The AI Filmmaking Checklist That Saved Me Weeks of Regeneration

The expensive part of AI filmmaking is not generating the shot. It is generating it again. A few months into making Lost Garden , my dark-fantasy anime series built entirely with an AI workflow, I generated a full scene. Eighty-something shots. It looked good in isolation. Every clip, on its own, was the kind of frame that gets likes. Then I cut them together. The protagonist’s face had quietly drifted across the scene. Her jacket went from deep violet to a washed-out grey somewhere around shot forty. The light that was supposed to come from a single lantern was coming from three different directions depending on the clip. None of it matched. I had a folder full of beautiful shots and not a single usable sequence. I threw most of it out and started again. Not because the model was bad. Because I had started generating before I had decided anything. That is the real lesson nobody puts in the tutorials: in AI filmmaking, generation is cheap and regeneration is the tax. You do not pay for the first version. You pay for the fourth, the seventh, the one where you finally realize the problem was never the prompt, it was that you never locked the world the prompt was describing. The film is not won at the generation step. It is won in the boring hour before it, the one where you decide what is allowed to exist. So here is the checklist I now run before I generate a single shot. It is not glamorous. It is the difference between a weekend of work and a month of re-rolls. \ Why pre-production matters more now, not less There is a myth that AI removes pre-production. You just type and it appears, right? In practice the opposite is true. When a camera crew costs thousands of dollars a day, you are forced to plan, because the shoot is expensive. When generation costs cents, the temptation is to skip planning and brute-force it with volume. That temptation is the trap. Volume without a locked plan just multiplies your inconsistencies. A hundred shots generated against a vague idea give you a hundred slightly different films. The tools themselves have been pushing in this direction. Character reference systems, what people in the space call cref , now let you anchor a generation to a fixed visual identity instead of re-rolling a face from scratch every time. Runway’s recent reference-driven consistency and Kling’s dedicated face-reference system both exist for exactly one reason: filmmakers kept losing their characters between shots. The tooling is admitting what the checklist already knows. You need a fixed thing to point at before you generate. \ \ The lock list: what to decide before you generate anything I think of pre-production as a set of locks. Each one is a decision you make once so the model cannot quietly re-decide it for you, shot by shot. Here is the order I run them in. 1. Lock the script as a shot list, not a screenplay. A screenplay describes a film for humans. A shot list describes it for a pipeline. Before generating, every beat should be broken into discrete shots with: who is in frame, the camera size (wide, medium, close), the action, and the emotional point of the shot. If you cannot say in one line why a shot exists, it does not exist yet. 2. Build a character bible. This is the single highest-leverage hour you will spend. For each recurring character, lock: A clean reference image of the face, front and three-quarter Wardrobe, with specific colors named, not “purple” but the actual swatch Distinguishing details: a scar, a prop they always carry, hair behavior The character’s default lighting, so they read the same across locations Feed this into your tool’s reference system instead of re-describing the character in prose every time. Prose drifts. References hold. 3. Write the world rules. Locations, recurring props, time of day, weather, and above all the color palette . A world that is allowed to be any color will be every color. Decide the three or four hues the film lives in and treat anything outside them as a defect to be re-rolled, not a happy accident to keep. 4. Choose one consistency strategy per element, on purpose. Identity and motion are different problems and the better workflows separate them. Decide before you generate whether a given character is anchored by a face reference, a seed, a full character-reference image, or a combination. Mixing strategies mid-scene is how faces drift. Pick the anchor, write it down, do not improvise it at 2am. Treat every parameter you leave undecided as a parameter the model will decide for you, differently, on every single shot. 5. Lock the grammar of the image up front. Aspect ratio, lens language, and grade are not per-shot choices, they are film-wide ones. A 2.39:1 anamorphic look does not survive being remembered for some shots and forgotten for others. Pick the frame, pick the focal-length feel, pick the color grade, and apply them as constraints on everything. \ The half of the checklist everyone skips The locks above are about the image. The next two are about everything else, and they are where most AI films quietly fall apart. 6. Plan the sound before the image, not after. Sound is not post. It is a pre-production decision. Knowing a scene is carried by a single sustained drone, or by silence broken once, changes how you shoot it. If you generate first and score later, you end up bending great audio around shots that were never built to hold it. Decide the sonic shape of the scene while it is still cheap to change. 7. Plan the assembly, because meaning lives between the shots. A beautiful shot means nothing until it sits next to another one. The cut is where emotion is actually manufactured. So your shot list should already imply the edit: this close-up exists because it follows that wide, this reaction exists because of the line before it. If you cannot describe how two shots talk to each other, you are generating decoration, not a scene. 8. Set a “good enough” bar and obey it. This is the discipline one. Because regeneration is nearly free, you can re-roll forever, and forever is not a schedule. Before a scene, decide what “done” looks like: continuity holds, the emotional point lands, no obvious artifacts in the focal area. When a shot clears the bar, move on. Chasing a marginally prettier version of a shot that already works is the most expensive habit in this medium, because it spends the one resource AI does not give you back, which is your attention. \ Where the tools fit I will name the obvious ones, because the checklist is tool-agnostic but the locks have to live somewhere. For generation and reference-driven consistency, Runway and Kling are the two I reach for most, the former for granular control and reference consistency, the latter when facial consistency across shots is the whole ballgame. For stylized motion and quick look-development I keep Higgsfield in the rotation. But none of those tools hold your plan . They hold your output. The character bible, the world rules, the shot list, the decisions, those have to live in one place that survives a single creative head juggling all of it. That is the gap I kept hitting on Lost Garden, and it is the reason I am building ScreenWeaver , a workspace where the script, the characters, and the world get locked before anything is generated, so the generation step inherits a fixed world instead of inventing a new one every time. The point is not the tool. The point is that the plan is a real artifact and it deserves a real home. A folder of loose reference images and a vague memory of “the palette” is not pre-production. It is hope. \ The actual checklist, on one screen If you copy one thing from this piece, copy this. Run it before every scene. Shot list done, every shot justified in one line Character bible locked: face refs, named wardrobe colors, signature details World rules written: locations, props, palette of three or four hues Consistency anchor chosen per character, on purpose, written down Image grammar fixed: aspect ratio, lens feel, color grade Sound shape decided before generation Assembly logic implied by the shot list, cut by cut ”Good enough” bar set, and obeyed It takes about an hour. It saves about a month. The filmmakers who win this era will not be the ones who generate the fastest. They will be the ones who decided the most before they pressed go. \ \ \