Chapter 29 · Editing AI footage, and judging the work
Part 6 · Edit, finish and delivery
Generated takes arrive as files that disagree with their labels, with holds nobody asked for and defects that start halfway through. None of that changes what an edit is for: an audience learning the right thing at the right moment. This chapter teaches how to log and cut such footage, how to judge it in the running cut, how to send a precise request back to generation when the cut is missing something, and when to stop repairing and lock.
In this chapter
Who decides what in the cut: the assistant, the director and the editing application
What must exist before the first cut, and a thirty-second commercial worked in frames
Logging a take by its usable interval, and curating a batch
Assembly A, the first causal cut; alternates; J and L cuts; joins across a contact; rhythm
Judging in the running cut: checking the file, the picture and the sound, and the verdict
The loop back to generation: the request, the missing-coverage list, handles
Diagnosing before you spend, the ladders, when to stop, and locking the picture
Before you start. Chapter 8 writes the obligations list, the audience information the cut must keep, against which every assembly here is judged. Chapter 14 shows what a clip gives the edit (its holds and its exit state) and how to probe a file, and Chapter 21 who owns each sound. Chapter 30 finishes the locked cut, Chapter 31 hands it over, and Chapter 32 is the Premiere Pro section.
29.1 What the edit decides, and who decides it
The edit lives in Adobe Premiere Pro (Premiere, from here on) on the director's machine, with an AI assistant, Claude or ChatGPT, working as assistant editor. The division of labour is strict, and it is the first thing to settle.
The assistant logs only frames it has actually sampled and audio it has actually heard; proposes structures and alternates; does the frame arithmetic; writes the edit plan; has a program build the timeline file from it; writes notes and replacement briefs; measures duration and loudness.
The director decides performance, taste, dialect, lip sync, emotional timing, the final select and the lock (the point after which the cut no longer changes).
Premiere executes and confirms. A timeline file that loads, a passed check or a contact sheet never passes a cut.
No maker documents editorial craft. What follows is ordinary professional practice: assemble for meaning, keep screen direction, treat J and L cuts as options, and let the audience's information, not the file count, say when a scene is done.
Word | In this book |
Probe | measuring a file with a tool such as ffprobe: size, frame rate, frame count, length, audio streams |
Interval | the part of a take the cut uses, written as frames: in (included) and out (excluded), counted from 0 |
Log | the record of a take's events at the frames where they happen, and of the frame where any defect starts |
Hold | the frames a model adds after the action has finished |
Handle | usable frames outside the interval, which a trim or a dissolve can borrow |
Edit plan | the written cut: every placement with its source range, sequence range, track and reason |
Assembly A | the first causal cut, on hard cuts, before any finishing |
Slug | a placeholder in the timeline marking coverage that does not exist yet |
Alternate | a duplicate of a cut with one thing changed |
J-cut, L-cut | a cut where the incoming sound leads its picture (J) or the outgoing sound carries past its picture (L) |
Join | the cut between two shots, looked at as a pair |
Picture lock | the point at which the picture is agreed and every later step works on it without changing it |
Four ideas govern the chapter.
The assistant chooses meaning and sequence; a program writes the timeline file; the editing application confirms the result. A project file typed freehand in a chat can fail in ways nobody sees until it is opened, so a small program writes the file from the edit plan, the same file from the same plan every time.
Keep three numbers apart: the seconds you asked for, the frames the file has, and the interval you use. No universal count exists. Five-second requests to Kling and Seedance have come back as 121 frames at 24 frames a second, not 120; that is an observation, not a constant, and other durations and models may differ. The interval you select is a third number again. Probe every file, and no edit plan can drift by a frame or ask for footage that does not exist.
Write a note as a cause, never an adjective. "Her recognition occurs after we have already shown the answer" tells the next step what to change; "pacing weak" does not.
A valid timeline is not a successful film. A file that parses or a check that passes says nothing about performance or idea. Grade, graphics and upscaling wait until the scene reads, because a dissolve or a grade can make a weak order look finished.
29.2 Before the first cut
What must exist.
The accepted treatment and the canonical script, word for word. The canonical line is the authority for every word spoken (21.4).
For each deliverable: runtime, aspect, exact copy and brand time. For each shot: who owns its sound.
A probed inventory: for every file its duration, frame size, exact frame rate (including rational rates such as 24000/1001, which a label rounds to 23.976), streams, checksum and accepted intervals.
Locked lines, beds, effects and music cues as separate files.
An edit plan: for every placement its source range, sequence range, track and reason.
The speaking-shot contract decides which audio the cut follows. Each project states at the start whether an exact take must ship or a new performance is acceptable (8.2, 22.2). On the exact-take route the locked take ships and the cut keeps its timing. On the new-performance route a video model performs the line and re-times it, so the clip's own audio is the audio that shipped and the cut follows it: cutting to the timing of the take it was given would put the words out of sync with the lips. Egyptian and Saudi dialogue always comes from a voice actor or ElevenLabs; a video model's native audio is a route for English projects only. Whichever route, the audio that ships owns the timing, and the captions follow it (Chapter 30).
Settings that live outside any instruction to the assistant.
The sequence timebase and aspect, taken from the delivery plan and never from the first clip that arrives.
The track lanes.
A flat folder of media with plain ASCII file names; the Arabic belongs in markers and titles (8.9, 31.2).
The slate on every review export (29.6).
The Premiere build, recorded for each project, because a timeline that behaves on one build is not proved on another (Chapter 32).
The track lanes. A fixed layout makes every project readable at a glance, by you, by the assistant and by the next editor: one purpose for each track, with dialogue, effects and music routed to their own submixes.
Lanes | Carry |
A1–A2 | dialogue: mono, one line per clip |
A3–A4 | Foley and effects |
A5–A6 | beds and room tone |
A7–A8 | music |
One placement, written two ways. The edit plan says "source frames 24 to 144, sequence frames 240 to 360". In the interchange file Premiere imports (Chapter 32), the same placement is in=24 out=144 start=240 end=360. The two pairs do different jobs: in and out select frames from the source; start and end place the clip in the sequence.
At normal speed the source span and the sequence span must be the same length, so compute end as start + (out − in); never type it. When the numbers disagree Premiere trusts start and end, and the source out-point drifts by a frame. A program that computes end from the other three cannot make that mistake; a person typing it can.
29.3 A commercial to cut
The methods below are easiest to learn on one film, worked in frames: a thirty-second spot for a fictional bottled hibiscus drink, set in a Cairo flat on a Sunday. A granddaughter carries a cold bottle up the stairs, opens it, crosses the living room and hands it to her grandmother on the balcony, who tastes it and is, before she speaks, back in her own mother's kitchen. It is arithmetic, not footage: every source range is planned to fit a five-second file of 121 frames, and a real take has whatever interval the probe finds.
Figure 29.1 — The thirty-second spot. S04, the handoff, is the shot the chapter keeps returning to. Dialogue has no lane: its position is taken from the take that ships, not from a plan.
Written for this book · food and drink · a fictional bottled hibiscus drink, thirty seconds, eight shots at 24 fps in a 720-frame sequence · a frame table, an audio map and one plan entry, not a prompt · not run
Example 29.1 — the thirty-second hibiscus spot: frame table, audio map and one placement
Shot | Picture | Sequence | Frames | Source | Handles before / after | What must read |
S01 | Stairwell: she climbs with the cold bottle | 0–72 | 72 | 10–82 | 10 / 39 | the bottle is cold: condensation, a wet palm |
S02 | Flat door: she twists the cap | 72–156 | 84 | 20–104 | 20 / 17 | the seal cracks; the sound has a source |
S03 | Living room: she crosses to the balcony | 156–252 | 96 | 8–104 | 8 / 17 | the geography: door, chair, balcony light |
S04 | Balcony: the handoff | 252–348 | 96 | 14–110 | 14 / 11 | the bottle changes hands once, label toward camera |
S05 | The grandmother tastes | 348–456 | 108 | 6–114 | 6 / 7 | her attention changes after the taste |
S06 | The granddaughter waits | 456–528 | 72 | 12–84 | 12 / 37 | she does not speak; the wait is the shot |
S07 | Both at the rail | 528–612 | 84 | 18–102 | 18 / 19 | the bottle set down on contact; the line lands |
S08 | Brand ending | 612–720 | 108 | 4–112 | 4 / 9 | label and name, held clean |
The audio map. Dialogue: one line, «لسه بنفس الطعم», spoken quietly after the recognition; its position is taken from the take that ships, never from this table. Effects: the seal cracking in S02 sits on the take's observed contact frame; the hand-to-hand Foley in S04 sits on the frame where support changes; the bottle set down on the balcony rail in S07 sits on contact. Bed: one continuous room bed under the whole scene, with a perspective change between the stairwell and the flat. Music: a restrained lift may begin after recognition; compare a version with no score. Brand ending: an audible decay or a deliberate clean stop, never a truncation.
One placement, as the edit plan records it. S04 is used at normal speed from source frame 14 to 110 and placed at sequence frame 252 to 348:
{
"sequence": {"name": "Hibiscus_30s", "fps_num": 24, "fps_den": 1},
"placement": {
"shot": "S04", "source": "S04_accepted.mov", "track": "V1",
"source_in_frame": 14, "source_out_frame_exclusive": 110,
"sequence_start_frame": 252, "sequence_end_frame_exclusive": 348,
"speed": 1,
"reason": "Complete the transfer once, then cut to the grandmother's tasting"
}
}The table gives every row what must read, so a note can point at a row.
Every source range fits inside a file of 121 frames and equals its sequence length, and the handles are written beside it: S04 has 14 frames before its interval and 11 after, and S05 has only 7 after, which limits how far its tail can be extended or dissolved.
Sequence end is start plus the length of the source span (252 + 96 = 348), so the plan and the file cannot disagree.
What changes when S04 fails depends on which obligation the failure touches, and each remedy protects a different one:
A hand mutates in S04's last second, after the bottle has visibly changed hands: trim the tail and start S05 earlier.
The transfer itself fails: a contact insert, or a corrected S04; never an unrelated beauty shot to cover it.
S05's face changes at the moment of recognition: a shorter valid interval, but only if it still carries the turn; otherwise replace the take.
29.4 Logging a take by its usable interval
A take becomes usable the moment it can be addressed. Logging turns a file into intervals, so that a note can say "frames 38 to 61" instead of "the second take". Do it whenever a take arrives, before it is placed.
Inputs. The probed file; its contact sheet (a grid of frames sampled across the clip); the clip at full speed; its audio; the events the shot must contain. The assistant works only from frames it has sampled and audio it has heard or measured; counts, rates and durations come from the probe, never from a label, a file name or a storyboard.
Logging a take
Probe the file; record its frame rate and frame count.
Mark five frames: the first usable frame, the onset of the action, the key contact or reaction, its completion, and the last usable frame.
For each rejected stretch, record the reason at the frame where the defect starts.
Recommend the strongest interval and the neighbouring coverage it needs, keeping events actually observed apart from uncertain ones. The director confirms.
Write the accepted interval into the run record and the edit plan, as in and out frames at the file's own rate.
Figure 29.2 — A log line, drawn. The interval is a decision; the five event frames and the defect frame are observations, each written at the frame where it happens.
Write down why an interval fails, at the frame where the defect starts. "Drift begins after the turn", "the label disappears in the grip", "the listener reacts before the key word", "the end hold contains a hand mutation". Each tells the next step exactly where the usable part ends.
A correct first frame is not a pass if the contact breaks later. On a start-frame route the opening frame is your approved still by design, so it proves nothing about the rest of the take.
A contact sheet locates events; it does not certify contact, performance or sync. Watch the interval at full speed before you accept it.
Count frames from 0, with in included and out excluded, at the file's own rate. Never convert a range from a file at one rate using the sequence's rate, and record rational rates exactly. Files do disagree with their labels (a standard size has come back a few pixels off, one clip at 30 frames a second among clips at 24), so the probe settles it.
A shorter valid interval can repair a take without a new generation, unless it removes a required action. It is the cheapest repair in the production, so it is the first to try.
Template: a log line (written for this book; not run)
Log line: [shot] · [take] · [in, out) at [fps] · onset [frame] · contact or reaction [frame] · completion [frame] · usable: [yes / no] · reason: [defect] from frame [n] · observed / uncertain: [which events]Check the log. Every cited frame exists in the file. Observed and uncertain events are kept apart. Every reject carries a reason tied to a frame. If a storyboard's timecodes have crept in, go back to sampled frames only; if no take contains a usable interval, the shot goes on the missing-coverage list (29.11).
29.5 Curating a batch
Curation is the same job for a batch: many stills or clips for a scene, of which a few will be used. Its danger is different. A folder full of good-looking files feels like progress; it is not a scene.
Keep raw outputs and rejects with their reasons, copy accepted candidates under stable shot and take names, and leave the originals untouched. A stable name (S04_T03, not "final_final2") survives every later step, and the untouched original is your provenance.
A scene is complete only when its selects work together in the rough cut, never when a file count is reached. Three accepted takes of a handoff are worth nothing if none of them joins the shot before it.
Treat a grid panel as a candidate only, and rebuild it at full resolution before it becomes an input. One cell of a 2 × 2 grid is a quarter of the frame; used as a start frame, it hands its softness to every clip made from it.
Cutting around a defect. Crop, mask, blur, shorten or cover, and use only the best part of the clip. A crop costs resolution the finish may need, a cutaway must not hide a broken join (29.8), and a change of speed never touches speech (29.9).
29.6 Assembly A: the first causal cut
Assembly A is the first cut made from real selects, in causal order, before any finishing money is spent. It is never skipped, because this is where takes are judged: a take that looked fine alone may fail next to its neighbours, and one that looked weak may carry the scene. From here on the cut keeps running, and it decides what is generated next (29.11).
Inputs. The obligations list; the interval log; the audio that ships (the locked take, or on the new-performance route the clip's own audio); room tone. Tools. The assistant writes the edit plan; a program builds the timeline file or a draft render; you watch it in Premiere or as a slated review render.
Building Assembly A
Place the intervals in causal order, on hard cuts. The result is the edit plan.
Keep the dialogue that ships at its own timing, linked to its picture.
Lay continuous room tone under the scene, so that silence cannot hide a broken join. Effects and lip sync wait: a first cut carries neither.
Slug missing coverage in the timeline, with its purpose written on the slug.
Render a slated review copy and return the applied cut list.
Watch three times, once for meaning, once for performance and once for continuity, and write notes. The director decides on each.
Cut on hard cuts first: no decorative transitions, graphics, grade or upscale until the scene reads.
Write each note as four things: the observed event, its effect, the smallest repair, and the upstream item it touches. The note then says where the fix lives: in the cut, in a take, or in a new generation.
Ask for the applied cut list (shot, source in and out, sequence in and out, reason), not a description of the intended sequence. The list can be checked against the render; a description cannot.
Slate every export. Work in progress carries "DRAFT — NOT FOR APPROVAL". A cut sent because a decision is wanted carries "ROUGH CUT v3 — FOR EDIT DECISION:" and the question, for example "keep the four-second brand ending, or cut to three". Every export carries its version number (31.2 sets out the layout).
The test cut and the finished cut. A test cut proves a route or a scene cheaply: a hard-cut assembly with room tone, slated as a draft, with no grade, upscale or graphics, one reference render, and the shot table row and run record filled in. It is never presented as the model of a finished version. The finished cut goes on through alternates, joins, rhythm, the verdict and the lock below, and then through Chapter 30.
Check the assembly. The order matches the obligations list; each note names a cut, a missing piece or a defect in a source take; the applied cut list matches the render. If a defect sits inside a take, try a shorter interval before any new generation. The director receives the edit plan, the applied cut list, the slated render, the notes and the missing-coverage list.
29.7 Alternates, and J and L cuts
After Assembly A, notes suggest other constructions: show the listener sooner, hold on the giver through the answer, drop the establishing shot, delay the score. Each is tested as an alternate. Not "a faster version" and not "a slower one": pace is a result, not a proposition.
Duplicate before you test, and change one proposition per alternate. If B differs from A in two ways and plays better, you do not know which change did it, and the next note has nothing to stand on.
Testing an alternate
Duplicate A as B.
Change one proposition in B. The assistant writes B's changed boundaries.
Compare A and B with the score muted, then with the score. Judge the construction on its own first. The director chooses.
Keep both, and record the decision.
Figure 29.3 — An L-cut carries the outgoing sound past its picture; a J-cut starts the incoming sound before its picture. In both, the arrow is on the picture track.
Move the picture boundary, never the audio. An L-cut can put the viewer on the listener without moving the approved line by a single frame. Never offset dialogue to fake an overlap while the speaker is visible: it breaks lip sync. J and L cuts are options, not automatic sophistication.
Build an overlap natively in Premiere until one has been imported and heard through the interchange file you use. Whether overlapping dialogue and music survive that file is not yet known; until one has been tried, build the overlap where you can see it (31.5).
With no usable listener handle, report the gap and request an insert; never loop the listener. Ask for the missing picture instead of repeating the picture you have.
The same instruction can be as short as one line of edit notes. For a J-cut across a clause break: "Edit instruction: use A's picture through the clause break; J-cut B's restrained inhale under the final words; cut to B's reaction before A's audio ends; let B's reply begin over the held reaction; preserve room tone and exact approved dialogue." Every clause names one boundary, and none of them moves a spoken word.
Check the alternate. Lip sync holds whenever the speaker is visible; no listener is looped; the duration is held. The output is both sequences, B's changed boundaries and the director's decision between them.
29.8 Continuity and contact joins
Any cut across an action (a handoff, a pour, a turn, a door) is a join, and a join is judged as a pair, never from one side. With generated footage the two sides were made separately, so nothing guarantees they agree on where the action is at the moment of the cut: audit the state at every cut.
Checking a join
Watch the outgoing and incoming action together, at normal speed.
Name the phase of the action at the out-point and at the in-point: approach, contact, support, release.
Find an out-point and in-point pair that shows the transfer once.
Check the label's orientation, the active hand and the direction of travel.
If no pair exists, keep the better performance and request only the missing insert (29.11).
Figure 29.4 — The join between S03 and S04. When the outgoing shot has already completed the transfer and the incoming shot starts it again, the audience sees the bottle handed over twice. Move the out-point earlier; a dissolve will not help.
Match event phase, screen direction, support and attention across the cut; show one complete transfer, never two. A wide can end just before contact and an insert complete it.
Track the object, not only the actor's body. For an object passing right, its travel carries the cut; the eye follows the bottle, not the elbow.
A dissolve never repairs a teleporting prop. It only shows the prop in two places at once, slowly.
Never mirror a frame to fix a join when text, screen direction or face asymmetry would reverse. A mirrored bottle reads its label backwards, and a mirrored face is a different face.
Check the join from both sides, not on one frame: the object's travel, and the hand that holds it. The output is a new out-point and in-point, or one insert brief on the missing-coverage list.
29.9 Rhythm and sound-led timing
Once the order works, tighten it, and let the performance and the sound set the cut. This is also where a runtime that is too long gets solved.
Tightening a cut
List trim candidates in this order: dead lead-ins, repeated information, tails after an action has completed.
For every hold, name what it does: anticipation, awkwardness, authority, or nothing.
Propose exact removals, with their effect. The director approves.
Listen across every splice, with picture and without.
If the runtime still cannot be met, show the smallest creative compromise.
Trim dead lead-ins, repeated information and completed-action tails before anything the audience needs. Models add holds after the action, often half a second to a second and a half after a spoken line. A hold that creates anticipation earns its frames; a hold that does nothing is the model's habit, not your film.
Let the actor's decision set a cut when it is stronger than the music; a key sound may land before or after a musical accent. Cutting to the grid of a track is a choice, not a law (25.4 teaches cutting to music).
Never speed up speech or squeeze a locked take to hit a length. Give the shot the time, or show the compromise. A squeezed take is a different performance nobody accepted.
Keep the pauses in dialogue trims. Removing every breath makes a performance sound assembled. Room tone bridges an edit; it must not hide a changed meaning.
Listen for Arabic consonants under music, with picture. Where music masks them, thin the competing instruments or duck that passage, and judge intelligibility by an ear that speaks the dialect: a language list never proves a dialect.
Put an effect's transient on the frame of contact, not on the start of the movement. A transient is the sharp front of a sound. A sound designed on the entrance of a card, a door or a bottle lands early; find the frame where the two things touch.
Template: reducing a duration (written for this book; not run)
Reduce this [N]-second cut to [M] seconds. Keep [the protected beats] and [the brand hold]. First list dead lead-ins, repeated information and completed-action tails; propose exact removals with their effect. Keep every approved word and the reaction's thought. If the length cannot be met without touching a protected beat, show the smallest compromise; never speed up speech.Check the result. Measure the runtime on the render. Every protected beat is present; every splice has been heard with and without picture; Arabic is intelligible under the cue. If the actor now sounds hurried, restore the pause and take the time from redundant picture.
29.10 Judging the work in the running cut
Judge a take in the scene, not alone, and generate only the coverage the cut is missing. A running rough cut is the cut that grows as takes are accepted: every accepted take goes into it and is judged there. The cut shows which beat is missing and which take carries the scene, so the next generation fills a gap instead of repeating a shot you already have.
Review the cut before any upscale or finishing spend. Assemble the accepted intervals on hard cuts, with the accepted dialogue and one bed, slate it, and judge it. Finishing starts only when the cut is accepted, and only on the intervals the cut uses.
The edit, not the shot list, decides what is generated next. A shot list says what the film was supposed to need; the running cut shows what it needs. A batch is read for what it can give the sequence: a longer take may yield several usable fragments, or a better angle than the planned one, if story, identity and continuity hold.
When a handoff, a fight or a reveal is too complex for one generation, give each short shot one new event and cover the contact with body-detail inserts. Then cut the short shots, check their joins, and fill only the coverage still missing.
The file, the picture and the sound
Every returned file is checked in the same order: the file, then the picture, then the sound. Each check can make the next one unnecessary.
The file. Probe every returned file before you look at it: duration, frame rate, frame size and audio stream, against what you asked for. A finished job proves nothing about the file. Record the provider's label separately from the measured file, and leave a file's native or upscaled provenance marked unknown until it is known. Label a provider's fault; never rewrite the prompt for it: a prompt cannot fix what the platform does to the file.
The picture. Inspect in this order: identity, product, geography, scale, light, beauty. A beautiful frame proves nothing about motion, and beauty judged first hides a wrong face.
Count the people against the cast in every take. A second copy of a person, a passer-by who arrives whole, or an uncast character at the end of the film fails however good it looks.
Judge identity at 200 % against the signed master, never against the previous output. Comparing with the last frame lets drift accumulate unseen.
After every edit, check what you did not ask to change. A watch changes wrists, the camera comes back slightly higher, a box shifts.
Re-judge the performance after any texture, look or finishing change. A change to a surface can change the acting. After a paint-over, a relight, a frame-rate boost or an upscale, watch the face and body again, not only the surface.
Make a contact sheet (six frames at 0, 20, 40, 60, 80 and 100 %) to compare takes, then judge motion, acting and sync in the clip itself. Inspect the middle of a start-and-end clip, not only its ends, and crop any text to the letter.
Look first where generated images break: food, hands, fabric and the ground.
The sound. On the exact-take route the check is the plate against the take: the onset and end of the speech, the lips closing on ب and م, the shape of ع and ح, nothing moving in the pauses, and the listener's mouth still. On the new-performance route the model's own audio is what ships, so score it word by word: the wording, the dialect, the pronunciation and the performance, then the sync. Never let a number pass a shot. An automatic mouth-movement score has rated a failed clip and a passed clip almost the same. A measurement can flag a take for attention; only a listener can pass it.
The verdict
Pass a shot only with your own eye and ear. "The job completed" or "the face held" is not a pass. Any technical read stays labelled a technical read until the director judges. Where a speech verdict is the director's acceptance and not a native reviewer's certification, record it as that.
Say only what the result covers. A result covers one task, one route, one asset and one duration. A run that passed once is an observation. Before a route becomes a project default, prove it on a small batch per condition (three outputs to pilot it, ten to confirm a protocol) and compare whole batches, not the best take of each: a hand-picked best hides the success rate the next budget needs.
Write the checks before the job runs. Each is a question answered on a frame, a tile or a measurement: "Is frame 0 the keyframe?", "Does the box stop against the laptop, not through it?" Then the result is judged against what the shot needed, not against what came back.
Record every run the day it runs, failures included. The run record (8.7) holds the exact prompt, the settings, the inputs and their hashes, the job number, the debit shown and the debit taken, the defect class, the verdict, and the accepted interval: source in and out at the file's frame rate, first frame included and last excluded. Log the model the platform's record returns, not the one you asked for.
The check sheet of 8.9 is the last gate before a shot's assets lock, and any failed check stops the lock. The returned file is checked against it now, block by block.
The returned shot against the sheet of 8.9
Prompt truth: the file is what the prompt asked, in its order, closing on its endpoint and its sound switch; camera and light gave the results written beside them.
Identity truth: judged against the signed masters, never the previous output; every reference did its one job; a real person shows no drift at all, which is the hard fail.
World truth: the geography agrees with the location masters and every noun belongs to the place and time; in-world text is honest to its distance and the editorial layer is absent from the pixels.
Delivery truth: the frozen systems held and the endpoint landed as a cuttable composition; the watch items are clear frame by frame at the contact points; duration, aspect, size, references and spend are logged; the shot answers its intention line, which no machine checks and which outranks the rest.
The sheet has no check on cost, rights or the accepted interval. Add these before the job, and log the interval and verdict after it.
Before the job: the gates
Spend gate: cost read on the day, and the spend authorised for this batch, with its cap.
Rights gate: rights status of every input recorded.
Sound switch, aspect and resolution set in the settings, not in the prompt.
Your checks for this shot written down.
For a hard choice between two takes, score each on rights and factual correctness first, then identity, product, exact text, dialogue and performance, sync, story, continuity, motion and physics, handles, file, repeat reliability, and cost and repair time. A rights or factual failure rejects a shot whatever its beauty.
29.11 Back to generation
A note from the running cut (the observed event, its effect, the smallest repair and the upstream item) does not always need a new generation. When it might, an edit diagnosis becomes a precise request, and only that is generated. Keep the approved version while the replacement is tested. The order of repair is a trim, an audio bridge, a child shot (a short new shot made to fix one join), and only then a global transform.
Figure 29.5 — The loop. The cheap question comes first, in the cut; a generation is the answer only when no select or trim keeps the obligation, and the approved take stays in the cut until the replacement has won its place.
From a note to a new shot
Write the note: observed event, effect, smallest repair, upstream item.
Try a changed select or a trim. If the obligation holds, stop.
Otherwise write one replacement request, and add it to the missing-coverage list.
Generate only that shot, inside the budget the director approved.
Place it in the running cut beside the approved take, and keep the approved take until the replacement is judged the better.
Try a changed select before a new clip. A different take, or a different interval of the same take, costs nothing. A new clip costs its credits and everything around them: retries, upload and storage, operator time, review and repair, and the finishing the shot then needs. A cheaper generation that needs more repair can be the expensive route.
Keep a missing-coverage list: the shot, what is missing, what stands in for it meanwhile, and what it would cost. It is the production's to-do list, written by the cut, and what the director approves spending against. Approval of a plan is not approval to spend: a batch of paid generations needs its own authorisation, with a cap (34.3).
Template: a missing-coverage entry (written for this book; not run)
[shot] · missing: [the event or state] · stands in meanwhile: [select, still or slug] · replacement brief: [yes / no] · entry state: [outgoing shot's exit] · needed interval: [seconds] plus handles · approval: [budget reference]Write the replacement request from the cut. It has four parts: the entry state (the outgoing shot's exit), the event (one), the order of the performance, and the outgoing handle (the hold the cut needs at the end). State what has already happened, so the model does not perform it again: if a new shot repeats the handoff, the request did not say the handoff had happened. For a continuation, restate the terminal state and describe only the next causal action, camera move and endpoint. Ask for a duration that covers the interval plus its handles, and write the outgoing hold into the request: the file has nothing beyond its last frame. The tool sections give each model's grammar (Chapters 16 to 18).
For a failed transfer, ask for a contact insert, not a beauty shot, and never regenerate both sides of the join. Keep the better performance and request only what is missing. If a hand switches sides between accepted shots, repair in this order: an alternate tail where the object is on the right side, then a shorter shot regenerated from the correct start state.
Written for this book · food and drink · a revision brief to the assistant editor, for shot S05 of the hibiscus spot in Example 29.1 · selection first, a replacement request only if selection fails · not run
Example 29.2 — a revision brief: make recognition causal
In the current thirty-second assembly, the grandmother looks moved before the bottle has reached her lips. Make recognition causal: keep S04 until the bottle has changed hands and she has looked at the label, then choose an S05 interval in which her attention changes after she has tasted. Keep the granddaughter's waiting shot. Hold the total runtime and the four-and-a-half-second brand ending by removing only redundant approach or completion time. If no S05 take contains that order, report the missing performance and draft the replacement request: entry state, one event, the order, the outgoing hold, and what has already happened.The defect is stated as a problem of order, in terms of what the audience sees first, not as "pacing".
The brief tries selection before generation and names the escalation, so nothing is regenerated by default.
It says what must not move: the runtime, the brand ending and the listener's shot.
Handles and length
A handle is a usable frame outside the interval you have chosen: a trim, a slip (moving the interval within the take without changing its length) or a dissolve borrows from it. On a shoot, handles come free: the camera rolled before action and after cut. A generated clip has only what the model made.
No universal handle duration exists; probe every file, and treat a handle as whatever usable interval the file holds. A generated file has nothing before its first frame or after its last. Interior holds are usable handles when your interval ends before them, if they are clean. A select that uses the whole file has no handles, and a range invented on paper is not footage: the hibiscus spot's S05 ends 7 frames before its file does.
Figure 29.6 — A handle check, drawn for two imagined clips of 121 frames. A hard cut needs no frames outside either interval; a dissolve needs usable frames on both sides of the cut; a trim or a slip needs frames beyond the interval.
A handle check
Name the operation at each cut: hard cut, dissolve, trim, slip or overlap (a J- or L-cut).
Compute the frames it needs outside the interval, on each side.
Compare with what the file has beyond the interval.
Check those frames are usable: no drift, no mutation in the end hold. A freeze or a loop is not a handle.
Put any shortage on the missing-coverage list, with a proposal: a different cut, another select, a native treatment in Premiere, or a new source.
Ask for length when you generate, not in the edit: a duration that covers the interval plus its handles, and the outgoing hold in the request (Chapter 14 sets the durations each model allows). If a handle is too short, cut hard, choose another select, or generate longer next time; never pass a freeze off as a handle. Premiere's Generative Extend adds frames to a clip; whether a job may use it, and its limits, are in 32.9. Record the handle status of every cut in the edit plan.
29.12 Diagnosing, the ladders, and when to stop
Most wasted credits come from repairing the wrong layer. Before any repair, ask three questions; they take a minute.
Is it already in the source? Open the keyframe, or listen to the take alone. If the defect is at frame 0 or in the take, fix the still or the line, not the clip. When a clip keeps failing, the still is the problem, and the fix belongs upstream, where a fix costs an image.
Does every attempt do it? The same defect every time points to the shot, the route or a sentence in the prompt; a different defect each time is variance, and another attempt may clear it. Allow exactly one unchanged re-roll for this reason.
Does the route own this behaviour? A mouth moving in a pause, a line re-timed by a re-performance, a lip-sync output shortened to the shorter of its inputs: these are how the route works, and no wording will move them.
Figure 29.7 — The levers of repair, ordered by cost. Climb one rung at a time, and only when the rung below has failed; read the true cost on the day.
Climb the smallest lever first: a different select or trim in the cut, a local fix in post, a one-change edit of the still, a new generation with one variable changed, another route, and finally a new plan for the shot.
The ladder for a failed clip is the seven rungs of 14.16, climbed in order: re-roll unchanged once, because variance is real; reduce the intensity; strengthen the protection on whatever broke; simplify to one motion system; re-phase; split the shot into two clips and a cut; and last, return to the keyframe, because if three steps fail the still is the problem. A failed take is diagnosed, never re-rolled blindly.
For any failed job, first check that the route, the task and the asset roles are compatible; after the last rung, check the stop conditions below.
Re-roll unchanged once, then change exactly one variable per attempt, and log it. Two changes at once hide which one worked.
Fix a frame that is 80 % right with an edit, not a re-roll. A re-roll renegotiates identity; an edit keeps it. One change per edit, the approved parent as the source, the preserve list repeated. Name the change and the effects it may bring: a relight moves shadows and reflections, a garment moves folds and contact shadows, a sign moves lettering. Where a region must stay pixel-identical, composite the approved edit back into the original. When an edit breaks more than it fixes, go back to the last approved parent; a drifting identity is re-anchored on the first-generation masters, never edited in place.
At the last rung, change what the model is given, not the words: a corrected first frame at the state before the action; a separate asset for a state the reference keeps overriding; a simple blocking clip as a video reference. Paint over a still that prompting cannot fix: have a painter correct the signed still, sign it, flag it "painted over" with its parent, protect the painted detail by name in the motion prompt, and re-judge the performance. Repair one missing event with a second generation and a composite in post rather than regenerating the shot (14.13).
Plan two or three named levers in the shot record before the first attempt. The second attempt should be a lever you chose in advance, not a hopeful re-roll.
Stop on conditions, not counts. A number of attempts says nothing about why the last one failed.
When an attempt fails the way the last one did, the diagnosis was wrong: change the diagnosis, not the wording.
Do not write a further wording for a behaviour the route owns, and do not lengthen the negatives when an artefact persists: if it survives one targeted exclusion, it is not a wording problem.
When exact lettering fails after the model was given the exact string, compose it in post. Editorial text is never generated in the frame, and lettering on an object is judged letter by letter (Chapters 10 and 30).
Change the shot when the ladder is used up or the remaining levers cost more than the shot is worth. Run the fallback you planned; do not rephrase. Count the cost with retries, review and repair time, and compare it with the fallback's.
Stop the affected condition when the route cannot be confirmed; a rights or privacy question appears; a schema or interface change invalidates the protocol; the spend cap is reached; three pilot outputs in a row are unusable; or a confound cannot be separated. Preserve the failure and restore the earlier rule. A failed candidate never replaces a working default, and a route never changes silently: a job that returns a different model from the one you asked for is the plain case.
Turn a fix you needed twice into a rule. A failure fixed once is production; fixed twice, it is a pattern, and patterns are written down with the date, the evidence and the correction, so that the next person does not learn them again.
29.13 Which tool for which job, and the records this stage writes
Job | Tool | Where |
Cut, log, conform, lock; build a handoff | Premiere Pro, with the assistant | |
A transcript-led rough cut of presenter footage | Premiere's Paper Edit (no Arabic); Arabic needs a transcript from elsewhere | 32.2, 32.9 |
Missing coverage: a replacement clip | Kling, Seedance 2.5, Wan 3.0 | Chapters 16, 17, 18 |
Fit a mouth to an exact take | Sync Lipsync 3 | |
A replacement line, or a new voice take | ElevenLabs | |
Job numbers, the ledger, the run record | Higgsfield | |
Money, rights and terms of a replacement | the tool's terms and the client contract |
The records this stage writes. The interval log; the edit plan and its applied cut list; the notes, each in four parts; the missing-coverage list; the director's dated decision on each rough cut, against the question its slate asked; the accepted interval and verdict added to the run record; and, at the lock, the list of transforms.
29.14 Locking the picture
Picture lock is the gate between editing and finishing. Until the picture is locked, every finishing step risks being made twice; after it, nothing changes without a new version.
Lock only when nothing is left unresolved, and keep the way back. Unresolved items prevent final lock. Every item on the timeline links to an immutable source; temporary text and temporary sound are gone; and a rollback to the previous lock exists.
List every transform. A speed change, a reframe, a stabilisation, an interpolation and any compositor effect changes the pixels the generator gave you. Each is logged with its frame range, so the next person knows which frames are generated, which are transformed and which are both. The audio status of each clip (native sound, approved take, bed) stays visible.
Interpolation is a versioned child, never a silent fix. Invented frames are a new file for that range; the original is kept; the child is inspected for new pixels with meaning in them (hands, text, product) and rejected if it introduces any.
A good edit does not make a generator reliable. Editorial transformation never retroactively proves a route: one that needed three rescues in the cut is still a route that needed three rescues, and its record says so.
Before a cut goes to review
Every interval logged in frames at the file's own rate, defect frames noted; the order matches the obligations list.
Hard cuts only; no grade, graphics or upscale yet; the dialogue that ships at its own timing; one room bed; missing coverage slugged.
Every join shows one transfer, nothing mirrored, no dialogue moved to fake an overlap, no looped listener.
Runtime measured on the render; the applied cut list matched against it; the slate carries its version and, if a decision is wanted, the question.
Before picture lock
Every timeline item links to its source; temporary text and sound are gone.
Every speed change, reframe, stabilisation and interpolation is listed, with its range.
A rollback to the previous lock exists.
The director has accepted the cut.
Open questions
Overlaps through the interchange file. Whether a J or L cut survives the interchange file is not yet known; one imported passage of overlapping dialogue and music would settle it. Until then, build overlaps in Premiere.
The 121-frame count. Whether five-second requests keep returning 121 frames at other durations and on other models is not yet known; probe every file.
What to remember
The assistant logs, proposes and does the arithmetic; the director decides the cut; Premiere confirms. A check that passes is not a verdict.
Keep three numbers apart: the seconds asked, the frames in the file, the interval used. Probe every file.
Log intervals, not files, with every defect at the frame where it starts. A shorter valid interval is the cheapest repair.
Assembly A: causal order, hard cuts, the audio that ships at its own timing, one room bed, slugs for what is missing, a slate with its version.
One proposition per alternate; move the picture boundary, never the dialogue; show every transfer once; never mirror a frame.
Trim lead-ins, repeats and finished tails first; never speed up speech; put a transient on the frame of contact.
Check the file, then the picture, then the sound. Only an eye and an ear pass a shot; the shot's intention outranks the rest.
A changed select first; one replacement request only when that fails: entry state, one event, order, outgoing hold, what has happened.
Diagnose before you spend, climb one rung at a time, stop on conditions, and lock only when nothing is open, with every transform listed and a way back.
Comments