TikTok Video Maker Workflow: Create Vertical Videos With AI

TikTok-style video rewards a narrow skill: saying one thing clearly in a frame that fills a phone screen, and doing it before the thumb moves. AI generation made footage cheap. It didn't make the decisions — what to show, in what order, for how long — any easier, and that gap is where most AI short-form fails.

What follows is a production workflow rather than a tool tour: how to plan a vertical video, how to prompt for 9:16 framing, how to hold a character or product consistent across clips, how to handle sound and captions, and how to cut for retention. It applies whether you publish to TikTok, Reels, Shorts, or all three.

Why vertical short-form breaks normal video habits

Most video training assumes landscape, a settled viewer, and time to establish context. Vertical short-form assumes the opposite on all three counts.

The frame. A 9:16 canvas at 1080×1920 leaves very little horizontal room, so wide establishing shots become unreadable. Faces, hands, and objects need to fill the frame more aggressively than in landscape work, and anything important belongs in the center band — roughly the middle 60% vertically — because the interface eats the top and bottom edges.

The attention window. The real decision point arrives in the first one to two seconds, before any narrative exists. If the opening frame is a slow drift toward nothing, the viewer is gone. Opening frames need motion, contrast, or a face.

The loop. Short-form video is often watched on repeat, and a clip whose last frame resembles its first gets replayed without the viewer consciously deciding to. That's free retention if you design for it and invisible if you don't.

All three constraints are cheap to handle during planning and nearly impossible to fix afterwards. Which is why the workflow starts well before generation.

Write the beat sheet before you generate anything

A beat sheet for a 30-second vertical video isn't a screenplay. It's five lines naming what the viewer sees and what changes.

1. Hook (0–2s). One visual that opens a question: an unusual object, a hand entering frame, a result shown before its process.
2. Premise (2–6s). What the video is about, communicated visually rather than through text.
3. Development (6–20s). Two to four escalations. Each one changes something — new angle, new scale, new setting.
4. Payoff (20–27s). The thing the hook promised, delivered.
5. Close (27–30s). A frame that resolves cleanly or loops back into the opener.

Write the hook last. Once you know the payoff, the hook becomes obvious — it's the payoff stripped of context. Hooks written first tend to be generic "watch this" statements that don't actually open a question.

For AI generation specifically, the beat sheet also determines clip count. Five beats usually map to six to ten shots, because the development section needs more coverage than the rest.

Build a shot list that already knows where it will cut

The most common failure in AI short-form is generating beautiful clips that don't edit together, because each shot was made in isolation and judged on its own.

Build the list with fixed columns: shot number, duration target, subject, action, camera move, lighting, and the transition into the next shot. That last column is the one people skip and the one that decides whether a video feels intentional.

Rules that survive contact with an editor:

• Never generate below two seconds. Motion coherence drops and you end up compensating in the edit.
• Cap single shots at four to five seconds. Longer reads as a stall in vertical. If a shot needs six seconds, it probably needs to be two shots.
• Alternate scale deliberately. Wide, close, wide, close. Alternating scale does more for perceived pace than faster cutting.
• Plan match cuts. A hand reaching for a mug in shot three that becomes a hand reaching for something else in shot four gives the edit a rhythm viewers feel without noticing.
• Generate payoff coverage. Two or three variations of the final shot. A strong 25 seconds ending on a weak frame wastes all of it.

If you'd rather start from a fixed structure than rebuild one per concept, video templates save time — the value is in the scaffolding, not the preset look.

Prompt for 9:16 framing that survives the crop

Many creators assume specifying a vertical aspect ratio is enough. It isn't. Models trained mostly on landscape footage still compose like a landscape camera, placing the subject small and centered with empty sky above. Forced into vertical, that becomes a strip of nothing with a subject at the bottom.

Prompt composition explicitly. A structure that works well:

Subject and wardrobe → action → camera position and lens → framing note → lighting and time of day → style and grain → duration and motion.

An example for a close product beat:

A woman in a rust linen shirt lifts a matte ceramic mug toward the camera, steam curling upward. Camera at eye level, 50mm equivalent, framed chest-up with the mug in the lower center third. Warm late-afternoon window light from the left, soft shadow falloff. Natural color, slight film grain, handheld micro-movement. Four-second shot.

Two details make that prompt work. The framing instruction ("chest-up," "lower center third") is explicit about the crop. And the motion is small — big camera moves are the fastest way to make a vertical clip look like a drone reel.

Language to treat carefully:

• "Drone shot," "sweeping aerial," "wide establishing" rarely survive a vertical crop intact.
• "Tiny push-in," "micro pan," "slow tilt" read as energy in vertical without destroying the frame.
• "Negative space" is ambiguous in vertical, since the top of the frame is usually covered by captions. Say where the space is if you need it.

Keep a personal prompt library of the seven or eight skeletons that reliably produced usable footage. Reusing structures is faster than inventing prompts, and it makes a series look coherent.

Keep characters, products, and style consistent

Consistency is what separates a channel from a pile of clips. Three techniques do most of the work.

First-frame anchoring. Generate or select a still of your subject, then use it as the reference or starting frame for every shot. This beats re-describing the subject in text, because descriptions drift — hair color shifts, the jacket changes, the mug loses its shape. If your tool supports image-conditioned generation, treat the still as source of truth, which makes an AI image generator part of the video workflow rather than a separate hobby.

A style bible of five attributes. Write down the five things that must stay identical across every clip: color palette, lighting direction, lens character, motion style, and grain or texture. Five fits inside a prompt without bloating it; anything beyond that is usually a detail viewers won't register.

Scene continuity by geography. If shots two, four, and seven happen in the same room, describe that room identically each time and name the light source. "Kitchen with a window on the left, warm afternoon light" beats "a kitchen" every time, because it gives the model a directional constraint.

Products are harder than people. A generated bottle will not match your actual bottle, and viewers notice. For product-accurate work, use AI footage for context, texture, and transitions, and insert the real product through stills, screenshots, or live-action cutaways. Blending generated and real footage isn't a compromise — it's how most commercial short-form is actually made.

Audio and captions are part of the edit

Two facts shape vertical audio. A large share of short-form viewing happens with sound off, and the first second of audio affects retention as much as the first second of visuals. Captions and sound design aren't polish; they're load-bearing.

Voiceover. Thirty seconds holds roughly 65–75 spoken words. Write to that budget and read it aloud before generating anything — sentences that look fine on screen collapse when spoken. If you're using text-to-speech, keep sentences short and skip parentheticals and adjective stacks.

Music. Trending audio is powerful and perishable; a sound that's everywhere this week is dated in three. If you're building a body of work, choose music that matches the energy arc instead of the current chart — a quiet opening into a percussive payoff supports almost any beat sheet.

Captions. Burn them in. Place them in the middle band, never in the top 15% or bottom 20%, keep them to two lines, and hold each phrase for at least a full second. Caption timing that runs slightly ahead of the voiceover reads as professional; captions that lag feel broken.

Silence is a tool. Cutting audio entirely for a quarter second before the payoff is the cheapest punctuation available, and it costs nothing.

Edit for retention, not for symmetry

The edit is where vertical video succeeds or fails, and the instincts that produce a beautiful landscape film work against you here.

Start with the first frame. Choose the most arresting frame you generated and put it at zero, even if it belongs later in the sequence, then build the opening two seconds backward from it. A clip that opens on motion, a face, or an unexpected object buys the next three seconds.

Cut on action — the hand entering, the turn of the head, the pour. Cutting on stillness feels abrupt in vertical and reads as a mistake.

Check the pacing curve. Short-form retains best when shot lengths shorten slightly across the runtime: 3.5s, 3s, 2.5s, 2s, 2s, then 3s. That final lengthening gives the payoff room to land.

Then design the loop, the step most people skip. If the last frame can cut to the first without a visible jump, the clip gets rewatched, and rewatches are the strongest signal a short-form feed has to work with. The simplest method: end on a frame similar in composition and color to the opener, so the transition feels intentional.

Finally, watch it three times in a row without pausing. Problems invisible on one pass — a caption lingering a beat too long, a shot repeating a composition you already used — become obvious on the second viewing, which is exactly how a chunk of your audience will see it.

A worked example: 30-second product teaser

Assume a ceramic mug brand, one character, two locations.

• Shot 1, hook (0–2.5s). Extreme close-up of steam rising off the mug against a dark background, shallow focus. Only motion is the steam and a tiny push-in. No text.
• Shot 2 (2.5–5s). Still vertical but wider: a woman at a kitchen window, mug in hand, three-quarters to camera. Same lighting direction as shot one.
• Shot 3 (5–8s). Close-up of hands wrapping the mug, thumb tracing the rim — the texture beat.
• Shot 4 (8–11s). Micro pan across a counter with the mug set down, morning light raking across it.
• Shot 5 (11–14s). Mid-close: she takes a sip, eyes closing for a beat. The human beat that keeps it from being a catalog.
• Shot 6 (14–18s). Insert of a slow pour, framing tight, sound of liquid filling.
• Shot 7 (18–22s). The mug on a desk by a window in a second location, same palette, different time of day — implies a daily ritual.
• Shot 8, payoff (22–27s). Slight pull-back revealing her at the desk, mug in frame, warm light.
• Shot 9, close (27–30s). Near-copy of shot one's composition. Steam, dark background. Loops cleanly.

Nine shots, roughly 65 words of voiceover, captions on the hook and payoff only. Every shot was generated from the same still of the mug and the same character, which is what makes the sequence read as one video instead of nine.

Common mistakes, batching, and what to measure

The recurring failure modes are predictable enough to check before publishing.

• Landscape compositions forced into vertical. Subject too small, too much empty frame. Fix it in the prompt, not the crop.
• Generating in isolation. Every shot looks great alone, nothing cuts together. Fix with a shot list and planned match cuts.
• Drifting subjects. Hair, wardrobe, or product changes between shots. Fix with first-frame anchoring and a five-attribute style bible.
• Over-long shots. Anything past five seconds in vertical reads as a stall.
• Captions in the interface zone. Text hidden behind UI elements at the top and bottom edges.
• Audio neglect. A clip muted in the first second never gets a second chance.

On batching: generate one concept at a time, not one shot at a time across many concepts. Ten variations of shot three teach you more about your prompt than one variation each of ten shots, because you can compare them directly. Batch the boring parts too — stills, voiceover takes, caption timing — and keep the edit as a single uninterrupted session, since edit decisions degrade when fragmented.

Three metrics matter almost everywhere: the share of viewers still watching at two seconds, average watch time as a percentage of runtime, and rewatch rate. A weak two-second hold means your opening frame is the problem. Steadily falling watch time means pacing. High completion with low rewatches means your ending isn't creating a loop. Change one variable at a time — changing hook, pacing, and music together tells you nothing except that something moved.

FAQ

Do Shorts, Reels, and TikTok need different exports? No. 9:16 at 1080×1920 covers all three. Check safe zones per platform, though, since interface overlays differ and captions near the bottom edge can get clipped.

How many generated clips does a 30-second video need? Six to ten usable shots, which in practice means generating two to three times that number. Budget on the assumption that roughly a third of outputs are usable.

Can AI footage look consistent enough for a series? Yes, if you anchor to stills and keep a written style bible. Consistency is a documentation problem more than a model problem — the creators who get it right are the ones who wrote their five attributes down.

Should everything be AI, or should I mix in real footage? Mix for anything product-accurate or person-specific. Generated footage is strongest for texture, atmosphere, transitions, and B-roll; real footage is strongest for identity and claims.

What's the minimum workflow if I have one afternoon? Beat sheet, shot list, eight prompts from a reusable skeleton, one generation batch, one edit session, captions burned in, one loop check. Skip the style bible the first time and write it right after — you'll know exactly what needs documenting once you've seen what drifted.

Make your next vertical video in Orelon

Everything above is deliberately tool-agnostic, because the decisions — beat sheet, shot list, framing, consistency, cut rhythm — are what decide whether a vertical video works. Orelon handles the generation layer of that workflow: Orelon's AI video generator turns a structured prompt into cinematic vertical footage, and the Orelon blog goes deeper on prompting, pacing, and production habits.

Start with one concept, one beat sheet, and eight shots. Publish it, look at the two-second hold, and change exactly one thing next time.