Skip to content
solomotion
AgentsAgent recipes

Prompts that work

Describe the result, the size and the feel. Name the things that matter. Ask it to look. Each recipe below is a prompt plus what the agent does with it, so you can tell when it went off course.

Built from one prompt, the second recipe on this page.

A title card

Text
1920×1080 title card. Kicker "CHAPTER 2" in caps, title "The long way round", a short accent rule. Left-aligned on a 12-column margin. Kicker fades up per letter, title unblurs per word, rule draws in. Soft pop per title word. Show me frames 0 to 2 s.

The agent stacks the text on one left edge, keys each line with an entrance preset and a stagger, adds pop sounds to the title words, then checks the motion with view_frames.

Stickers that fall

Text
Twenty round and pill-shaped stickers with tiny words (idea, draft, again, nope, v2, final) rain onto a paper desk, bounce and pile up. Real physics, no keyframes. Last one is a big "ship it" sticker that lands on top. 7 s, square.

Each sticker becomes a group with a physics_body; the comp gets a floor and walls with physics_scene; clip starts are staggered as spawn times. Nothing is baked, so you can move a sticker and the pile rebuilds.

One shape, many materials

Text
Take the sun-and-hills illustration and morph the whole frame through paper, foil, stained glass and chrome, a stop every 1.2 s, each change 700 ms.

The agent uses effects_morph with one stop per material, then view_frames at each stop and halfway between.

A character that talks

Text
Detect the parts of the fox group and bring it to life: blink, breathe, a small nod, and lip sync to the voice line on track 3. Dry run first and tell me the plan.

character_detect names the eyes, mouth and head; character_bringToLife with dryRun shows the plan, then applies it as one undo step. Without a recorded line, the offline Test Voice speaks one with exact mouth timing.

Captions for a talking head

Text
Transcribe the interview, cut every pause longer than 600 ms, then add three-word captions in the lower third with the current word highlighted.

media_transcribe and media_remove_pauses need a transcription model connected under Agents › Connect a model. The captions arrive as styled layers you can edit.

A logo reveal

Text
Here is our logo as SVG. Draw the outline on over 900 ms, fill it in, sweep one shine across it, then bring the wordmark in. 3 s, 1080 square, black ground.

The mark becomes path layers that draw on, then fills fade over the strokes. The agent checks the final frame at rest with view_render.

When it goes wrong

  • It guessed sizes. Give the comp size and where things sit: "bottom-left, inside title-safe".
  • It moved too much. Say "one hero move per shot".
  • It did not look. Say "show me three frames", then point at what is off.
  • It cannot do something. It says so and files a report. That is the intended behaviour, not a failure to hide.