Skills

Install

In Claude Code, run these two lines.

With a Spark API key, Claude also makes the clip for you and saves it, after telling you the price and waiting for your yes.

Why the clips look real

Structure beats vibes

Every prompt names the format, who is on screen, the camera, the light, what happens second by second, the exact words, the sound, and what must never change.

Realism is directed

Phone-held sway, a glance away and back, a breath before a line, real window light, one small unforced expression. Dialogue is sized to the clip, about two and a half words a second.

Checked before it is spent

The plan is checked in code: the timeline fills the clip, something happens in the first second, every spoken line is written down, and words a model cannot act on, like cinematic, are refused.

From one sentence

“Me in my nail studio showing a fresh set of almond nails and telling people I have two slots left this Saturday.”

Vertical 9:16, 10 seconds, single continuous take, no cuts. The person from the reference image, a nail technician in their studio, showing off freshly manicured almond nails. Camera: medium shot from the waist up, hands in the foreground. Lighting: daylight from a window to the left, no ring light. Timeline: 0-2.5s: hands display the nails, a slight head tilt toward the camera (pleased). 2.5-5s: a glance at the nails and back to the camera, one small nod (confident). Lip-synced dialogue: “Just finished this set. I have two slots left this Saturday if you want to book.” Keep consistent: the person matches the reference image exactly; hands look natural with five fingers; no text on screen.