10. A Production Folder You Can Rerun
Summary
Finding and redoing one shot
The courier short is finished, and a note comes in a week later: the name plate insert should be brass, not painted. The chapter shows how to trace that shot, redo it and drop it back in without disturbing the rest. It is a walkthrough on the show's illustrative courier project. Every name and figure is an example, not a record of finished work.
Shots live in one folder per scene and shot, not sorted by date or tool. File names follow the fixed-field pattern in Netflix's VFX shot and version naming guidance: project, scene, shot, a marker for generated clips, then a zero-padded version. Each clip gets a sidecar text file with the exact prompt, model and version, input files, date, parent take, credits and licence or plan. Runway's hosted models may not return the same clip from the same inputs, so the accepted file stays the master and is never overwritten.
For the brass plate, you change only the plate description, keep the lighting words from the production bible, and save the new take as v002 with its own sidecar. Rejected takes are kept and marked, since they count toward cost per usable shot. In Shotcut, Replace swaps the clip on the timeline and keeps its filters. v001 stays in its folder in case the note is reversed.
Saved exports and the vertical cut
Export settings are made once as a custom preset, following YouTube's recommended upload settings: MP4, H.264, AAC. Shotcut's export preset thread covers saving one. Resolution, aspect ratio, frame rate, colorspace and audio channel lines are deleted so each project supplies its own. Advanced mode can switch off after a restart, so check the fields before exporting.
The vertical version is a copy of the project. Set Vertical HD in Video Mode first, then reframe each shot. The wide walk needs keyframed position or a blurred pad, and titles need checking. Exports are named by project version and shape, with a log line for each.
Three ways to rerun shots later are compared. Runway's Workflows can lock a node's output, but availability needs checking on your plan. The Runway API reports credit costs and has tier limits. ComfyUI stores the workflow in each output image and reruns most faithfully with a fixed seed and the same model files, at the cost of the most upkeep.
Multi-model platforms this month
No major release with a primary source turned up for early November 2026. One weak-sourced roundup describes platforms bundling several models behind one credit balance. A2E AI's press release claims a generator with 151 video models, which is the company's own claim. Because a platform can hide which model made a shot, the sidecar's model line matters. The suggested test is one take from the name plate's inputs, logged the same way and checked against v002.
Tracing and rerunning one shot
The courier short is finished. Now suppose a note arrives a week later: the name plate insert at the lit door should look different, maybe a brass plate instead of a painted one. Can you find that one shot, redo it, and drop it back into the edit without breaking anything else? You can, if two things are true. Every accepted clip sits in a named folder with a written record of how it was made. And your exports are saved settings you can call up again, not choices you rebuild from memory each time. This chapter sets up both on the show's illustrative courier project. The walkthrough describes what you would do. It is not a record of work already done, and every name and figure in it is an example.
Start with the folders. Organize by scene and shot, not by date and not by which tool made the clip. Dates scatter one shot across many folders. Tool names stop making sense the moment you switch tools. So inside the project folder, make a folder called shots, and inside that, one folder per shot: s01 for scene one, then sh01 through sh09 for the nine shots in the order of the shot list. The name plate insert is shot nine, so its folder is courier, scene one, shot nine.
Next, file names. Film and television visual effects teams solved this long ago. Netflix's Backlot guidance builds a shot name from fixed fields, the show, the episode, the scene and the shot number, and then adds a task and a version number padded with zeros, like v001. The fixed fields come first, so files sort together by shot. The version comes last, so takes sort in order. Adapt that pattern here. The first accepted name plate take becomes courier underscore s01 underscore sh09 underscore gen underscore v001. "Gen" marks it as a generated clip. The References still that started it gets the same name with "ref" in place of "gen". Netflix's guidance also notes that editorial usually assigns shot names. In a one-person studio that is you, and the shot list is where the numbers come from. Pick your field order and padding once and never change it.
Now the record. Beside each generated clip, save a small text file with the same name, ending in dot json or dot txt. People call this a sidecar file, because it rides alongside the clip. There is no published standard for what an AI clip's sidecar should hold, so here is a practical set. Write the prompt exactly as sent. Write the model name and version, for example Gen-4.5 Image to Video. List the input files by name, such as the References frame and the reference sheet crops. Add the date, the take this one came from if it was a revision, the credits it cost, and the plan or licence you generated under. That last line matters later. If someone asks whether you had commercial rights on the day you made a clip, the answer is sitting next to the clip.
One more thing the sidecar cannot do: guarantee a perfect rerun. Runway's hosted models may not give you the same clip twice from the same inputs, and there is no seed setting you can count on to lock the result. A seed is the number some tools use to repeat the same random starting point. So the original output file is the master record. Never overwrite it. The sidecar tells you how to make something close. The file is what you actually accepted.
Here is the trace for the brass plate note. Open the shot list and find row nine, the name plate insert. Go to its folder. The sidecar for v001 shows the motion prompt, the References frame it started from, and the lighting words from the production bible that fixed the wrong-side light in chapter seven. You change only the plate description in the References prompt, keep the bible's lighting words, and generate a new first frame. Then you animate it with the same motion prompt. Say the first take passes inspection for light, framing and duration. Save it as v002 with its own sidecar. That sidecar names v001 as its parent and logs its credits. A take you reject gets saved too, marked rejected, because it counts toward cost per usable shot.
In Shotcut, open the finished project. Select the shot nine clip on the timeline, and use the Replace command from the Timeline menu, or right-click the clip, with v002 loaded in the Source player. Replace keeps the clip's place and its filters, so the correction and the shared grade stay on it. Check that v002 is the same length. If it runs long, trim it to the original in and out points. Watch the cut across the door, then save the project with a new version number. Keep v001 in its folder. If the note gets reversed, you are one swap away.
The second half of the folder is exports. Every time you deliver, you make the same choices about codec, bitrate and audio. Make them once and save them. In Shotcut, click Export in the toolbar, then Advanced at the bottom of the panel. Match YouTube's recommended upload settings. Its help page asks for an MP4 file with H.264 video at High Profile, progressive, with the index at the front of the file, which YouTube calls Fast Start. For 1080p at twenty-four to thirty frames per second it suggests about eight megabits per second. Audio should be AAC at forty-eight kilohertz, stereo, at three hundred eighty-four kilobits per second. Color for ordinary video is the standard called Rec. 709. Shotcut's field names don't line up one for one with YouTube's list, so check each tab. Format and codec set H.264 and AAC. Video holds the frame settings. The rate control tab is where the bitrate goes.
Then press the plus button under the preset list. A text box opens showing every setting. Delete the lines for resolution, aspect ratio, frame rate, colorspace and audio channels, so each project supplies its own. Name the preset something clear, like YouTube H264 AAC, and click OK. If the name is empty, nothing saves. The preset now appears under Custom. Two cautions. Advanced mode can switch itself off after a restart. And settings you change after picking a preset aren't saved unless you save the preset again. Before each export, glance at the fields to confirm the preset actually loaded.
The vertical cut-down was left as an option in the last chapter. Here it is. Save a copy of the finished project under a new name, so the landscape edit stays untouched. Then, before you change any clip, open Settings, Video Mode, and under Non-Broadcast choose Vertical HD at thirty frames per second. That gives a frame of one thousand eighty by one thousand nine hundred twenty, a nine by sixteen shape. If your landscape project runs at a different frame rate, use Custom and Add instead, and set the same size at that rate, because YouTube asks you to keep the original frame rate. Set the mode first. Changing it after you reframe can leave Size, Position and Rotate filters in the wrong place. Then reframe each shot. For most of them, use Size, Position and Rotate to scale up and slide the frame onto the courier. The wide walk is the hard one, because she crosses the frame toward screen right. Either keyframe the position to follow her, or use Blur: Pad, which keeps the whole landscape frame and fills the top and bottom with a blurred copy. Titles and credits need checking too, since text sized for a wide frame can run off a narrow one. Export with the same custom preset. Because you deleted the resolution lines, the vertical project sets its own size. For YouTube Shorts, one thousand eighty by one thousand nine hundred twenty fits. The maximum length has changed over time, so check YouTube's current Shorts help page before you upload. The courier short sits under a minute either way.
Name each export like courier underscore v003 underscore youtube16x9, and the vertical one with 9x16 in place of 16x9. Put both in an exports folder. Add one line to the production log for each: which project version, which preset, and the result of a watch check on phone and computer. That log is the routine. The next delivery means picking the preset, exporting, naming and watching.
Now try the routine on one clip of your own. Pick a clip you have already accepted, give it a shot folder and a versioned name, and write its sidecar from whatever its generation details still show. Where a field can no longer be recovered, write unknown rather than guessing. Then export it once with your saved preset, name the export and add its log line. If you can find that clip and its record again tomorrow without searching, the folder works.
Who could rerun a shot like the name plate for you later? There are three candidates, and they differ in setup, repeatability and upkeep. Runway's in-app Workflows connect inputs, models and language-model nodes, boxes that each do one step and pass their result along, on a canvas with no code. You can lock a node so a good output is reused instead of regenerated. Access was limited when it launched, and its current availability and pricing on your plan need checking in your account. The Runway API is the second choice. It has an image to video endpoint, reports the actual credit cost of each task, and bills per second of video on every attempt, rejected ones included. It also has tier limits. At the entry tier you get one or two tasks running at once and a monthly spend cap of one hundred dollars. Extra tasks wait in a queue. You write the code and keep the inputs yourself, and there's no seed to rely on. ComfyUI is the third. It's a free node-graph program that runs models on your own graphics card or a rented one. It saves the whole workflow inside each output PNG image, so dragging that image back onto the canvas restores the graph. With a fixed seed and the same model files and node versions, it reruns most faithfully of the three. It also takes the most setup and upkeep, since updates can break old graphs. That last option is where this goes next: a saved ComfyUI graph that makes approved variations from the courier's reference pack.
Where AI video access is heading this month
The news this month is thin. No major release or policy change from Runway, Google, Luma, Pika, ComfyUI or Adobe surfaced for the first half of November 2026 with a primary source behind it. What does show up is a direction. Production access is moving onto platforms that bundle many models behind one credit balance. A startup blog's roundup from November sixth, a weak source, describes Higgsfield offering Seedance, Kling, Veo, Wan and Sora models together. It also describes VEED letting you switch models with free quotas and paid credits beyond them, and invideo connecting to Veo 3.1. Separately, A2E AI says in a press release that it launched a video generator on November fifteenth, 2026, with a hundred fifty-one video models. Treat that number as the company's claim until you see it working.
Why this matters for the folder you just built: when one workspace hides many models, the sidecar's model and version line is what lets you know which one made a shot. The smallest useful test is to take the name plate's References frame and motion prompt and run one take on a multi-model platform you already have access to. Log it in the same sidecar format, with the platform, the model, the credits and that plan's licence terms. Then compare it against v002 on the same checks: light from screen right, a readable plate, and a clean join at the door.
Want a Series of your own?
Gnothi wrote and narrated this chapter. Name a subject and it writes you a course, or a story that teaches it, one chapter at a time, delivered to your podcast app.
Get started free