AI for Filmmaking and Vertical Video: A Creator Workflow

For most of cinema history, the cost of a shot was proportional to its ambition. A crane move, a crowd, a rain-soaked street at night — each required equipment, permits, people, and time. That equation has not disappeared, but it has loosened. Generative video tools, better phone cameras, and faster editing software have collapsed the distance between an idea and a watchable frame.

At the same time, the place where most video is watched has changed shape. Vertical feeds dominate short-form platforms, and audiences now expect a story to establish itself in under two seconds. A filmmaker who once only thought in 16:9 frames has to think in two aspect ratios, two paces, and often two entirely different audiences for the same footage.

The practical result is that AI is not a replacement for craft. It is a set of shortcuts, cleanup tools, and expansion options that sit inside an otherwise normal production pipeline. The creators getting the best results are not the ones generating the most clips. They are the ones who plan properly, shoot deliberately, and use generated material only where it genuinely improves the cut.

This guide walks through a complete workflow: story first, shoot planning second, capture third, then an AI-assisted post pipeline, vertical-specific editing, tool selection criteria, and the mistakes that quietly ruin otherwise good projects.

Start With Story Logic, Not With Tools

The most common failure mode in AI-assisted video is starting with the tool. Someone opens a text-to-video generator, writes a prompt, gets a beautiful eight-second clip, and then has no idea what to do with it. The clip is impressive and useless.

Reverse the order.

Define the single promise of the piece

Before anything else, write one sentence that describes what the viewer gets. Not the topic — the payoff. "You will understand why your edit feels slow even when the cuts are fast" is a promise. "A video about editing" is not.

That sentence determines length, aspect ratio, tone, and whether you need a talking head at all. A 20-second vertical explainer with one idea beats a four-minute horizontal essay with six half-ideas, especially in feeds where the first three seconds decide everything.

Map beats, not shots

Write a beat sheet with five to nine beats. Each beat is a change: new information, new emotion, new visual. Beats are format-agnostic, which matters because you will likely deliver the same project as a vertical cut and a wide cut.

A simple three-column table works well: beat, what the viewer learns, what the viewer sees. If a beat has no visual answer, it is usually a candidate for narration over b-roll or a text card rather than a shot you have to produce.

Decide what must be real

Generative video is strongest at texture: skies, crowds, environments, abstract motion, impossible camera moves. It is weakest at specific human performance, precise brand assets, and anything requiring an exact match to a real location or product.

So mark each beat as real, generated, or hybrid. Real means you shoot it. Generated means you produce it with an AI video model. Hybrid means you shoot a plate and extend, clean, or enhance it. This single classification saves more time than any prompt trick, because it tells you what to schedule and what to generate while you are still calm and objective.

Planning a Shoot That Works in Both Formats

If you want one shoot to serve both a wide format and a vertical format, plan for it. Retrofitting is possible, but it wastes resolution and framing.

The dual-frame shot list

For every setup, note two things: the wide composition and the vertical safe area. On a phone, that usually means shooting wider than you think and keeping key subjects near the vertical center. On a mirrorless or cinema camera, it means recording in the highest resolution you can reasonably handle so you have vertical crop room later.

A useful discipline is to design each setup so the vertical version still reads without the edges. Faces centered, hands and props inside a central column, and important text never near the frame boundary. If a wide shot only works because of what happens at the far left of frame, plan a separate vertical shot rather than cropping.

Location and light planning

Apps that track the sun position are more valuable than most gear. Knowing when the light will rake across a wall lets you schedule the shots that need it and batch the flat-light shots together. For vertical work, backlight is your friend: a bright background with a shaded subject reads beautifully on a small screen and adds separation that a flat frame lacks.

Bring practical lights if you can. A single small LED panel with a softbox, positioned just off-axis from the phone, transforms a phone interview more than any camera upgrade.

Build a shot budget

Instead of promising yourself a hundred setups, decide how many finished shots the piece actually needs. A tight 60-second vertical video often needs 12 to 20 shots. A three-minute narrative short might need 40 to 60, plus coverage.

Write the number down. It prevents the classic trap of shooting endlessly and then discovering in the edit that you have 90 clips and no coverage of the one beat that matters.

Capture: Getting Clean Footage on a Phone

Modern phones produce genuinely broadcast-usable images in good light. What they punish is movement, mixed lighting, and bad audio.

Lock exposure and focus

Auto exposure will hunt the moment a subject walks past a bright window, and auto focus will drift during a slow push-in. Lock both before the take and re-adjust between takes. Most phones expose this through a long press on the subject in the camera app.

Shoot at the highest frame rate you need and no higher. Higher frame rates cost light and storage and can make motion look soapy when played back at normal speed.

Stabilization and movement

Small sensors exaggerate shake. The cheapest upgrade is a gimbal; the next cheapest is a shoulder rig or a simple two-handed grip with elbows tucked. If you have neither, lean on techniques that stabilize naturally: slow walks with bent knees, panning from a fixed pivot, and static shots cut together rather than one long moving take.

Be wary of in-camera electronic stabilization at the widest crop setting — it can warp edges and make vertical crops even tighter than expected.

Audio is half the film

Viewers forgive soft focus. They do not forgive hollow, echoey dialogue. Get a lavalier mic on the subject, or a shotgun mic on a boom just out of frame, or at minimum record a separate audio track on a second phone placed close to the speaker.

Always record 20 to 30 seconds of room tone. It is the single most useful asset for smoothing audio edits later, and it costs nothing.

Shoot vertical natively when it is the primary deliverable

If vertical is the main format, rotate the phone and shoot vertical. Cropping a horizontal capture to 9:16 throws away more than half your pixels and leaves you with a soft image on large screens. Only shoot wide-first when the wide version is genuinely the flagship cut.

The AI-Assisted Post-Production Pipeline

Post is where AI saves the most time, provided you keep it in a supporting role.

Step 1: Ingest and organize

Create a project structure before importing anything: footage, audio, graphics, generated, exports. Rename key clips descriptively the moment they land. Use markers for the best takes instead of relying on memory.

This step feels administrative and is actually creative. An organized timeline lets you find the emotional beat you remembered instead of settling for the clip you can locate.

Step 2: Radio edit first

Build the story with audio only. Cut narration, interview bites, and sound design into a sequence that works with your eyes closed. Then lay pictures on top. This prevents the seductive trap of cutting to a pretty shot that does not serve the story.

AI transcription tools make this stage fast. Transcribe everything, search for the exact phrase you want, and drop it into the timeline with one click. Transcript-based editing has become one of the biggest real-world time savers in video work.

Step 3: Generated inserts and cleanups

Now bring in generative video where the plan says "generated." Typical uses:

Environment establishing shots that would cost a permit or a travel day.
Transitions and textures — smoke, particles, light leaks — that need to be custom rather than stock.
Extension of an existing shot when you need two more seconds of a push-in.
Cleanup: removing a boom shadow, a reflection, or a passing car.

Generate several variations per shot, not one. Then pick in context. A clip that looks astonishing in isolation often fails next to real footage because of a subtle mismatch in grain, motion blur, or lens character. Matching grain and adding a light vignette can do more for believability than regenerating the clip ten times.

Step 4: Voice, music, and sound design

Synthetic voice is useful for scratch narration and internal review versions, and it can work in final delivery for informational content where the voice is not the personality. For anything character-driven, a real performance still wins.

Music should be chosen after the radio edit, not before. When the story exists, you can hear what the music needs to do — lift, unsettle, or stay out of the way. Then duck it properly under dialogue rather than riding faders by feel.

Step 5: Finishing

Grade for consistency across real and generated shots, then export multiple versions: 9:16, 1:1, and 16:9 if needed. Keep a version without burned-in subtitles so you can restyle later.

Editing for Vertical: Pacing, Type, and Retention

Vertical video is not a cropped movie. It is a different grammar.

Pace is faster, but not frantic

Average shot length in vertical feeds is short — often one to three seconds. That does not mean every cut must be a jump cut. Vary rhythm: three quick cuts, then one longer hold to let a moment land. Constant maximum speed is exhausting and flattens emphasis.

Type has to be legible

On a phone at arm's length, safe text height is far larger than most editors assume. Keep captions in the middle-upper third, away from platform UI at the bottom and the top. Use one typeface, two weights, and never more than three lines at once.

Burned-in subtitles are close to mandatory. Most vertical viewing happens muted, and auto-generated captions from platforms are inconsistent. Burn them in, then correct every line manually — including names and jargon.

Keep something moving, but not everything

Loops, parallax, slow pushes, animated text — these help retain attention in a narrow frame. But a frame where everything moves is noise. Choose one motion motif per video and stay consistent.

Design the first frame and the last frame

The first frame must work as a thumbnail and a hook. The last frame should either loop back to the first or hand the viewer a clear next action. Cutting off mid-sentence is a wasted ending.

Scaling Production Value Without a Big Crew

Ambition is cheap now. Production value comes from consistency and detail, not crew size.

Repeat locations. A three-person apartment shoot five times across a month reads as a world. Twelve random locations read as a mess.

Own a kit you understand. One camera, two lenses, one mic, one light, one diffuser. Knowing your kit cold beats owning a large bag you have to think about.

Build a reusable look. A LUT or grade preset, a title package, a subtitle style, and a consistent sound palette give small projects a house identity. Audiences recognize consistency long before they recognize budget.

Batch production. Write four scripts at once, then shoot all four in one day. Setting up lights is the expensive part; shooting the second video after the first is nearly free.

Keep a stock library of your own. Every shoot, grab 30 seconds of sky, texture, traffic, and hands. Over a year, that becomes the cheapest b-roll library you will ever own.

Tool Selection: A Practical Decision Framework

Tool lists age quickly. Criteria do not. Judge any AI video tool on these axes:

Controllability. Can you specify camera motion, subject placement, and duration? Tools that offer image-to-video, motion controls, or reference inputs are far more useful in real projects than prompt-only generators.

Consistency. Can it hold a character or a location across multiple shots? For anything narrative, consistent characters matter more than maximum realism.

Resolution and licensing. Check output resolution, commercial usage terms, and whether generated content can be used in paid work without restriction.

Speed and iteration cost. You will generate ten clips to use one. A tool that takes two minutes per attempt fits a fast workflow better than one that takes twenty, even if the slower tool is prettier.

Integration. HTML-free, plain exports matter: can you get ProRes or high-bitrate H.264 out? Does it hand off cleanly to your editor?

A practical stack for most solo creators looks like this:

Capture: a recent phone plus a lavalier and one small LED.
Generation: one general-purpose image-to-video model and one specialized model for the trickiest shot type.
Editing: a fast editor for assembly and a color-capable editor for finishing, or one tool that does both.
Audio: a transcription tool, a music library, and a small set of whoosh and impact sounds you reuse.
Delivery: an encoder or export preset set that produces all aspect ratios in one pass.

Publishing, Testing, and Iterating

Publishing is part of the workflow, not an afterthought.

Prepare a small metadata package for every video: three title options, a written hook, a thumbnail or first-frame still, and two caption variants. Then publish consistently and track two numbers that actually mean something — average view duration and completion rate — rather than raw views.

If completion rate is low, the problem is almost always the first three seconds or the pacing in the middle. If completion is high but reach is low, the issue is usually packaging: title, thumbnail, or topic fit.

Run one experiment at a time. Change the hook in one video, change the length in the next, and keep everything else stable so you can attribute the result.

Common Mistakes and an FAQ

Mistake: generating before writing

If there is no beat sheet, generation becomes aimless. Write the structure first, then generate against it.

Mistake: mixing lighting temperatures in generated and real shots

A generated shot with cool daylight next to a warm practical shot reads as fake instantly. Grade both to a common neutral before you judge the cut.

Mistake: over-relying on spectacle

Awe lasts about two seconds. Story, contrast, and clarity hold attention for minutes.

Mistake: ignoring audio

Bad audio is the number one reason viewers leave a video that looks good.

Mistake: one aspect ratio for everything

Deliver each major platform its native shape. Cropping at the last minute degrades both versions.

Frequently asked questions

Do I need a cinema camera to make films with AI assistance?
No. A recent phone with locked exposure, decent light, and clean audio covers most formats. Spend your money on sound and lighting before sensors.

Can generated footage be mixed with real footage?
Yes, and it usually looks best when it is used for environments, textures, and inserts rather than for human performance. Match grain, motion blur, and lens character to blend the two.

How long should a vertical video be?
As long as the idea needs and no longer. For single-idea content, 20 to 45 seconds is a reliable range. Story-driven vertical pieces can hold attention for two to three minutes if the pacing varies.

Is it better to shoot vertical or crop later?
Shoot natively in the format that matters most. Crop only when you have spare resolution and a generous vertical safe area.

How many generated shots should a project contain?
If viewers can tell which shots are generated, it is too many, or the blending is too rough. Aim for generated material to be invisible — a supporting element, not the headline.

What is the fastest way to improve?
Finish and publish something small every week. A completed 30-second video teaches more than three unfinished ambitious ones, because only finished work shows you where your pipeline actually breaks.

The through-line across all of it is simple: decide what the viewer gets, plan the beats, shoot what only real footage can do, generate what only generation can do, cut for the shape of the screen, and publish consistently enough to learn from the numbers. Tools will keep changing. That workflow will not.