Skip to Content
The AI assistantPrompting it well

Prompting it well

What it is good at

  • Building structure. “A 6-second SaaS product intro with a kinetic title, three feature cards staggering in, and a light sweep.”
  • Bulk mechanical work. “Stagger the five cards in from the left, four frames apart, easing out.”
  • Wiring things you know the name of. “Give the headline a per-glyph fade-up with a range selector.”
  • Camera moves. “Add a two-node camera and push it through the stack over three seconds.”

What it is worse at

  • Taste. It will produce something coherent; making it good is direction.
  • Very long single prompts. Two focused requests beat one sprawling one.
  • Anything that depends on seeing your reference. Attach the image instead of describing it.

Be specific about the four things that matter

Duration

“6 seconds”, “90 frames” — not “short”.

Structure

Name the elements: title, three cards, a background, an emblem. The director plans against roles.

Motion

Say how things should enter and how they should feel. “Stagger four frames apart, ease out” is actionable; “make it snappy” is not.

Constraint

Anything that must not change. “Keep the existing background”, “do not touch the logo layer”.

A worked example

Make a 6-second SaaS product intro with a kinetic title, three feature cards staggering in, and a light sweep.

What happens:

  1. The router classifies this as generative, so it goes to the caster.
  2. The model writes a short creative brief, then casts — picking technique and layout ids from a capped, pre-filtered candidate list.
  3. The cast is validated; anything unknown, forbidden, clashing or too short for its slot is replaced with the top-ranked valid candidate.
  4. The deterministic libraries emit the tool calls: add_background, add_title, add_kinetic_title, add_cards, stagger_in, add_light_sweep.
  5. A fit critic judges the result against rendered evidence.
  6. A preview appears with Apply and Decline.

Apply lands the whole thing as one undo step.

Because validation replaces bad casts deterministically, a run rarely fails outright — it degrades. If a result looks generic, the cast probably fell back; a more specific brief (energy, mood, a named look) gives the model a narrower candidate list to choose from.

Iterating

Work in Manual mode while you are finding the right prompt. Decline rolls back cleanly, so a bad result costs you nothing but the time.

The productive loop is:

  1. Ask for the structure in one prompt. Apply.
  2. Refine by hand — timing, easing, colour. This is faster than re-prompting and it is where the quality comes from.
  3. Ask for a specific mechanical change if you need one: “stagger these four layers 3 frames apart” rather than “make it better”.

Everything it made is ordinary layers, keyframes and effects. Press U and edit.

Image attachments

The composer accepts image attachments. Use them for:

  • A reference look you want matched.
  • Artwork you want brought into the scene — the create_media_from_attachment tool exists for exactly this.

Reading what it did

After an Apply:

  • Press U on each new layer to see the keyframes it wrote.
  • Open the Effects tab to see what it applied.
  • Check the History panel — the whole run is one entry, which you can rename (“AI first pass”) and jump back to later.

Quick-fire prompts worth stealing

PromptTools it reaches
“Stagger these layers in from the left, 4 frames apart, easing out”stagger_in, set_keyframes, set_easing
“Add a per-glyph fade-up on the headline with a range selector”text_animator
“Push a two-node camera through the stack over 3 seconds”create_layer, add_camera_move
“Put a soft glow on every card and grade the whole scene warmer”add_effect, update_effect_param
“Rig this logo with puppet pins and give it a gentle sway”create_puppet_rig, set_puppet_pin_keyframes
“Cut the card entrances to the beat of this track”analyse_audio, stagger_in
“Make that settle with a spring instead of an ease”set_spring

Naming a look

Casting is filtered by look pack, so naming one narrows the candidate list before the model chooses. The six are grotesque, swiss, editorial, geometric, technical and humanist — “make it editorial” or “keep it swiss and restrained” is a more actionable instruction than “make it clean”.

Last updated on
Connecting a providerTool reference