6.7 Editing and Final Delivery
The clips exist. Assembly turns them into something publishable, and the difference between a generated sequence and a piece of content is almost entirely editorial.
This chapter is the last mile: the tool choice by scale, the edit skeleton that works for a speaking persona, the two cheap multipliers most operators skip, and export per platform.
Choosing an editor by scale
| Scale | Tool class | Why |
|---|---|---|
| A few posts a week, phone-first | Mobile editor (CapCut-class) | Fastest path; captions and music built in; vertical by default |
| Daily posting, multiple accounts | Desktop editor (Resolve-class, free tier) | Keyboard-driven, templates, batch export, no per-export cost |
| Client work, multiple formats | Professional NLE (Resolve, Premiere) | Project files, versioning, delivery specs a client can accept |
| Automated, template-driven | Scripted pipeline (7.3) | When the edit is the same every time and volume justifies code |
The rule: choose the tool that removes the most manual steps at your current volume, and change it only when a specific step becomes the bottleneck. An operator producing four posts a week does not need a professional NLE, and an operator producing twenty does not need to drag clips on a phone.
The one capability worth checking before you commit: does the tool support templates with replaceable media? A saved project where you swap the clips and the captions and export is the whole efficiency argument for the desktop tier. Without templates, you are re-deciding the same layout every time.
The edit skeleton
The structure that works for short-form, in order:
0–2 seconds: the hook. The single most important element. It must be visually legible at a glance and it must be a shot you generated deliberately (6.4), not the first A CLIP-style text encoder that responds best to comma-separated fragments; used by older model families. you happened to have. In practice: a striking frame, a movement, a question in text, or a face doing something specific. Not a logo, not a slow fade, not an establishing shot.
2–5 seconds: the promise. One line of context, spoken or on screen. Why keep watching.
Body: a cut every 1.5–3 seconds. Fast enough to hold attention, slow enough to be readable. The cuts come from the material: five-second clips cut into pieces, stills with motion, cutaways.
Regular pattern interrupts. A text card, a change of location, a sound effect, a different framing. Every five to eight seconds. This is the mechanism that keeps a scrolling viewer.
Final 2–3 seconds: the close. A line, a question, or a call to action. Not a hard sell unless the piece is promotional.
On pace, and where it comes from. Fast cutting is the default in short-form, and it also solves a generation problem: since The slow migration of a face away from the reference plate, invisible in one frame and obvious across a year. concentrates in longer continuous shots (6.1), cutting frequently hides it. A fast edit is both a retention technique and a consistency technique, which is why it is worth adopting even if you prefer a slower aesthetic.
Captions: generate or burn in
Captions are effectively mandatory. Most viewing happens muted, and captions are read where a spoken line is skipped.
Two approaches:
Burned into the video. Permanent, guaranteed to display identically everywhere, and it survives re-uploads and screen recordings. Costs nothing at view time, and it cannot be turned off. The right default for short-form.
Platform captions or a sidecar file. Editable, translatable, accessible to screen readers, and platform-native. The right choice when you need a transcript, and required in some accessibility contexts.
The practical combination: burn in captions for short-form, and keep a separate transcript file. The transcript costs an extra export step and gives you three things: an accessible version, a translatable asset for the seven languages this book's own subject will ship in, and a searchable record of what you published.
Two caption rules from experience. Keep them in the safe zone — platform UI covers the bottom of vertical video, and captions placed there are the most common amateur error. And keep the line lengths short: three to five words per line, not a full sentence.
Sound design: the cheapest realism multiplier
The single highest-return editing investment for synthetic video, and the one most often skipped.
Generated clips are silent, and silence is unnatural in a way viewers register immediately. Layering audio onto a generated clip does more for the impression of reality than any visual improvement at a comparable cost.
The layers, in order of impact:
- Room tone or ambience. A continuous quiet bed: a kitchen hum, street distance, room reverb. This one layer removes most of the "studio void" that makes generated video feel synthetic.
- Foley. The small sounds of the action: a mug set down, fabric movement, footsteps. Wrong or missing foley is noticeable; good foley is invisible.
- Music bed. Low, unobtrusive, matched to the pace of the edit. Music carries emotion that a five-second clip cannot establish on its own.
- The voice. On top of everything, with the other layers ducked under it.
Where the assets come from: royalty-free libraries, or your own recordings. A single session recording a dozen household sounds — a kettle, a keyboard, a chair, footsteps on two floors — builds a usable foley library for a year.
One warning about music licensing. Commercial use of music requires a license that covers it, and platform-provided music libraries are typically licensed for use on that platform only. A track that is fine on one surface can be a copyright claim when reused in a paid ad. Read the terms, and keep a record of which tracks you used where.
Export per platform
[AS OF 2026-10] Specifications change constantly. Verify before a campaign; the table is directional.
| Target | Aspect | Practical notes |
|---|---|---|
| Instagram Reels | 9:16 | Safe zones top and bottom; cover frame chosen deliberately |
| TikTok | 9:16 | Same safe zones; the first frame matters most |
| YouTube Shorts | 9:16 | Slightly longer tolerance; captions often read |
| Feed video | 4:5 or 1:1 | If a vertical piece must also work in feed, export separately |
| Paid ads | Per placement | Often needs a cleaner version without organic captions |
Bitrate and codec: export at platform-recommended bitrate or above, H.264 or the platform's preferred codec. Over-compressing damages motion smoothness, which is the first thing a viewer perceives as low quality.
One export per platform, not one file reused everywhere. The cost is minutes; the benefit is that the safe zones, the cover frame, and the caption placement are correct on each surface.
The cover frame. Platforms let you choose it. Choose it deliberately — a frame where the persona's face is visible, legible at thumbnail size, and not mid-blink. Chapter 6.3's frame-stepping procedure is how you find it; the middle of the clip is usually a better cover than the first frame.
Video upscaling: when it is worth it
Video upscalers can raise resolution and add detail to generated clips.
Worth it when: the clip will be shown large (a hero piece, an ad, a website embed), or when the source model's native output is below the platform's recommended resolution.
Not worth it when: the clip will be viewed at phone scale in a vertical feed. The perceived difference at that size is small and the processing time is significant.
And never as a fix for identity drift. An upscaler enlarges what is there. If the face drifted, an upscale gives you a sharper drifted face.
Delivery checklist
Before publishing anything, in order:
- Hook legible in the first second, frame-stepped
- Captions in the safe zone, short lines, checked against the transcript
- Audio layers present: ambience, foley where relevant, music, voice
- No clipped audio; voice intelligible on a phone speaker
- Correct aspect and bitrate for the target platform
- Cover frame chosen deliberately
- Identity frame-stepped at the start, middle, and end (6.3)
- Disclosure present where required (11.3) — the step most often skipped under deadline
- Provenance record written: The validated still a clip is generated from, carrying the identity the video model would otherwise sample., The text describing movement alone, kept short when identity conditioning is heavy., voice take, edit version
Item eight is the one that creates legal exposure, and item nine is the one you will want in three months.
Where this goes wrong
No hook. A piece that opens with an establishing shot will not be watched.
Slow cutting. Costs retention, and leaves identity drift visible.
No captions. Most viewing is muted.
Captions in the platform UI's zone. Covered by the interface.
Silent clips. The most common reason generated video reads as synthetic.
Music without a commercial license. A copyright claim waiting for a paid campaign.
One export reused on every platform. Wrong safe zones, wrong cover frame.
Raising resolution on pixels rather than latents, so it runs after composition, identity, and repair. to fix drift. Sharpens the problem.
Skipping the disclosure check under deadline. The legal exposure is not worth the minutes saved.
The gate
- Does your edit open with a deliberately generated hook inside two seconds?
- Is there a pattern interrupt every five to eight seconds?
- Are captions burned in, in the safe zone, with a separate transcript kept?
- Are ambience, foley, music, and voice all present?
- Do you export per platform with a chosen cover frame?
- Is the disclosure check a fixed item in your delivery list?
What you leave with
An editor choice keyed to volume with the one capability that matters, a six-part edit skeleton where fast cutting doubles as a consistency technique, the caption decision with the transcript as a by-product, four audio layers that make generated video believable, per-platform export with the cover frame as a deliberate choice, a rule for when video upscaling is worth it, and a nine-item delivery checklist that ends with the two items operators skip. Part VI is complete: motion, voice, sync, and delivery.