ai-film-director

The one rule that matters

Each stage takes the previous stage's finished file as its only input. If that file does not exist, the stage cannot run. Say so and go back. Do not improvise a substitute.

This is the whole reason the pipeline works. A model asked to "just make the video" will happily generate 40 clips that do not cut together. A model that cannot start Step 6 until Step 5's music timeline exists will generate the right 12.

The pipeline

Routing

Read what the user is actually stuck on, then load exactly one stage skill.

• "Where do I start" / a subject but no story → Step 1, film-research.
• Facts but no story → Step 2, film-story.
• "My character looks different in every shot" → Step 3, character-sheet. This is the single most common failure and it is never fixed downstream.
• "My images look flat / not cinematic / like a stock photo" → Step 4a, cinematic-frames.
• "The colour looks cheap / washed out" → Step 4b, film-lut.
• "It looks like a slideshow" → Step 5, film-music. A slideshow is a pacing problem, not an image problem.
• "How do I animate these" / "which video model" → Step 6, film-motion.
• "Every character sounds different" → Step 6, film-motion, voice section.
• "I have clips but it does not feel like a film" → Step 7, film-edit.

The three failures this pipeline exists to prevent

1. Character consistency. The hero's face changes between shots. Fixed only at Step 3, by a character sheet built from measured facial geometry.
2. Voice. Every generated clip gives the same character a different voice. Fixed only at Step 6, by dubbing and voice-converting one performance, never by trusting the model's audio.
3. Shots that ignore the direction. The camera obeys and the hair does not, or the crowd turns and stares into the lens. Fixed at Step 4a by stating camera position, framing and eyelines as positive prose inside the prompt.

If the user is describing one of these three, name it, then route to the stage that owns it.

Order exceptions worth knowing

Music comes before video. This looks wrong and it is correct. Cutting the stills to the music first tells you exactly which shots are missing and which movements each shot needs. Generating video first means paying for clips that the edit then throws away. AI video generation is a lottery: every clip costs money and roughly one in three is usable.

Research outranks the story. A beat that turns out to be historically false has to be cut after everything downstream has been built on it. Cheapest place to catch it is Step 1.

When the user wants to skip

They will want to skip Step 1, Step 3 and Step 5, in that order. Each skip has a specific downstream cost. Name the cost, offer the fast version of the step, and let them decide:

• Skipping research → the film contains something false and a comment section finds it.
• Skipping character sheets → every later image is unusable and must be regenerated.
• Skipping music-first → clips get generated that the edit deletes.

State the cost once, plainly. Then do what they decide.