Everyone has seen the demo: a typed sentence becomes a five-second clip of a neon city. It looks like magic. Then you try to make a sixty-second product video with three shots that connect, one character who stays the same, and a voiceover that lands on the beat, and the magic collapses into forty browser tabs.
The gap is not model quality. It is pipeline. After shipping a few hundred short vertical videos, the workflow that reliably produces publishable output looks less like prompting and more like production management. Here is the version we use, step by step, including the places where it still breaks.
1. Compress the script before you touch a model
A sixty-second video holds roughly 140 to 160 spoken words. A blog post, a landing page, or a founder's brain dump is usually ten times that. The first job is subtraction, not generation.
Write the narration as plain text, read it aloud with a stopwatch, and cut until it fits. Every sentence you keep should either set up a visual or pay off one. If a line needs a diagram to make sense, it belongs in a blog post, not a short.
2. Build a shot list, not a prompt list
Before opening any generator, write a table with four columns: shot number, what the viewer sees, what the narrator says, and how long the shot lasts. Three to five seconds per shot is the practical ceiling; longer shots need camera movement or they read as a frozen image.
This table is the contract. It tells you how many generations you need, which shots can be reused from a previous video, and where a screen recording will beat a generative clip. Most bad AI videos are not badly generated. They are badly planned.
3. Route each shot to the right model
No single model wins at everything. In practice, we route by shot type:
- Photoreal people and product close-ups: a model with strong image-to-video conditioning, driven by a clean reference frame.
- Stylised or impossible scenes: a text-to-video model with more motion latitude.
- Text, UI, and screen content: don't generate it. Record it. Generated interface text is still unreliable at reading size.
- Logos and packaging: composite them in post. A model that redraws your logo will redraw it wrong.
The economics matter here. If a model costs roughly three dollars per five-second 1080p clip, a forty-shot video can burn a hundred dollars in bad takes before you have a single usable cut. Image-to-video with a good first frame cuts the retry rate dramatically, because you are no longer asking the model to invent composition and motion at the same time.
4. Lock identity with reference frames
Character consistency is the most common failure. The fix is boring: generate or photograph one strong reference image per character, then drive every shot from that image plus a tight motion instruction. Keep the same seed family where the model exposes one, and keep wardrobe and lighting identical in the reference.
Accept that wide shots hide inconsistencies that close-ups expose. If a shot does not need a face, frame it so the face is not the subject.
5. Generate audio before you edit picture
It is tempting to cut picture first and drop music in later. That produces videos where the visuals and the voice fight each other. Generate or record the voiceover first, then cut picture to the audio waveform.
Two practical rules: keep sentence-level clips separate so you can re-time individual lines, and leave 200 to 300 milliseconds of room tone between sentences. That gap is what makes a synthetic read sound like a person who is thinking.
6. Burn captions, but design them
Most short-form viewing happens muted. Captions are not accessibility garnish; they are the primary read. Two lines maximum, six to eight words per line, high contrast, and positioned above the platform's UI chrome rather than under it.
Auto-transcription is good enough for a first pass and wrong often enough that you must read every line. Product names, numbers, and acronyms are where it fails.
7. Run the QA gates that catch the embarrassing stuff
Before export, watch the cut three times with three different questions in mind:
- Continuity: does the light, wardrobe, and set direction match across cuts?
- Anatomy and text: count fingers, read every sign, check every logo.
- Claims: does the narration promise something the footage does not show?
The third pass matters more than the first two. AI video makes it cheap to illustrate a claim you cannot support, and that is a legal and reputational problem, not a craft problem.
8. Ship variants, not one master
A single export is a wasted production. From the same timeline, produce:
- A 9:16 master for Reels, TikTok, and Shorts.
- A 1:1 or 4:5 cut for feed placements.
- Two hooks: the first 1.5 seconds decide most of the retention, so test a question hook against a result hook.
- A silent version with larger captions for in-feed autoplay.
This is where a structured pipeline pays for itself. If the shot list, voiceover, and captions live in one project, variants are exports. If they live in five tools and a chat log, variants are a rebuild.
9. Measure the first three seconds, then the rest
Platform analytics will tell you where viewers leave. In our experience the drop is concentrated in the first two to three seconds and again at the first hard cut. Fix the hook before you fix the colour grade.
Keep a simple log: hook type, length, model used per shot, and retention at three seconds. After twenty videos you will know which combinations work for your audience, and the pipeline stops being a gamble.
The pipeline above is the one we run at Pavo AI, a script-to-vertical-video tool: the shot list, references, voiceover and captions live in one project, so a reshoot is a row change rather than a rebuild. Even if you build the same flow out of separate tools, the order of operations is what saves the budget.
Where this still breaks
Two honest limits. First, dialogue-heavy scenes with multiple characters remain expensive, because every cut needs its own reference discipline. Second, anything requiring precise physical interaction, like hands assembling a product, still fails often enough that live action or 3D is cheaper.
The good news is that the models are improving faster than the process around them. The teams getting consistent output are not the ones with the best prompt; they are the ones who stopped treating generation as a slot machine and started treating it as a production line with a shot list, a budget, and a QA gate.
If you are building this workflow, the tooling matters less than the order of operations. Script, shot list, routing, references, audio, captions, QA, variants, measurement. Skip a step and you will feel it in the edit.
Top comments (0)