Chapter 47. Assembly: From Generated Clips to a Watchable Scene#

47.1 Organize first; do not drag straight onto the timeline#

Every generated clip carries a shot ID, take, usable interval, state and QC result. The editor builds selects first: best performance, best action, repairable versions and insurance shots. Rejected material never enters the main pool.

47.2 Sync and naming#

Dialogue, narration, Foley, music and video sync to one time base. The final dialogue audio is the source for lip sync and subtitles. Proxies and masters share asset IDs so nothing goes offline.

47.3 The rough cut protects causality first#

The first pass assembles by beat and shot ledger, checking only whether the audience can follow the action, the evidence and the shift in power. Do not chase speed on every beat yet. If causality is missing a shot, return to shot design or use approved coverage — do not patch it with narration.

47.4 Master and coverage#

The master holds the geography; coverage emphasizes lines, hands, proof and reactions. When a generation flaw appears, you can cut to coverage inside the action — without teleporting a character or moving a prop between hands.

47.5 Cut points#

Cutting on action preserves momentum. Cutting on an eyeline shows the audience the object. Cutting to a reaction completes an emotion. An insert supplies fact. J- and L-cuts maintain audio continuity. Cut points are decided by action phase, sense and breath together.

47.6 Eliding dead process#

Delete walking to the door, sitting down completely, waiting for a computer to load — anything with no dramatic information. Keep the intent and the result of an action. When the process contains a choice, an obstacle or evidence, it cannot be crudely elided.

47.7 The limits of fixing things in the edit#

A brief hand error can be covered by an insert. A face changing at the end can be cut before it. A background jump can be covered by a reaction. Core identity errors, wrong text and a fully reversed axis cannot be concealed indefinitely by fast cutting.

Log every problem a repair covers, so the material does not re-expose it when reused for advertising later.

47.8 The edit timeline#

clip:
  timeline_in: 00:25.400
  timeline_out: 00:28.100
  source: GEN_E001_S09_004
  source_range: 1.1-3.8
  purpose: authorization proof insert
  audio: SFX_PAPER_SLIDE_02
  continuity_note: document at table center, right hand leaving
  hidden_defect: background text drifts on the last frame; cut before it

47.9 Edit versions#

Keep assembly, rough, fine, picture lock and platform versions. Record the reason for each significant change. Changing duration after picture lock affects sound, subtitles, graphics and cues, and must be propagated.

47.10 Building a cuttable material pool#

The editor does not receive one final video per shot. They receive a set of usable segments with boundaries. Each clip carries six markings before entering the project: usable in and out points, the best performance region, the action phase, entry and exit states, visible defects, and which shot functions it may carry. Without those, the editor repeats the selection the generation station already did, and may use a face change or prop error hidden outside the usable range.

One shot ID can have candidates with different jobs. Take A has the most accurate identity and suits holding. Take B has a complete hand action and suits a short cut. Take C has the best performance with background drift at the end, so it must be cut away before the reaction finishes. Selects are not crowning a winner; they state what problem each candidate solves.

Proxies, audio and high-quality masters share stable IDs and timecode. Relinking must never depend on folder position or the word "final." Keep source hashes inside the project, so a re-download or re-transcode from a generation platform cannot substitute different content under the same name.

47.11 The four-pass causal assembly#

The first pass lays events only: who did what, and what it caused. Do not add decorative reactions, musical hits or complex transitions yet. If the audience cannot retell the causality in one sentence, return to script, shot design or a pickup route.

The second pass lays knowledge: when each character saw, heard and understood what. The audience having seen a document does not mean the antagonist has read it; the protagonist suspecting something does not mean her ally knows why. Reaction shots and eyeline cuts update knowledge states that are not synchronized.

The third pass lays power: who controls the floor, the space, the evidence and the pauses. The same material can be cut into different power relationships. Cutting to Zhou Lan's answer immediately after Lin Yun finishes leaves Zhou Lan controlling the rhythm. Holding half a second of silence, cutting to a bystander, then returning to Zhou Lan says the evidence has already changed the room.

Only the fourth pass adds perceptual reinforcement: rhythmic fine-tuning, sound bridges, music, graphics and visual transitions. The four passes stop a team from using more effects to manufacture the impression of completeness before the causality holds.

47.12 The continuity triangle at every cut#

Check three lines at every cut: picture state, sound state and audience knowledge. If the document moves from the right hand to the table in picture, sound needs contact or a plausible elision, and the audience must know the action completed. A break in any one of the three produces an unintended jump.

Action continuity does not require pixel identity across two shots. It requires that direction, speed and phase can be completed by the brain. Cutting as the hand begins to descend and resuming as the document contacts the table elides the middle successfully. If the outgoing shot has not yet raised the hand while the incoming shot already has the document down, the jump needs a sound, eyeline or time-elision signal to explain it.

47.13 Off-screen space and reaction coverage#

AI production over-relies on whoever is currently speaking. Genuinely cuttable dialogue needs listeners, bystanders, hands, evidence and spatial coverage. A reaction shot is not there to cover every error; it shows what a line caused. A reaction with no consequence is rhythm filler.

Establish each coverage shot's purpose at shot design: proof_readability, listener_hit, spatial_reset, defect_cover. A phone insert that exists only to hide a face change, while showing the wrong time or message state, fixes a technical problem and creates a narrative one.

47.14 Assembling the boardroom scene#

The first version cut in dialogue order: Lin Wei announces the dismissal, Lin Yun responds, the authorization appears, Zhou Lan questions it, counsel confirms it. The information was complete and the reversal of power was flat. The re-cut lets Lin Wei's announcement land on Lin Yun's still reaction. Lin Yun does not answer immediately; she glances at the door. The room follows her eyeline and turns, and Shen Yan enters the off-screen space. Then the document lands, the title insert plays, Zhou Lan's fingers stop tapping — and only then does Lin Yun say the acquirer's representative will attend.

No lines were added and no shots were shortened. By delaying the answer, redirecting shared attention and showing other people's reactions first, the edit turned "she produced a piece of paper" into "the whole room is forced to accept a new source of power."

47.15 SOP for assembly#

First, organize selects and usable intervals. Second, sync the final audio. Third, assemble by causality. Fourth, establish space with the master. Fifth, add coverage and inserts. Sixth, handle action, eyeline and sound cut points. Seventh, elide dead process. Eighth, log editorial covering. Ninth, complete picture lock and notify downstream.

47.16 Fault tree#

Symptom: it cuts fast and cannot be followed. Coverage has replaced establishing and consequence. Restore the causal shots.

Symptom: the action jumps. The cut is not on a shared phase. Adjust the source range or redo the handoff.

Symptom: space is worse after the repair. The covering shot does not belong to the current axis. Return to the spatial ledger.

Symptom: post keeps reworking. Picture lock came too early, or versions are unclear. Establish edit states and dependency notifications.

47.17 Checklist, exercises and deliverables#

Check that clips have IDs and usable intervals; that the rough cut protects causality first; that the master establishes space; that coverage preserves state; that editorial repairs are logged; and that picture lock triggers downstream updates.

Exercise one: cut a clearest version and a fastest version from the same material. Exercise two: find five processes that can be safely elided. Exercise three: write a timeline record for one early cut that removes an error.

Deliverables for this chapter: selects, the assembly, rough and fine cuts, the edit timeline, the hidden defect log, and picture lock.

47.18 Missing-shot diagnosis: do not use the edit to hide upstream holes#

When the rough cut stalls, classify first: is the gap a missing fact shot, a missing spatial shot, a missing action phase, a missing reaction, a missing time marker, or insufficient material quality? Different gaps need different fills. Using one reaction face to cover everything makes a character look permanently blank; explaining a missing action in narration turns a visual scene into radio drama.

Raise a missing shot request stating what the audience currently cannot understand, the minimum information required, acceptable shot forms, the states either side, and a cost ceiling. A pickup is not a re-shoot of the whole original design: a hand insert, an empty plate of the doorway, or off-screen sound may complete the causality — provided it demonstrably fills the gap.

missing_shot_request:
  gap: viewers cannot tell why security suddenly steps aside
  required_information: security stands down after Zhou Lan confirms the authorization
  minimum_solution: the guard withdraws his arm; "let her in" heard in the background
  state_in: arm_blocks_door
  state_out: passage_open
  alternatives: [insert_reaction, reshoot_medium, composite_arm]

47.19 The timeline must retain provenance and intent#

Every video clip on the timeline carries a shot ID, version, source hash, the range used, speed changes, crops, stabilization, compositing and approval state. Audio clips likewise record the dialogue take, cue, Foley and rights provenance. A timeline is not a black box containing the final truth; it is a traceable assembly list.

Fix the track layout by function: master picture, coverage, inserts, temporary repairs, subtitle reference, dialogue, narration, music, ambience, Foley and technical markers. Mark temporary material conspicuously and scan automatically before exporting a release candidate. Editors are free to create, and provenance must never be lost.

Record intent alongside decisions — "cut to the reaction eight frames early so the audience sees the fear first." When someone later extends the previous shot, they can tell whether those eight frames were incidental or a narrative choice. Without intent records, fine cuts are frequently reset to the material's default length by accident.

47.20 Picture lock is a controlled state, not a verbal announcement#

Before locking, complete six review passes — causality, performance, rhythm, continuity, evidence legibility and cut points — and freeze shot order and durations. Locking produces a version hash, EDL/XML, a reference video, frame rate and start timecode, and sound, subtitles, grading and graphics all reference that version.

Changes after lock are still possible through a change request stating the reason, the affected shots, dialogue, subtitles, music, graphics, grading and delivery date. Replacing even one shot urgently can change sound bridges and subtitle timing; you cannot drag it on the master timeline and inform downstream afterwards.

When a released version is modified, create a new locked version rather than overwriting the old one. Only after each department reports completion and regression evidence does the version re-enter release candidacy. The value of picture lock is not prohibiting creativity — it is making change visible, costable and verifiable.

A note on sources#

AI video is raw material. Reliable editing uses coverage and sound to connect fragments while respecting the state and space already established upstream.