Chapter 72. Edit and Finish: From Usable Shots to a Deliverable Narrative#

Editing is not arranging eighteen approved shots in numerical order. Generated shots typically carry excess wind-up, uneven action speed, eyelines half a beat late and short usable ranges. The editor's job is to rebuild attention, causality, breath and audio continuity while protecting the facts locked upstream.

72.1 Narrative assembly, the sound-picture interface, and tightening#

Assemble for story first; do not package.

V0 uses only reference audio from the shots and scratch dialogue, with grading, transition effects and complex music forbidden. The editor checks four things: whether the shift of power is comprehensible with the sound off; whether the conflict is comprehensible without picture; whether every shot has an irreplaceable job; and whether Gu Zhou's informational advantage is established before his closing line.

V0 ran 81.6 seconds. The problem was not that the whole piece was slow, but three specific redundancies: Lin Xia holding for 1.1 seconds after security finishes speaking; the directors' toast and Lin Wei's smile carrying the same victory information; and two similar reactions before the folder appears. Removing them produced 75.2 seconds.

J-cuts, L-cuts and reaction shots.

Lin Wei's challenge begins four frames before she is on screen, laid over Lin Xia's entrance so the language blocks her first. After Lin Xia says the answer is inside, the picture cuts early to the folder while her line's tail continues, moving the audience from words to evidence.

When Gu Zhou opens the file, paper is heard before the first page is seen; sound establishes the action so the graphic shot can be shorter. Reaction shots appear only where a judgment changes — not distributed evenly as a nod or a stare after every line.

Four passes of tightening the rhythm.

The first section, deleting the nameplate, cuts fast without a musical accent, so the audience asks why for themselves. The second slows the boardroom celebration by 0.6 seconds to establish the order about to be interrupted. The third accelerates shot exchange through Lin Xia's entrance until she sits and everything abruptly steadies. The fourth reduces cutting during the handoff, letting the physical movement complete. Twelve frames of black after the closing line let the question land.

Microdrama being fast does not mean every shot is equally short. Evidence of action needs time to complete; the pause after the information is understood is the redundancy worth deleting.

72.2 Subtitles, graphics, and transitions driven by causality#

Subtitles are not an automatic transcript.

Subtitles break by semantic unit, one line preferred and two at most. Key verbs and numbers never split across screens. They never sit over characters' eyes or the contract's title. Phone safe areas and platform button zones are reserved in advance. One long Lin Wei line splits into two subtitles without breaking the phrase naming the acquirer's representative.

subtitle_style:
  font_asset: licensed_sans_v02
  max_lines: 2
  max_chars_per_line: 15
  safe_bottom_percent: 18
  speaker_color_policy: none
  emphasis: weight_only
  punctuation: conversational_standard
  burn_in_master: false

Automatic captioning mis-transcribed the phrase confirming the offer's validity. Because that phrase belongs to the commercial fact dictionary, a QC rule blocked it outright rather than relying on a reviewer noticing by chance.

Phone and contract graphics.

The phone screen tracks a blank panel first, then composites the delivered-offer state, held for 1.25 seconds. The contract's first page tracks the paper's four corners, with motion blur and depth of field matched to the source shot. The body text need not all be readable; the title, the amount and the signature state must be accurate.

A failed version rendered the amount as large glowing type. It was easy to read and looked like advertising packaging, which destroyed the credibility of an object inside the story. The final version raises local contrast only, and lets Gu Zhou's line confirming validity carry the meaning.

Transitions follow causality, not plugins.

The episode uses hard cuts, matched action and sound bridges. Cutting from the nameplate being scraped off to the phone vibrating contrasts identity lost with identity confirmed. Cutting from the toast to the door handle turning interrupts the order. Cutting from the folder landing to Gu Zhou's eyes turns physical evidence into a change of knowledge.

There are no flash frames, spins or random light leaks. If removing a transition leaves the narrative unaffected, it was usually covering a shot interface that was never designed.

72.3 Visual unification, versioned export, and an edit diagnosis#

Grading and visual unity.

Technical normalization comes first: exposure, white balance, skin tone and black level. Then scene unification: a cooler corridor, a neutral-to-warm interior, blue hour outside the windows. Then narrative emphasis: the red folder holds its dark wine tone and never jumps to bright red. Grading cannot repair face-shape drift, and it cannot turn a wrong light direction into plausible space.

Lin Xia's skin in shot 015 read more magenta than the adjacent shots. The team corrected the skin locally rather than adding a stronger stylistic LUT across the whole episode. Unification outranks making each shot individually more cinematic.

Versions, picture lock and export.

Edit versions run cut_v01_story, cut_v02_pace, cut_v03_sound and cut_v04_client_notes. Each round addresses one primary kind of change and stores a change list. After picture lock, any duration change is assessed against music, bar positions, subtitles, lip sync, graphic tracking and delivery subtitle timecode.

Delivery exports include a high-bitrate master, the platform vertical file, a textless clean version, subtitle files, audio stems, thumbnails and a checksum manifest. Re-import after export and inspect; the export program reporting success is not acceptance.

SOP and deliverables.

Complete an unpackaged V0. Do a muted read and an audio-only listen separately. Trim by causality. Build J-cuts, L-cuts and sound bridges. Produce subtitles and the graphic fact layer. Normalize technically before unifying style. Execute picture lock. Generate deliverables from the locked timeline. Re-import for technical checks.

Deliverables: edit_project/, change_lists/, subtitle_master.srt, graphics_sources/, color_reference_stills/, picture_lock.json, export_presets.yaml, delivery_checksums.txt.

A three-pass edit diagnosis exercise.

On the first pass, play muted and write down, at every pause, who the audience is watching, what they know and what they are waiting for. If the answer does not change for ten seconds, check whether shots are repeating. On the second pass, turn off the picture and listen: is the space continuous, is it clear who is being addressed, is the music explaining ahead of the picture. On the third pass, watch at double speed for structure only, checking that each causal exit meets the next entrance. That pass cannot judge performance detail and does expose long set-ups.

Then duplicate the timeline and build a shortest-comprehensible version, cutting boldly to the edge of comprehension, and add back the breathing you genuinely need one item at a time. That process separates information from habit far more reliably than shortening a long version gradually. Every restoration needs a reason: an action completing, an emotion landing, a beat waiting, a fact becoming readable, or a sound connecting. A pause with no reason does not earn an exemption because the shot was expensive.

A note on sources#

This chapter reconstructs an edit pass to show the order — causality, then rhythm, then packaging — and the checks that keep locked facts intact while the cut changes.