Animated Story Ad . Agency OS

What it does

It makes a vertical 3D animated ad in which characters talk to each other about your product: a short story with a problem, a turn and your offer, ending on a still end card with one real result. It writes the script, plans the shots, builds a cast sheet, generates one lip-synced clip per spoken line, gives every character a fixed voice, adds scene sound and captions, and delivers the finished ad at 1080.

The value is in the checks, not the generation. Every take is transcribed twice and compared with the script word by word, every scene is measured for sync and pauses and reviewed by a model that watches and listens, and the finished cut is measured for loudness, joins, captions and the end card. A line with a stutter, an extra word or a clipped ending is regenerated, never patched.

Say this to start

Just say what you want. Any of these will do it:

> make me a 60 second animated story ad for [your product]

> Pixar style ad with two characters talking

> make an ad like this one [paste a reference ad]

> cut a 30 second version of the ad we made

When to reach for it

you want a story ad where characters talk, not a presenter or a voiceover
you saw an animated skit ad that works and want one for your own offer
you want 30 second cutdowns or new hooks from an ad you already made

When NOT to use it

If you actually want	Use this instead
an avatar or talking-head presenter video	`heygen`
a single video prompt to run in another tool	`video-prompt-builder`
footage you recorded yourself, edited	`auto-editor`
short clips from a finished long video	`video-repurposer`
images or carousels	`social-visuals` or `carousel-builder`

Before you start

This is the most tool-heavy skill in Agency OS. It runs in Claude Code on your own computer, and every key is your own.

What you need	Why
A video model that speaks a line with lip-sync from a still (tested with Seedance 2.5)	one clip per spoken line is the only reliable lip-sync	Required
An image model that takes reference images (tested with GPT Image 2.5)	the cast sheet and every shot's still	Required
An ElevenLabs API key	cleans each line and converts it to the character's fixed voice, so voices do not drift between clips	Required
A Gemini API key on a paid tier	the review that watches and listens to every scene before you see it	Required
ffmpeg and whisper.cpp with the large-v3-turbo model	all the measuring, cutting and transcription, free and local. See the setup guide's toolchain page	Required
Your verified claims and offer	the offer line and the one result on the end card come only from here	Required
Your fonts and colours	captions and the end card are built in your brand. The build stops without them	Required
A video upscaler	for the 1080 file. Without one you deliver the 720	Optional
A track you have the rights to use, or a music tool	a soft bed under the ad	Optional

How it runs

ScriptA short story with every question a viewer would ask answered, lines of 14 words or fewer, and your offer word for word. Both claim checks run before you see it. You approve it.
Storyboard and costEvery shot has a still, a line, a clip length and a background sound. Runtime comes from the words, not the panels. You get a cost per scene before anything is spent.
Cast sheetOne locked sheet of the characters and a fixed voice for each. You approve the faces.
Scenes, one at a timeStills, then one clip per line, each checked word by word, each voice converted, each scene measured and reviewed. You lock each scene before the next one starts.
The cutThe scenes joined with sound that crosses each cut, captions in the script's wording, and the end card with your offer and one cleared result. Measured before you watch it.
1080 deliveryThe picture alone is upscaled, then captions and the end card are drawn fresh at 1080 so the text stays sharp. Every shot change is checked against the 720.
Cutdowns and new hooks, when you askBuilt from the scenes you already locked. No new clips.

What you get

The finished ad at 1080 and at 720 (the 720 is small enough to send to a phone)
A measured QC report for each
The script, storyboard and project file, so any version can be rebuilt
Cutdowns and hook variants on request

Honest limits

Read this before you rely on it

It costs real money: every clip and still is paid generation. You approve each batch and see the price first. The first six-scene ad cost about USD 45 in generation including retakes; budget about 20% for retakes.
Only Seedance 2.5 has been tested for the talking clips. Another model is a new test: run one scene end to end first.
The voice and review tools are fixed by the scripts. Changing them means changing the scripts.
No cleared result means no proof panel on the end card. It never invents one.
Each scene takes several rounds of generation and checking. It is a production, not a button.

Where people go wrong

The mistake	Do this instead
Asking it to skip straight to generating	Approve the script and storyboard first. A change after the clips exist costs a regeneration.
Fixing a bad line by cutting inside it	Regenerate the line. A cut inside speech is heard.
Adding room tone under clean voices	Leave it out. A generated tone can carry a buzz that plays under every pause.
Paying for a new upscale after trimming	A version that only trims scenes reuses the old upscale at no cost.

Worth knowing

Set your own rules once in brand/_animated-story-ad.md (your media folder, music, how you approve spend) and per client in brand/[workspace]/animated-story-ad.md. The skill reads both at the start of every run.