, write the chain: "when Card is clicked, play Intro, and when Intro ends, show Caption." For things that should react together, ask for a signal: "clicking the Start button sends a signal; the three panels listen and slide in."

06Style: use the skill words

Skills are briefs on a look, a layout or a kind of motion. Naming one — or pinning it under ▾ › Skills — gets the whole brief, not just the word.

Make it look nicer. Better: Editorial: oversized headline, a small uppercase kicker above it, one supporting line, ink on paper.

Words that route reliably: liquid glass / frosted, grain / gradient wash, neon / glow, editorial / headline, bento, kinetic type, cinematic / intro / stagger, logo reveal, hover / press / micro, scroll / parallax / section, screen transition / push, chart / count up, demo loop, soft / card, layout / grid / hero, Apple. The full list is on The AI assistant. Auto never mixes two looks in one turn — say which one you mean.

07Whole pages: brief the page, then refine

For a full composition, brief it like a page and let the assistant draft the layout first — it composes the structure as a page and converts it into real layers, then adds motion on top.

Make a landing page. Better: A dark landing page for a synth plug-in, 960 wide: nav with a logo and two links, a hero with a big headline and one accent button, a three-card feature row, a footer. Then a staged intro and a hover on the cards.

After the first pass, refine one region at a time: "the hero headline is too small — 72px, tighter leading".

08Iterate in small turns

One change per message is faster than one big message that gets half right. Each turn is its own undo step, so small turns also mean precise undo.

Slower, and make the button blue, and add a shadow to every card, and… Better: Slower — 1.5s total. Then: Button fill #0a84ff. Then: Soft shadow on the cards.

If a turn goes wrong, Ctrl+Z on the canvas and rephrase, or say what was wrong: "the caption should end at full opacity, it stops at 0.6".

09Ask it to look

After a creative change, the assistant can look at the rendered canvas and judge its own work — at several scroll positions and moments in time, and just after a hover or click. Invite it when the result matters: "build it, then look at it and fix anything that reads wrong — spacing, contrast, the first frame."

10Point at the moment

When the problem is at one spot in time, park the playhead there and say so — "here the title is too big". With the Frame chip on auto, words like here and this frame send the canvas at the playhead along with your message; the chip's dot fills when it will. Set it to on to send the frame every time.

11Ask how, instead of asking for it

Want to learn where a control lives? ▾ › Mode › Show me how and ask "how do I make this card lift on hover?". Nothing changes: you get numbered steps, Show me highlights each control in the app in turn, and Do it for me applies the same steps if you change your mind.

12Motion: let the Motion view do the mechanics

For a whole page, Motion › Rough in gives you an animatic in one click and Animate by section has the assistant work through a long page one band at a time. Then prompt only what's left: "the cards should overlap more". See Choreography & the Motion tab.

13What to expect, and what not to

• It edits the open scene — ask it to draw, animate, style and wire, not to save, export or open files.
• It applies directly; the action card, the ✦ AI preview strip (keep ⏎ / discard) and Ctrl+Z are your review.
• It can generate a real image, video, sound clip or 3D model when you ask for one — "make a hero image of X" runs the same Make the chip does, keeping Ref and Look. See The ✦ verbs.
• It can't hear or scroll: to try what it built, use Play or Preview yourself.

14A prompt checklist