Module 7. Every take you own, in order, on a timeline, without you dragging a single file. And then the tool stops, because the next part is the part you are actually for.
Not a cut. The film in order, at full length, with nothing chosen yet.
Every shot in shot order, its selected take, its dialogue, its cue, on a timeline you can scrub. In a normal post workflow an assistant builds this and it takes a day.
You have a file that already knows all of it.
You have 300 files with names like kling_20260803_0417_v2_final.mp4.
The timeline needs them in story order, with the right take of each.
That is a lookup, not a decision.
The bridge is an edit list. Every NLE on earth imports one:
"Build an EDL fromshots.json: shot order, the latest approved take of each, its duration, and the audio layers for that shot. Write it toedit/assembly-v1.edland tell me anything you could not resolve."
"Tell me anything you could not resolve" is the load-bearing clause. A shot with no approved take must appear as a hole you can see, not as a silent omission that shortens your film by four seconds.
Generative footage arrives in whatever the model felt like. Mixed codecs on one timeline is how an edit starts stuttering and you blame your machine.
"Make proxies of every approved take at a single edit codec and resolution, into proxies/, keeping the original filenames. Skip anything already done. Report what failed."
Same queue pattern, third time. By now this should feel unremarkable, which is the point.
If your film has dialogue or narration, do not open the NLE first. Cut it in text.
"Transcribe every voice take with timecode. Then give me a paper edit: the transcript in shot order, so I can delete lines to build the cut."
You delete the lines that do not earn their place. Then:
"Build a new EDL from what survived in the paper edit."
You cut the film in a text file, in a sitting, and the timeline follows. This is how documentary has been cut for thirty years and it maps onto generative work perfectly.
Never point two agents at one NLE instance. Not two tabs, not two sessions, not one agent while you are also clicking.
An editing application holds your project in memory and writes it when it feels like it. Two things driving it at once do not collide loudly. They collide quietly, and you find out three days later that a timeline is missing shots you know you cut.
Whatever is driving the NLE, it is one thing at a time, and it is yours or it is the tool's, never both.
Your takes are already versioned. The project file is the thing you have not been protecting.
Before any automated pass touches the project: copy it, dated, and say so. Then work on the copy.
"Duplicate the project file to edit/versions/ with today's date before you change anything, and tell me the path you wrote."
A project file you can roll back turns every scary automated step into a cheap one.
The limit, one last time, and this is the module where it matters most.
It cannot feel a scene run long. It cannot tell you the film dies in the middle, that the turn lands two shots too late, or that the best take is the one with the flaw in it.
It has no taste. It has never been bored.
Plenty, and none of it is the cut:
assembly-v3 and assembly-v4"Compare my current timeline against shots.json and tell me what is on the timeline that should not be, and what is missing that should be there."
Continuity checking, which is a real job, done in seconds and done exactly.
The filmmaker cuts.
The tool prepares, organises, drafts, checks, converts, verifies and accelerates. It does not have taste and it does not make the decision.
Your film from Module 6. Do not open your NLE until step 4.
proxies/. Skip what is done."shots.json and report what is missing or superseded."Steps 3 and 6 are the film. Everything else is the assistant work you no longer do.
proxies/ so the timeline actually playsedit/assembly-v<n>.edl, versioned, each one rebuildable from the fileYou have a cut, and the day of assistant work in front of it did not happen.
Sell. Delivery specs that pass, the versions every platform wants, and getting the finished thing in front of people without a week of exporting.
Verified against Claude Code 2.1.220 · 2026-07-30.