Case study · studio.santiagomorel.dev

Sidequest Studio: how much of a creative studio can AI agents run?

A side project that started with one capybara plush: four AI-run characters, each with a written “bible” for how it looks and talks, posting to three platforms most days — with a human approving every single video before it goes out.

Characters 4 Platforms Instagram, YouTube, Facebook Stack Claude Code, Claude, OpenAI images, ElevenLabs, Remotion, Python Approval gate every post, by email
4
Characters
3
Platforms
165+
PRs merged
0
Posts without a human yes
The pipeline

Idea to published post, seven steps

Every character has a written “bible”: how it looks, how it talks, what it would never say. AI agents follow it at each step, and check each other's work — but nothing reaches a feed without a person clicking approve first.

01 · Idea

An idea goes on the schedule for a character.

02 · Script

Claude writes it in the character's voice and checks it against the bible.

03 · Pictures

An image model paints the scenes. Another AI checks them for wrong details, like a missing leg or stray text.

04 · Voice & motion

Each character has its own AI voice, and the scenes get moving.

05 · Edit

Code (Remotion) puts it all together: cuts, captions and on-screen text.

06 · A human says yes

The finished video arrives by email. Approve or reject. Nothing is posted without that click.

07 · Posted

It goes out to Instagram, Facebook and YouTube on schedule, and the numbers come back on their own.

The cast

Four characters, four bibles

Each character's bible is what keeps its voice consistent across a hundred-plus independently generated videos — the same way a style guide keeps a brand consistent across writers.

Pancracio

Calm Capybara
Calm Tea Hustle
“Other people's choices are wind. Stop fighting. Sit down. Drink your tea.”

Eusebio

Black Sheep Explorer
Curiosity Maps read upside down Following the flock
“Eusebio asked why. Then he asked why not. The second answer was more interesting.”

CoderAssist

Proud AI Assistant
Typing speed Knows the coder is the hero Pride
“Writing code was never the goal. Solving problems was. Code was just the most direct way to do it.”

Gervasio

Gentle Giant
Strength Hugs Rushing
“Strong people get tired too. Rest is how the strength comes back.”
What broke, and what it taught

Outtakes from the build

Real lessons, not a highlight reel. Nobody got it right the first time — not even the AI.

The AI couldn't spell

Asked to draw the character and the quote at once, the image model mangled the words and moved the layout on every fix.

Fix: the AI paints only the picture now; code types the words.

The most expensive incense stick

One small stick of incense kept smoking from the wrong place, once straight out of a bonsai. A post went out like that and had to be deleted.

Fix: after seven rules just for smoke, the incense was retired.

Eye contact was scary

A close-up of a plush toy looking right at you feels like a horror movie.

Fix: eyes-open shots are now always wide or medium, never a close-up.

The sheep lost his legs

In a side view, Eusebio's front legs quietly vanished.

Fix: every prompt now asks for all four hooves, and the first frame is checked before paying for the clip.
Synthesis

Why this pipeline, not a simpler one

Every stage earns its place against a specific failure mode the earlier version of the pipeline actually hit — not added because it sounded sophisticated.

A written bible per character
Keeps voice and visual identity consistent across a hundred-plus independently generated videos, the way a brand style guide keeps consistency across writers.
AI checking AI
A second model reviews every generated image for wrong details before it moves downstream — catching the missing-leg and stray-text failures instead of shipping them.
Code, not a model, for on-screen text
Image models can't reliably spell. Moving text rendering into Remotion (deterministic code) instead of asking the image model to paint it removed an entire class of bugs.
A human approval gate, no exceptions
Every video arrives by email for a yes/no before it's posted anywhere — the incense-stick outtake is exactly the kind of mistake this gate exists to catch, even when it occasionally doesn't.
Check the first frame before paying for the rest
Video generation is the most expensive step in the pipeline. Verifying the first frame (after the sheep's legs vanished) catches a bad generation before spending on the full clip.

The human in the loop

Sidequest Studio started with one capybara plush and a question: how much of a small creative studio can AI agents run, if a person still makes the calls? Four characters and 165+ merged PRs later, the answer so far is most of it — with a lot of careful checking.