How to make your first AI video in an afternoon

Most first attempts at an AI video end the same way: a folder of clips, each one quite good, and no video. The person generated first and planned never, so the clips do not go together, nothing says what the video is for, and the afternoon ends with a shrug.

This is the other way round. It produces a one-minute video with a voice, pictures that match the voice, captions and music, exported and watchable, in an afternoon. The tools are the easy part. The order is the whole trick.

1. Write one sentence (ten minutes)

Before anything is generated, write down what the video is for, in one sentence, with a verb in it. "Show that our espresso machine heats up in forty seconds." "Explain why the moon has phases to a nine-year-old." "Make people want to visit the coast path in October."

Every later decision will be measured against that sentence, which is what makes them quick. A shot either serves it or it is cut. If you cannot write the sentence, you are not ready to generate, and generating will not help.

Then write the narration: six to ten short sentences that say the thing. Read it aloud and time it. Narration runs around two and a half words a second, so a hundred and fifty words is about a minute, and a minute is the right length for a first video. Longer is not more impressive; it is more shots to get right.

2. Make the stills, not the clips (forty minutes)

Here is the step that beginners skip and experienced users never do. For each sentence of the narration, generate a still image that could sit under it. Four variants per run, pick one, move on. Images are cheap and fast; a wrong one costs you seconds.

Give the stills a shared look from the start: the same lighting phrase, the same lens phrase, the same palette words in every prompt. "Soft morning light, 35mm, muted warm tones" at the end of every prompt does more for consistency than any amount of fixing later.

By the end of this step you have a storyboard that already exists as pictures. Lay them on the timeline in order and press play. Even as still images with no sound, it will tell you whether the video works. If a sentence has no picture that fits it, change the sentence. This is the cheapest moment to find that out.

3. Record or generate the voice (twenty minutes)

Now the narration becomes a file. Two ways, and both are fine.

Record it yourself. The recorder is in the Assets panel; do it in a quiet room, at a conversational pace, and do not worry about the ums. (They can be cut later, and a first video does not need them cut.)

Or generate it: choose a voice, paste the script, run it. Spell out numbers and abbreviations the way they should be said ("forty seconds", not "40s") because the voice reads what is written. Generate it in two or three pieces rather than one if the script has sections; a fix later then re-reads one piece.

Drop the voice on an audio lane. Its length is now the length of the video. Everything else is cut to it.

4. Animate the stills that need it (thirty minutes)

Not every still needs to become a clip. A still with a slow Ken Burns move over it holds for four or five seconds perfectly well, costs nothing, and is how most documentary footage of photographs has always been done. Use it for at least half the shots.

For the shots that need real motion, the subject doing something, send the still to the video generator as the seed for Animate, and describe only the motion: "she lifts the cup and steam rises", "slow push in toward the lighthouse". The picture is already decided by the still, which is why this step goes right far more often than a prompt from scratch would. Run the batch and keep editing; they render in the background.

Expect one in four to need a second attempt. That is normal and is why the budget for this step is thirty minutes rather than ten.

5. Cut to the words (thirty minutes)

With the voice on its lane and the clips in the library, the edit is mostly arithmetic. Each clip starts where its sentence starts. Drag it there; snapping will pull it to the playhead. Where a clip is longer than its sentence, park the playhead at the end of the sentence, press Split, and delete the tail. (Pulling a video clip's edge changes its speed rather than trimming it, which surprises everyone once.)

Leave no gaps and no shot longer than about six seconds. If a sentence runs long, give it two shots: a wide and a close, or the still and then the animated version of it. Add a dissolve only where time passes; everywhere else, cut.

Watch it through once with the sound on. It will feel nearly done, which is the right feeling at this point.

6. Captions and music (twenty minutes)

Select the voice clip and generate captions. They arrive as text layers timed to the words; fix any spelling, and make them larger than looks right on a monitor, because most people will watch on a phone. Captions are not optional for anything that will be seen in a feed with the sound off, which is most things.

Then a music bed. Generate one from a short style prompt and the length of the video, drop it on its own lane, and flag it as Duck in its Audio tab so it dips under the voice automatically. Start the music a little quieter than feels right; it will be too loud on headphones otherwise.

7. Export and watch it on a phone (ten minutes)

Export at 1080p, H.264, 30 fps. The render runs on the server, so the tab can close. Then watch the file on a phone, not on the monitor you edited on, because the phone is where its faults show: captions too small, a shot too dark, music too loud.

Fix the one worst thing, export again, and stop. A first video does not need to be good. It needs to be finished, because the second one is where you get good, and nobody makes a second one with a folder of clips and no first.

What this afternoon teaches

The order was: words, stills, voice, motion, cut, dress, export. Notice that the expensive and unpredictable step, generating video, came fourth, after everything it depended on was decided, and that the thing that fixed the length was the voice, not the clips. That order does not change when the video is ten minutes long and the stakes are a client. It only gets more important.

The order that works

1. One sentence, then the narration, timed by reading it aloud.
2. A still per sentence, four variants at a time, with a shared look in every prompt.
3. The voice as a file, recorded or generated, on its own lane; its length is the video's length.
4. Animate only what needs motion; Ken Burns on the rest.
5. Cut each shot to its sentence, nothing over six seconds, dissolves only for time passing.
6. Captions from the voice, music ducked under it.
7. Export, watch on a phone, fix one thing, stop.