Lab · Claude Skill · 2026

Clayline Ads

From a rough idea to a finished claymation ad. With an agent that waits for your approval after every phase.

Agent DesignAI VideoWorkflowOpen Source
6phases, each with approval
36 sfinished spot “Grandma Elise”
1clip per image pair instead of one long one
0third-party images or voices
Overview

I don't build the film. I build the tool that makes it.

Clayline Ads is a Claude Skill: an agent with a fixed process. You put in an idea, and it guides you through brief, plan, style, stills, video clips and sound to the finished spot. After every phase you stop, take a look and approve.

Built and tested on a real spot: Grandma Elise, a seagull and a handbag full of surprises.

Define

The problem: characters that change along the way.

AI video rarely fails at a single image. It fails over the distance. After three scenes, grandma has different glasses, the seagull different feathers, and a long clip with several actions falls apart.

So the question wasn't “How do I make a nice image?” but: how does a character stay the same across a whole spot?

Design

Six phases, each with a stop.

  1. Brief

    The rough idea becomes a structured brief with a storyline.

  2. Plan

    Characters, product and world are locked as fixed descriptions and reused word for word in every prompt. Plus a scene table.

  3. Style & characters

    Reference images for the style and every character. Everything else builds on them.

  4. Stills

    One image per scene, with a consistency check against the locked descriptions and a contact sheet for approval.

  5. Video clips

    One short clip per pair of neighbouring images, one after the other, with a check in between. Honest about sound: voice yes, music only if you bring it.

  6. Edit

    Assemble clips, mix sound, burn in subtitles, deliver the final file.

Character
Adopt

Tested on a real spot.

Grandma Elise sits in the Café du Coin. A seagull steals her handbag, eats a pill from it and suddenly glows with health. Eight scenes, two characters, one world, and the same grandma in every image.

Grandma
01 Café
The
02 Theft
A
03 Bag
The
04 Pill
The
05 Effect
Grandma
06 Spotted
Grandma
07 Reconciled
Grandma
08 Together
Result

Short clips beat long ones.

The key insight: one short, tightly guided clip per image pair instead of one long clip with several actions. That's where quality problems arise. With approval after every phase, a mistake shows up before it gets expensive.

The skill is open on GitHub. It's a remix of “claymation-ads-claude-skill” by Mike Futia, rebuilt on the Higgsfield server and extended with a consistency check, contact sheet, approval per clip and stricter rules against third-party content.