A practical workflow for producing 90-second videos with AI: planning, prompt design, shot continuity, sound, and quality control that keeps output cinematic.
A 90-second video is not a shortened long video. It is a compressed argument: one idea, one visual logic, one emotional turn, and an ending that lands before the viewer's thumb moves. That compression is where AI video tools help most — and where careless workflows fall apart, because every inconsistency has nowhere to hide in ninety seconds.
Most teams do not lose time on rendering. They lose it on re-deciding: re-deciding the concept, the look, the shot order, the voice, the ending. The system below front-loads those decisions, keeps generation time predictable, and leaves you with something that reads as intentional rather than assembled.
Why 90 seconds is the hardest length to produce well
A 90-second piece typically needs 12 to 20 shots at 4 to 7 seconds each. That is a lot of visual decisions for a very short runtime, and it creates a specific failure mode: a montage with no spine. Pretty frames, no through-line.
Longer formats forgive this. A five-minute explainer can carry a weak middle because the viewer is already invested. Ninety seconds cannot. Each shot has to do at least two jobs — advance the idea and establish place, or carry emotion and plant the next cut.
The practical consequence for AI production is that shot count, not runtime, drives your workload. A 30-second spot with 14 shots is more work than a 90-second piece with 8 longer shots. Plan the shot count first, then let the duration follow.
The efficiency metric worth tracking: concept to finished cut
Render speed is the least interesting number in your pipeline. What matters is the elapsed time from a blank page to a file you would actually publish. Call it concept-to-cut time.
For a solo creator working with an AI video generator, a realistic target breakdown looks like this:
| Phase | Target | Most common failure |
|---|---|---|
| Brief and angle | 10 min | Skipping it, then rebuilding the concept mid-edit |
| Beat sheet | 15 min | Writing prose instead of beats |
| Shot spec | 25 min | Vague prompts that force regeneration |
| Generation | 20–30 min | Regenerating shots that were never specified |
| Assembly | 30 min | Discovering the story does not cut together |
| Sound and voice | 20 min | Leaving audio to the end |
| QC and variants | 10 min | Exporting one aspect ratio only |
Two levers move this the most. The first is reducing regeneration cycles: going from four takes per shot to 1.5 takes per shot roughly halves the whole schedule. The second is locking the look before generating motion, which prevents the quiet disaster of a video where shot 3 is warm daylight and shot 9 is cool overcast for no narrative reason.
A repeatable production workflow for short-form AI video
The workflow below assumes you are generating clips and assembling them in an editor. It works whether you are making a product teaser, a social explainer, or a narrative micro-story.
Step 1: Write the beat sheet in six lines
Six lines, one per beat. Not paragraphs — beats. Something like: cold open on a problem, establish the world, first attempt fails, turn, resolution, final image. If you cannot write six lines, you do not have a 90-second video yet; you have a mood.
For each line, write the emotional state you want the viewer in at that moment. This matters later, when you are choosing between two takes that are technically similar. Pick the one that serves the beat.
Step 2: Turn the beats into a shot spec table
Every shot gets a row with fixed columns: shot number, duration, subject, action, environment, camera, lighting, and audio intent. This table is your production contract. It stops the mid-generation drift where you start improvising shots because the prompt field is empty and inspiring.
Keep durations realistic. Four to six seconds is the sweet spot for AI-generated motion, because longer clips tend to introduce unwanted drift in hands, faces, and background geometry. If a beat needs nine seconds, split it into two shots and cut on movement.
Step 3: Lock the look with stills before generating motion
Generate three to five reference frames for the whole piece. Pick one. Write a fixed descriptor block — lens, light quality, color palette, texture, grade — and paste that block into every motion prompt. This single habit does more for perceived production value than any post-processing trick.
An AI image generator is genuinely useful here: iterate on stills cheaply, then commit to motion only once the frame is right.
Step 4: Generate in batches by scene, not by shot order
Generate all shots that share a location, lighting setup, and wardrobe in one sitting. Context switching is expensive: every time you change the descriptor block, you risk small inconsistencies that will read as errors on the timeline.
Generate two takes per shot maximum. Review them side by side at full size, not as thumbnails. Choose on the basis of the beat, not the novelty of the motion.
Step 5: Assemble against a scratch audio bed
Put a rough voiceover or a temp music track on the timeline before you place a single clip. Cutting to audio is the fastest way to discover that shot 7 is redundant and shot 2 is too long. Silent assembly produces videos that feel like slideshows, because the rhythm was never defined by anything.
Step 6: Finish, then export variants
Captions, safe zones, and aspect ratios come last. Export a vertical master, a square crop, and a 16:9 version from the same timeline, checking that text and key action survive every crop.
Prompt patterns that survive the edit
A prompt is not a wish. It is a specification with a limited number of slots. Prompts that survive the edit share three traits: they name one primary action, they describe the camera explicitly, and they repeat a fixed style block.
Name one action per clip
"A cyclist turns a corner and accelerates" beats "a cyclist rides through a city thinking about the day." The second prompt gives the model interiority it cannot render. One subject, one verb, one direction of movement.
Specify camera behavior, not just camera position
Slow push in, static locked-off frame, handheld follow, lateral tracking. These are different visual grammars. Mixing them randomly inside a 90-second piece is one of the most common reasons AI video feels unprofessional — the camera has no consistent personality.
Use a fixed style block and a short negative list
Paste the same 15 to 25 words of style language into every prompt: lens, light, palette, grain, contrast. Then keep a small negative list for the problems you actually see — warped hands, text artifacts, floating objects, oversaturated skin. Keep it short; long negative lists tend to fight the positive prompt.
If you are starting from scratch, a prompt library is faster than inventing phrasing from zero, and you can adapt existing structures to your subject instead of rebuilding the vocabulary each time.
Keeping visual continuity across a dozen shots
Continuity is where AI short-form lives or dies, and it is mostly a planning problem rather than a model problem.
Anchor frames. Keep the chosen reference still open beside your editor. Every new prompt gets checked against it for palette and light direction.
Screen direction. If your subject moves left to right in shot 2, keep moving left to right until you deliberately reverse it. Reversals read as mistakes unless they are motivated.
Wardrobe and props. Write them into the style block, not just the shot description. "Same jacket, same backpack" repeated is not redundant; it is the mechanism that keeps a character recognizable.
Cut on action. Place cuts mid-movement rather than at rest. A cut during a turn, a reach, or a step hides small inconsistencies in pose and framing far better than a cut between two static frames.
A color script. Assign each beat a dominant tone — cool for the problem, warm for the resolution — and hold it. Deliberate color movement reads as craft; random color movement reads as a rendering error.
Sound and pacing: the parts AI still cannot decide for you
Ninety seconds of silence is a lifetime. Audio is not a finishing step; it is half the pacing.
Record or generate the voiceover first and cut the picture to it. Human speech has its own rhythm, and trying to fit a voice to a locked picture produces either rushed lines or dead air. If there is no voice, use a music bed with a clear structural change at the turn of your story — most tracks give you one around 45 to 60 seconds, which maps neatly onto a 90-second arc.
Then add punctuation: a whoosh on a transition, a click on a product reveal, room tone under quiet moments. These small sounds are what make generated footage feel grounded rather than floaty.
Pacing rule of thumb: never let a shot outlive its information. If the viewer has understood the frame, cut. Most amateur short-form is 20 percent too long, and the fix is almost always in the first ten seconds.
A pre-export quality control checklist
Run this before any export. It takes ten minutes and saves an afternoon of re-uploading.
- Mute test. Watch with sound off. Does the story still read?
- First-frame test. Is the opening frame legible as a thumbnail at small size?
- One-speed watch. Watch at normal speed, not scrubbed. Judder and awkward motion hide in scrubbing.
- Caption check. Captions inside safe zones, no clipping, correct line breaks on short words.
- Ending hold. Give the final frame half a second to breathe. Cutting instantly to black reads as an error.
- Crop check. Verify vertical, square, and widescreen versions individually.
- Audio peak check. Normalize so dialogue sits consistently above music across the whole piece.
Mistakes that quietly cost an hour per video
Generating before specifying. The single biggest time sink. Every unspecified shot becomes three or four exploratory generations.
Chasing a perfect take. In 90 seconds, a good shot that cuts well beats a perfect shot that does not. Set a two-take limit and move on.
Changing the style block mid-project. It feels like an improvement in isolation and looks like an error in sequence.
Writing a script instead of a beat sheet. Prose invites performance direction the model cannot follow; beats invite visuals.
Skipping the scratch audio. You will re-time the entire edit later, and the second pass is always slower than the first.
Producing one aspect ratio. Most distribution happens vertically. Plan for it in the shot spec so key action stays in the middle of frame.
Choosing your production stack
You do not need many tools. You need one generator you understand deeply, one editor, and one place to keep your prompts and reference frames.
When evaluating an AI video tool for short-form work, weigh these criteria above raw feature lists:
- Shot-level control. Can you specify camera movement and duration per clip, not just a global style?
- Reference consistency. Can you feed the same reference frame across multiple shots?
- Iteration speed. How fast is a rejected take, and does the interface make side-by-side comparison easy?
- Aspect ratio output. Vertical-first, not vertical-as-an-afterthought.
- Cost predictability. A workflow you cannot afford to iterate in is a workflow you will not use.
A workable pipeline looks like this: draft the beat sheet, iterate on reference stills, then generate clips, assemble, and finish audio last. Orelon is built for that middle stretch — cinematic ideas in motion, with shot-level prompting rather than one-shot roulette. You can try the AI video generator for clip generation, iterate frames with the AI image generator, and start from a proven structure using video templates when you need to ship on a deadline rather than experiment.
If you are comparing options, the AI video generator alternatives overview is a useful sanity check, and the Orelon blog covers specific techniques for pacing, prompting, and continuity in more depth.
FAQ
How long should each AI-generated clip be in a 90-second video?
Four to six seconds is the practical sweet spot. Longer clips tend to accumulate drift in faces, hands, and background geometry, and they give you less editorial control. If a beat needs more time, split it into two shots and cut on movement.
Do I need a script or a shot list?
A shot list, every time. A script describes performance and subtext, which a video model cannot act on. A shot list with subject, action, camera, and lighting gives the model something executable and gives you something to check the output against.
How many shots does a 90-second video need?
Usually 12 to 20. That number drives your workload far more than runtime does, so decide it early. A calmer piece with nine longer shots can be stronger than a 20-shot montage, and it is significantly faster to produce.
How do I stop the look from shifting between shots?
Write a fixed style block — lens, light, palette, texture, grade — and paste it into every prompt unchanged. Keep one chosen reference still open while you work and compare each new clip against it before accepting it.
Can AI handle the voiceover too?
Yes, and it is a reasonable option for explainers and social cuts. But decide the voice before you edit picture. Cutting picture to a human or synthetic voice produces better pacing than fitting a voice to a locked edit.
What is the fastest way to cut my production time in half?
Reduce regeneration. Specify every shot before generating, cap yourself at two takes, and batch generations by location and lighting. Most wasted hours are not rendering hours; they are indecision hours.
Start with the beat sheet, not the prompt
The fastest improvement to your short-form output is not a new model. It is deciding what the 90 seconds are actually about, specifying twelve to twenty shots precisely enough to generate once, and cutting to audio you chose in advance. Do that three times and the workflow becomes muscle memory.
When you are ready to put frames on the timeline, start creating with Orelon — bring a six-line beat sheet and a fixed style block, and the tools will do the rest of the heavy lifting.

