Skip to content

When the AI gets stuck

As you make video, you hit spots where the AI workflow itself stalls. It’s not the video content that’s stuck, it’s the flow that stopped. Here are a few you’ll meet often, and the way out.

This is the spot that stops most often on the remix track. If the source has no license label, the gate refuses to proceed. This isn’t a bug; it’s a deliberate halt that blocks copyright accidents.

There are two ways out. Switch to a source with a clear license (e.g., a CC-licensed video, your own video), or explicitly accept responsibility. Responsibility acceptance is planted in license_responsibility.json, meaning the copyright responsibility is on you. Forcing a video of unclear source through responsibility acceptance can get you a copyright claim later, so it’s safer to confirm the source first.

When the AI stops and asks on ambiguous input

Section titled “When the AI stops and asks on ambiguous input”

If you say only “make a short-form video,” the AI doesn’t proceed and asks back. Is there a URL? Is it a character and scenario? This isn’t stuck; it’s a deliberate spot. A wrong track choice throws off the whole flow, so it doesn’t enter on a guess for ambiguous input.

The answer is simple. To rework an existing video, attach a URL; to build a new video, give a signal like “original” or “character.”

When image style breaks from scene to scene

Section titled “When image style breaks from scene to scene”

If a character comes out with a different face from scene to scene in the generative track, the character sheet may not be set. When there’s a recurring character, you have to register the character in scenes.json and make the sheet first, so the same face is kept across every scene. Say “make this character’s sheet first” to set the sheet and regenerate.

Codex calls are set up to gather in one place, image-director, and this is the spot that keeps style consistent. If images keep breaking, check the sheet first.

If the first video takes too long, check whether you’re in preview mode. Preview runs at a small resolution with a fast image backend, but if it accidentally runs in final, it’s much slower. Say “run it in preview” to return to the fast mode. Setting REPLICATE_API_TOKEN makes preview images faster. Final runs once after all decisions are done.

When a session broke and you don’t know how far you got

Section titled “When a session broke and you don’t know how far you got”

The session-start hook pulls the active video’s progress automatically. If you’re still confused, ask “which video, which stage am I at?” It pulls together videos.json and the active video’s progress.md to tell you. Writing the next starting point into active.md when you finish for the day makes it smooth.

Copying the error message verbatim, pasting it into Claude Code, and asking “why is this happening?” is often the fastest. If that still doesn’t work, find the same symptom in troubleshooting, and if it’s not there either, open a GitHub issue.

A collection of common sticking points is in troubleshooting.