How to Make a Video with AI (Without It Looking Like AI)

Making a video with AI is easy. Making one that does not announce itself as AI is the actual skill, and almost nobody writes about that part.

This guide covers both: the workflow that gets you from an idea to a finished video, and the specific tells that make generated video look cheap, because knowing what to avoid is more useful than knowing what to type.

Start with the sentence, not the tool

The single biggest predictor of whether this goes well is whether you can say what the video is for in one sentence before you start.

Not "a video about our product." That produces a video about your product, which is to say something generic and forgettable. Something closer to: *"A 30-second explainer for people who just landed on our pricing page and are confused about the credit system."*

That sentence contains four things the tool needs and a vague brief does not: a length, an audience, a moment in their journey, and a specific confusion to resolve. Everything downstream gets better when those are decided.

If you cannot write that sentence, you do not have a video problem yet. You have a positioning problem, and no generator will solve it for you.

The order that works

Working in the wrong order is the most common way to waste time. Structure first, texture last: the same discipline as any edit.

1. 01Describe the whole thing. One or two sentences. Let it build a complete first cut, however rough. You are looking at shape, not quality.
2. 02Fix the structure. Is it saying the right things in the right order? Cut, reorder, or replace whole sections before touching anything else. Polishing a section you later delete is pure waste.
3. 03Fix the script. Now that the structure holds, the words matter. Tighten the voiceover, cut the throat-clearing, get to the point faster.
4. 04Fix the visuals. Shot by shot. This one is too generic, that one is the wrong mood, the third is fine.
5. 05Layer the audio. Music, and how it sits under the voice.
6. 06Caption it. Last, because captions follow the final timing.

The temptation is to fix the first shot you dislike immediately. Resist it. The first cut exists to show you the shape, and the shape is what determines whether the rest is worth doing.

The tells that make generated video look cheap

This is the part worth internalising. Generated video has a small number of recurring giveaways, and most of them are avoidable once you know to look.

Shots that hold too long

Generated clips are often produced at a fixed length and used at that length, which produces a video where every shot lasts the same four seconds. Real edits have wildly varying shot lengths: a half-second cutaway next to an eight-second hold. Uniform pacing reads as machine-made even when each individual shot is good.

Fix it by asking: "make the cuts less even" or trimming specific shots. Varying rhythm alone does more for perceived quality than upgrading any single shot.

The dead-eyed stock look

Generated people in generic settings (the smiling team around a laptop, the thoughtful person at a window) land in the same uncanny valley as bad stock footage, for the same reason. The content is unspecific, so the viewer registers it as filler and stops looking.

The fix is specificity. "A person at a desk" produces filler. "A close-up of hands scrubbing a timeline on a cluttered desk, late evening, screen light" produces a shot. Detail is what separates an image from a placeholder.

Music that ignores the voice

Generated scores tend to sit at one volume for the whole video. Real mixes duck the music under speech and let it come up in the gaps. Without that, the music fights the voiceover the entire time and the whole thing feels amateur, regardless of how good the track is.

Always ask for it explicitly: "duck the music under the voice."

Captions that are technically correct and visually wrong

Default caption styling is almost always too small, too centred, and timed to phrases rather than speech. It reads as an afterthought because it is one. Say where the video is going and ask for the style that fits: captions for TikTok and captions for a webinar are different objects.

Perfect, characterless framing

Every shot dead-centre, every subject perfectly lit, no handheld drift, no lens imperfection. Individually fine; collectively sterile. Asking for "handheld" or "shot on an older lens" or "slightly underexposed" adds the irregularity that reads as real.

Mix in something real

The most effective thing you can do to stop a video looking generated is to include something that is not.

A few seconds of you actually talking. A real screen recording of the real product. A photograph someone on your team took. One authentic element re-frames everything around it: the generated shots stop reading as substitutes for reality and start reading as illustration, which is a role they play well.

This is also why the generate-and-edit combination matters more than generation alone. A tool that only generates pushes you toward fully synthetic video, which is the version most likely to look cheap. Mixing your own footage with generated shots in one project is the workflow that produces something watchable.

Say what it is for, not how to build it

The instruction should describe the outcome, not the mechanism. You do not need to know how the music is produced or which model made a shot.

Weak instructions describe process:

• "Add a transition"
• "Insert b-roll"
• "Apply captions"

Strong instructions describe intent:

• "Make the jump between these two points less abrupt"
• "Cut away while I explain the pricing, the talking head is going flat here"
• "Caption it TikTok style, keep them clear of the bottom third"

The second set gives something to aim at. The first set gets you a technically correct result that misses the point. There is a fuller collection of these patterns in the prompt guide.

Point when describing gets awkward

Language runs out when you need to identify a specific thing in the frame. "Remove the object on the right" is ambiguous if there are three objects on the right, and describing which one takes longer than the edit is worth.

That is the case for drawing on the frame instead: pause, circle the thing, say what should happen to it. Removing a stray object, blurring a screen, keeping the right subject in frame during a vertical crop, all faster to point at than to describe.

Expect to iterate, and budget for it

The single most useful mental adjustment: you are not writing one perfect prompt. You are having a conversation.

The first result is a draft. You react to it, change one thing, look again. Five small instructions almost always beat one elaborate one, because you can see the effect of each and course-correct. An elaborate prompt gives you one shot and no diagnosis when it misses.

This is also why per-use pricing matters more than it sounds. If each iteration draws down a monthly quota, you feel the cost of every experiment and you stop experimenting, which is precisely the behaviour that produces worse videos.

A realistic first project

If you want to try this, do not start with something important. Start with something small enough that you will finish it:

1. 01A 30-second explanation of one thing you do. Describe it, get the first cut.
2. 02Cut it to 20 seconds. Length pressure forces every useful decision.
3. 03Caption it for wherever it is going.
4. 04Add music, ducked under the voice.
5. 05Watch it on a phone with the sound off.

That last step matters more than the four before it. Most video is judged in exactly those conditions, and most video fails there for reasons that are invisible on a desktop monitor with headphones on.

Then make the next one. The workflow gets fast quickly: what takes time at first is not the tool, it is deciding what you actually want to say.