How Is an AI Film Made? The Planet Hell Production Workflow

How do you make an AI film end to end?

You build a planning layer first, then generate, then fix what the models got wrong. For Planet Hell episode 1, that meant an outlined 10-episode roadmap, a Gemini pass to organize two years of notes, an asset-tracking "second brain" that a coding agent could read, video generation driven through OpenArt's MCP, and post-production repairs in After Effects. The sequence is planning, references, generation, rerolls, and cleanup. Most of the saved time came from the first and last steps.

The creator behind the series described this pipeline in a behind-the-scenes breakdown. OpenArt sponsored the film, so treat the platform praise with that in mind. The workflow lessons stand on their own.

How do you plan a series before generating anything?

The creator outlined a ten-episode roadmap. It is not fully scripted, but each episode works as its own mini movie with a narrative thread running through all of them. The model is closer to a comic book than a TV schedule, and there is no fixed release cadence. The source cites Warhammer and Metal Hurlant (Heavy Metal) as influences.

The raw material was the problem. Two years of tests, ideas, and documents had piled up, much of it generated with Claude. The creator handed it all to a Gemini model (named Gemini 4 in the video) and used it as an editor. Their view is that Google's models are good writers with strong storytelling instincts. The human still shaped the result, but the model turned volume into structure.

What does the asset tracking setup look like?

The creator calls it a "studio" widget: a second brain plus asset tracker. It stores the script, character and location references, a lookbook, and a preferred prompt style. A coding agent such as Claude Code, Codex, or Google's Antigravity can read all of that and then act on it.

The point is context. When the agent drives generation, it already knows what the scene needs and which references apply. You are not re-explaining the film in every prompt.

What is OpenArt's MCP and how did it fit in?

MCP (Model Context Protocol) lets an agent drive a platform's tools directly. In this production, OpenArt's MCP let the agent run generations while the second brain supplied the script, references, and prompt style.

The creator did not use it for everything. Some jobs were faster by hand in the web interface. One example was a reference-to-video fix where a character held a bullet-shaped pen. Uploading the clip and asking for a black pen in the man's hand was simpler than routing it through an agent, and it saved tokens.

The source also mentions a few other OpenArt features:

• Director mode: a built-in alternative to a custom second brain, with a full timeline. The creator says it avoids rebuilding your own studio brain repeatedly.
• Arena: a model leaderboard useful for tracking what is currently strong, including for video editing.
• Audio tools: voice-overs, music, sound effects, voice cloning, and voice changing.
• 10-bit output: the creator reports Seedance generations on the platform come back 10-bit, which helps if you plan to color correct.

Which reference tools helped with look and consistency?

Several tools fed the lookbook and the shots:

The ShotDeck step is easy to copy. Real stills with real technical metadata give both you and your agent something concrete to match instead of vague style words.

What went wrong, and how was it fixed?

Character likeness. The returning lead was the hardest to get right. He read as "Tom Hardy adjacent," and the video model pushed against him. The fix was odd: running the original character image through Nano Banana. The creator admits not knowing why it worked, but it did.

Recasting. A previously tested actor for the character Sunny did not work in real production. The channel's recurring character Flamethrower Girl took the role after a wardrobe change in OpenArt. The creator also thought the recast helped a plot beat in the first half.

Look changes. The trailer's biomechanical, Giger-like creature design was swapped for something more organic and stranger.

3D previs. Early scene tests in Cozy Clay and three.js disappointed. The creator felt they had let Claude and Codex take too much control, and they were spending more time fiddling with previs than making the film. Most of it was dropped.

Stray objects. The video model gave a character a slung rifle, and another shot had an oversized bullet. Rerolling risked losing a good take, and the model might not even respond to the correction. These went to After Effects, where Claude handled the rotoscoping. The result was close to clean, with one frame where the object flickered back in. The creator chose not to chase it.

How many clips does an AI short actually take?

The project bin ended with 79 clips, and 59 were used in the cut. A good share of the 79 were false starts and tests. Not every used clip was a full-length generation, either. Action beats were often cut from shorter 10-second generations, and the creator tries to get as much as possible out of each one.

On rerolls, the creator says Seedance usually delivers something workable in about two rolls, sometimes two or three. It costs more per generation, but fewer rerolls offset some of that.

One craft note: the creator deliberately gave the actors "business," meaning props and blocking, to avoid two characters standing still and talking. A boot gag and a coffee mug gag are the examples cited.

Should you copy this workflow?

Parts of it travel well. A structured asset tracker, a reference lookbook, and agent-driven generation help any multi-shot project. The skepticism about heavy 3D previs is also useful if you find yourself tuning a pipeline more than making shots.

Two caveats. The platform advice comes from a sponsored video. And the source material available here does not cover the color grading and audio passes in detail, so this article does not describe them beyond the 10-bit note.