1. Your First Usable Shot

The shot is five seconds long. Early on a grey morning, a bike courier in a grey rain jacket crouches and leaves a small paper parcel on the stone step of a bakery that has not opened yet. One person, one plain physical action, one place. Watch the first take closely and her hand goes half an inch into the parcel at the moment she puts it down. Catching flaws like that, and deciding what to do about them, is most of the work of getting a usable shot.

The courier, the bakery and the takes described here are invented; I made them up for teaching, and yours will differ. The course keeps coming back to this courier, because one subject and one clear action make problems visible. When the jacket changes colour between shots, you can see it.

Notice how little is in that brief. No dialogue, no second person, nothing that asks the model to understand a relationship or a joke. When you are learning, a small brief is the instrument you are measuring with.

The thing that makes the shot is a video generation model: a program trained on an enormous amount of existing film and footage, which produces new footage from a description. Some of these models run on a computer you own, which needs a powerful graphics card and some patience with software. Others are hosted: they run on somebody else's computers and you reach them through a web browser, the same way you reach your email. We start with a hosted one, because you can be generating today without installing anything.

Compared with filming or animation, generation is fast and cheap, and you pay for that in control. You cannot tell the model to move the actor's hand two inches left. You can only describe, look at what came back, and describe again. That is why the habit of looking matters more than any clever wording.

The service we will use is Runway. You log into its web app in a browser rather than installing a program, and its documentation recommends Google Chrome. There is no graphics card requirement and no code. You register with an email address, or by signing in with a Google or Apple account, and no credit card is required to start.

The exercise uses Gen-4.5, which requires a Standard plan or higher. You can register for free, but the free account does not unlock this model. Each generation spends credits, including takes you reject. Before starting, check the charge shown for a five-second clip and budget for three attempts. If you are not ready to pay, follow the example first and decide afterwards whether to try it.

In Runway's Tool Mode, select Video and choose Gen-4.5 from the model selector. Gen-4.5 can also animate an uploaded image, where the image supplies the appearance and composition and the prompt describes the movement. That is a different starting point. Today we are using text-to-video, so leave the image input empty.

Set the duration to five seconds. Gen-4.5 supports two to ten; five gives our small action room to finish without making the first attempt too expensive. Keep the widescreen frame, the standard seven-twenty picture and the default MP4 download. Leave upscaling and other delivery formats until you have a shot worth keeping, and leave sound design until you know which take you will keep. We are judging the picture first, and a convincing image can make you imagine the paper crunch and the birds before you have added either. Check the credit charge before generating.

The input is a text prompt: a plain-language box where you describe what appears in the frame and how it moves. Before writing it, two terms, because from here on they are how you and I will talk about a shot.

A shot is one continuous run of camera, from the moment it starts recording to the moment it stops. Cut, and you have a new shot. A scene is bigger: the action in one place and one stretch of time, usually made of several shots. The courier leaving the parcel at the bakery is a scene. The five seconds of the courier's hands putting it down is a shot. Everything a generation model gives you is a single shot, which means a scene is something you assemble, not something you ask for.

A prompt for a shot has named parts. First the subject: who or what the frame is about. A young bike courier in a grey rain jacket. Second the action: the one physical thing that happens. She sets a small paper parcel down on a stone step. Third the framing: how much of the subject the frame contains. A medium shot takes a person from roughly the waist up. A close shot is head and shoulders, or a pair of hands. A wide shot puts the whole figure in its surroundings. Framing is a decision about what the audience is allowed to notice.

Fourth, camera movement: whether the camera holds still or travels, as in a slow push-in or an orbit. Still is called static, and for a first shot it is the right choice, because a moving camera gives the model one more thing to get wrong. Describe what you want to see: write static camera rather than no camera shake. Fifth, light: where it comes from and what it feels like. Soft light from the left, early morning. Naming a direction gives the model something to be consistent about. Sixth and seventh are duration and aspect ratio, which you have already set: five seconds, widescreen.

Here is the prompt, at forty-four words:

Medium shot of a young bike courier in a grey rain jacket. She crouches and sets a small brown paper parcel down on a stone bakery step, then straightens up. Empty city street at dawn, wet pavement, soft light from the left. Static camera.

Each sentence has a job: subject and framing, action, setting and light, then camera movement. There is no magic word count. A longer prompt can be useful, but extra detail can also introduce competing requests, and you want each instruction easy to find and change.

Look at the action once more before generating, because it is the difference between a nice clip and a useful one. A shot can look handsome and still be no use to whoever cuts the film together, and often that is because the action does not finish. If the courier begins to crouch and the clip ends while she is still bending, an editor has nothing. They cannot cut away from a half-finished movement without it looking like a mistake. They need to see the parcel land. That is why the action is crouch, place, straighten. A complete action in five seconds means one small movement, not three, and that simplicity is what makes the footage cuttable.

Press Generate.

Then look at what came back. This is where you find out whether the credits were well spent.

The temptation is to watch it once, feel pleased or disappointed, and move on. Instead, watch it at least three times and give each pass a job. That habit is what separates directing from gambling.

First pass, watch the action only. Does it start and does it finish inside the five seconds? In the example take, the courier crouches at about half a second, the parcel is on the step by three and a half, and she is upright again by four and a half. There is half a second of nothing at the end. That is fine. A little air at the end is useful.

Second pass, watch the contact points. Where a body touches a thing, generation models are weakest. Do the fingers close around the parcel, or does the parcel simply travel with the hand? Do her feet meet the pavement, or float a few pixels above it? In the example take her right hand passes into the corner of the parcel as she lets go. Not badly, but if you froze the frame and pointed, someone would see it. That flaw has a name: broken contact at the moment of release.

Third pass, watch the face. Generated faces drift. Does she look like the same person at second one and second four? In the example, her face holds until she straightens, and then it changes: the jaw narrows and the eyes sit differently. That is called identity drift, and it matters because a story needs the same person to keep being the same person.

Then check the light. Does it keep coming from one direction? The example's does, mostly, though the puddles brighten oddly halfway through as if a second sun arrived. And check for text. Models love to invent lettering, especially on signs and shopfronts. There is a smear of not-quite-letters across the bakery window, not words, just the shape of words. In a finished film that is a distraction. Then listen once, so you know what sound work the saved clip will need.

So one generation, four flaws noticed, none of them fatal. The seconds you decide not to use are not failure. They are rejected work, and rejected work is a normal cost of producing one accepted shot. A working production counts credits and seconds spent per usable shot, not per generated shot. If you throw away four clips to keep one, your cost for that shot was five clips. That is a number, not a shame. People who hide their rejected work cannot estimate anything, and they end up believing every generated second is footage.

Now, one change. Exactly one.

Go after the broken contact first, because that is what an audience notices without knowing why. To address it, change one element of the brief: the framing. Go from a medium shot to a close shot on the hands and the parcel. A close frame gives the model far more pixels for the thing that has to be right, and it takes the face, the other flawed thing, out of the frame entirely.

Keep the action, light and camera movement as close as possible to the first request. The new generation may change other things too. This comparison will tell you whether the new take works; one pair of clips cannot prove that framing alone caused the improvement.

The revised prompt, at thirty-nine words:

Close shot on the hands of a young bike courier in a grey rain jacket. Her fingers set a small brown paper parcel down on a wet stone step and release it. Soft light from the left. Static camera.

Generate again.

In the example, the second clip fixes the targeted flaw. The fingers open, the parcel sits, and the shadow under it lands where the light says it should. The face problem is gone because there is no face. And the revision brought in a new problem: the parcel is already resting on the step when the clip begins, and her hands only let go. The placing is not there. As storytelling it is weaker, because a hand releasing something already at rest does not read as delivery.

A targeted revision can come back worse, and often it comes back worse in exactly this way: you fix what you named and lose something you were not watching. When that happens you have three options, and which one you pick is a budget decision as much as a craft one. You can accept the earlier clip, flaws and all, if its flaws are small enough to survive on screen. You can run the revision once more with the same single change plus one sentence protecting what you lost; here, adding that the hands lower the parcel onto the step. Or you can accept the new clip and adjust the brief around it, deciding that this shot is about release and something else will carry the placing.

Run the third version: close shot, hands lowering the parcel onto the step and releasing it, same light, static camera. In the example, that take has the full movement, clean contact and consistent light, with a faint wobble in the paper's edge in the last few frames. That is the accepted take. Three generations, fifteen seconds produced, five seconds kept.

Finish with two things side by side.

The first is the clip. Download it with the download arrow on the preview pane; it comes down as an MP4. Put it in a folder you will find again, with a name that says what it is rather than what the service called it, something like courier, bakery step, close on hands, take three. Keep the rejected takes in the same folder.

The second is a short written record beside it, in a plain text file or a note. Write down the service and the model version. Write down the prompt, exactly as you typed it. Write down the duration, aspect ratio and generation cost. Then one line on what you rejected and why: two earlier takes, the first for a hand passing through the parcel and a face that changed, the second because the parcel started already on the step. Keep the prompt and settings for those rejected takes too.

Six lines, two minutes of typing. That record is the first item in a library you will keep adding to. The moment you want this shot again, or a shot that rhymes with it, the note lets you start from where you finished instead of from a blank prompt box. It lets you resume the work without reconstructing your decisions; it does not promise an identical new generation.

Where generated sound is arriving

We left sound design for later in this exercise. Other workflows combine picture and sound during generation, so when you compare services, check which exact model produces the audio and what you can control.

Hailuo AI released a model called MiniMax H3 at the end of July twenty twenty-six. Its headline is what they call an omni-pass: stereo audio generated in the same text pass as the picture, including room tone, footstep-and-object sound, and dialogue. It also runs single clips up to fifteen seconds at two-K resolution, across six frame shapes from very wide to upright phone. The production task that changes is delivery of a standalone shot, with no separate sound step and no separate upscale. The smallest useful thing you could do with it today: pick MiniMax H3 in the Hailuo web console, set fifteen seconds at two-K, and add sound words to the end of your prompt the way you added light words, such as wind howling or gravel crunching. Then listen to what comes back and ask whether the crunch lands on the frame where the foot lands.

Kling AI has done the same thing on a different schedule. Its VIDEO three-point-zero Turbo, updated through the middle of June twenty twenty-six, has native synchronised audio, clips up to fifteen seconds at ten-eighty or two-K, and it reads camera paths directly out of your wording. It has also scheduled its older models, versions one through two-point-one, for full retirement on the fifteenth of September twenty twenty-six, which matters if you have prompts you rely on there. The smallest next action: switch the model selector to three-point-zero, turn the audio toggle on, and write your camera move as an explicit direction, such as a slow push-in or a low-angle tracking shot, rather than hoping.

We used Runway's Gen-4.5 for text-to-video. Its Gen-4 Turbo workflow starts from an image instead. If you try that next, choose the starting frame deliberately: it supplies the composition you were describing in words today.

Every one of those tests is the same loop you just ran. Small brief, one change, and look hard at what actually came out.