First Last Frame Video Skill — Development skill for Claude Code

First + Last Frame -> Video

A reusable skill for placing yourself (or any subject) into a generated video using the **first-frame + last-frame** keyframe method, then generating the in-between motion and stitching.

Instead of hoping a single text-to-video prompt lands your subject correctly, you build the **FIRST** frame (start pose) and the **LAST** frame (end pose) yourself, then let the model generate only the motion that connects them. This gives precise control over placement and pose.

The 5 steps

#	Step	What you do
1	Brief	Pick the scene + a single action arc. Anchor identity with a Soul ID.
2	Place me	Build the environment still, decide exactly where you appear in frame.
3	Slice it	Generate the FIRST and LAST keyframes (same wardrobe/lighting/lens).
4	Middle	Feed both frames to the video model as start_image + end_image; it generates the motion.
5	Stitch	Concatenate segments into one clip (music/captions in post).


The middle-generation prompt

The key rule: the prompt describes **only motion + camera + world** -- never the face/identity (the two keyframes already carry that).

Start on the FIRST frame and end exactly on the LAST frame.
[SUBJECT] [motion]. Camera [move]. [World: crowd, light, dust, motion blur].
Smooth continuous real-time motion, no cuts, photoreal, wardrobe and lighting consistent throughout.

See [SKILL.md](SKILL.md) for the full workflow, the filled-in football example, the exact generation call, and the pitfalls.

Pitfalls

• FIRST and LAST frames must share wardrobe, lighting, lens, scene -- any drift causes mid-clip warping.
• Don't re-describe identity in the middle prompt; the keyframes own it, the prompt owns the motion.
• One clean action arc per segment; multiple actions -> slice into multiple first/last pairs.