7 Prompts You Need to Create Better AI Videos (ASAP)

0:00 Over the past few months, I've generated hundreds of AI videos, and I can confidently say that the difference between a clip that falls flat and one that feels cinematic comes down to the prompt. And if you keep writing prompts the way most people do, you'll get outputs like these that fall flat no

0:16 matter which model you're using. So, in this video, I'm going to give you seven prompt techniques that cover every major type of AI video shot that you'll ever need so that your outputs feel professional. The tool we're using for everything today is called Higgs Field. It's an all-in-one platform that handles

0:31 both the image generation for our character and location and the video generation for all seven clips. So, inside Higgs Field, we're heading into Cinema Studio. Make sure you're in image mode and select GPT Image 2 as your model. That's the one that gives the cleanest output for character

0:47 generation. Now, I'm going to upload a photo of myself here. What this does is give the model a face reference so that the character it generates actually looks like me. Then, I'll paste in the character prompt and generate. What comes back is still me, but now styled exactly the way the prompt described. It

1:03 looks like a proper editorial shot with clean studio lighting and sharp focus on the face. This is our character, and she's going to be in every single video we generate. Now, for the location, we stay inside Cinema Studio, but switch the model from GPT Image 2 over to Cinematic Locations, specifically the 2K

1:21 option. This model is built specifically for environmental scenes, and the quality difference compared to a standard image model is immediately obvious. Paste in the location prompt and generate. What we get back is this golden hour rooftop with warm amber light and a city skyline glowing in the

1:37 background. This is the perfect backdrop for everything we're about to make. After it generates, click save location so you can reference it directly inside your video prompts. And the reason we went through all of that is so that every single clip we're about to make uses the exact same face and the exact

1:53 same rooftop. That way, the only thing that changes between each video is the prompt itself, which is exactly what we need to actually see what these techniques do. Now, let's get into our first prompt. If you want to follow along with me, I've left a link for Higgs Field in the description below.

2:08 Upload your character, and once the image is uploaded, it will instantly check for eligibility. Sometimes it fails, but just try two or three more times and it'll go through. And after that, you can reference it directly in your prompt. As you can see down here, our location is already saved from

2:23 earlier, so it's ready to go. No need to upload it again. That's the base setup we'll use for most of the generations today. Now, here's the first prompt. back is a smooth deliberate dolly in that pushes toward the character over 4 seconds, then moves into a medium close-up with a subtle handheld drift

2:39 and a rack [music] focus. That's a complete two-shot cinematic sequence from a single prompt. And this is actually a good moment to talk about something that most people never think about when they sit down to write a prompt. AI video models like SeaDance were trained on enormous amounts of

2:54 visual data, which means that they don't just understand what things look like. They understand the language that filmmakers use to describe them. Phrases like dolly in, rack focus, medium close-up are the precise technical instructions that the model was trained to execute. Most people write prompts

3:11 the way they'd type a Google search, something like woman on a rooftop, camera moves toward her. And the model gives them exactly that, a vague flat clip with no real intention behind it. It did what it was told, but it was told almost nothing. The gap between those two outputs isn't the model, it's the

3:27 language. When you use the actual vocabulary of filmmaking, you're not just describing a scene, you're directing one. And the model has a precise brief [music] to execute rather than a vague idea to guess at. That's why that dolly in feels deliberate and that rack focus lands exactly where it

3:42 should. SeaDance didn't get creative on its own. It did exactly what the prompt asked for. And by the way, every prompt I use today is in the description below, so you can copy them directly. Now, this next technique is about something that happens before the camera even starts rolling, and it changes the entire

3:59 emotional tone of the clip. Think about the last time you watched a scene in a film that genuinely made you feel something. Chances are the emotion wasn't something that built up slowly over the whole clip. It hit you almost immediately within the first few seconds. That's not an accident.

4:13 Directors establish the emotional tone of a scene before anything else happens, and that's exactly what this technique does inside a prompt. The first line of what you write sets the tone for everything that follows. When you open with a specific mood, the model aligns every visual decision to serve that

4:29 feeling. The lighting, the motion, the pacing, all of it bends toward the emotion you named first. But when you open with an action instead, you get movement that feels mechanical. The model builds the scene around what's happening physically rather than how it should actually feel, and the output

4:45 ends up looking stiff and disconnected. This prompt opens with three words, raw, broken grief. Everything after that is just the physical expression of those emotions. That way, the model isn't guessing at the vibe. It's been told exactly what this scene feels like before it's been told anything else. And

5:02 every visual choice it makes from that point is in service of those three words. Even though it's the exact same setting as the first clip, this is a completely different scene. And without a single word of explanation, you can feel what's happening just from watching it. So, next time you're generating any

5:17 scene with emotional weight, try writing the feeling before you write the action. The shift in output is immediate. Emotions set the tone of a scene, but the next technique focuses on how you word the action. This one is simple in theory, but it makes a massive difference in practice. The verbs you

5:33 choose in your prompt directly determine the quality of motion in your video. Weak verbs like walks, looks, stands, and moves give the model almost nothing to work with. They're so broad that the output ends up vague and low energy because the model has no information about the force, the speed, or the

5:49 intention behind the movement. But specific physical verbs like sprints, slams, whips, and heaves communicate all of that. They tell the model not just what the body is doing, but how it's doing it and what's motivating it. To show you the difference, here's what a weak version of this scene would look

6:04 like as a prompt. She runs to the railing and grabs it. The wind blows her coat. She turns around. That's technically the same scene, but compare it to what actually happens when you replace every one of those verbs with something specific and physical. She doesn't run, she sprints. She doesn't

6:19 grab the railing, she hurls herself against it. Every single verb is doing real work, and the output reflects every bit of it. You can feel the impact of her hitting that railing. That's the model responding directly to the physical language in the prompt. Go back through your old prompts and look at

6:35 your verbs. That's usually where the energy is being lost. The next element is very important, and most of the time people underestimate it or ignore it completely. [music] Lighting is one of the most powerful tools in filmmaking, and it's also one of the most underused elements in AI video prompting. Most

6:51 people either don't mention it at all or they write something vague like cinematic lighting, which tells the model almost nothing. What actually works is treating light the same way a cinematographer would. You name the source, describe its color, and specify its quality, if it's hard or soft, warm

7:07 or cool, where it's coming from, and how it affects the subject. When you answer those questions in your prompt, the model can build a scene where the lighting is doing real emotional and visual work rather than just filling the frame. This prompt uses two light sources that work against each other on

7:22 purpose, lighting from the sun and a soft light from the sky. That contrast is a deliberate choice. Warm and cool tones pulling in opposite directions on the same face creates visual tension and dimension. It makes the image feel three-dimensional in a way that flat single-source lighting never does. Every

7:39 light element is named, described, and given a specific job to do. The difference in visual quality here is immediately obvious. That contrast between the warm and cool tones on her hair isn't something that happened by accident. It's exactly what the prompt asked for. So, from now on, you should

7:55 treat lighting as a character in your prompt. Name it, describe it, and give it something specific to do. Now, the next technique focuses on details so small that most people would never think to include them. There's a specific quality that separates an AI video that feels rendered from an AI video that

8:11 looks realistic, and it almost always comes down to one or two tiny physical details that most people would never think to put in a prompt. I call these sensory anchors, which are basically micro details that ground the scene in physical reality. It's not a major action or a dramatic moment. It's a

8:27 loose strand of hair lifting in the breeze or the last of the sun glinting off a metal railing. These details are small, but they're the things that make a viewer's brain register a scene as real because they're relatable to how people or the environment actually act. The reason they work is actually pretty

8:43 straightforward. The model has been trained on real footage where these kinds of details exist naturally. When you name them explicitly, you're essentially telling the model to operate at that level of realism rather than at the level of a clean polished render. This prompt has no dramatic action, no

8:59 emotional arc, and no camera movement, just a series of small precise physical observations layered on top of each other. And the result is quiet and still, but it doesn't feel empty. The way the breeze moves through her hair, it feels like a real moment rather than something that's generated. One or two

9:15 of these details in any prompt will immediately change the texture of what comes back. That technique works beautifully for short clips, but the next one is specifically for when you need longer videos without having them fall apart. Here's something worth knowing about how AI video models handle

9:30 longer durations. When you give a model a 15-second generation without any structural guidance, it tends to fill the time with slow drifting motion that doesn't really go anywhere. It looks fine on the surface, but it doesn't feel like a scene. It feels like footage. And the difference between the two is

9:46 narrative. A scene has a beginning, a middle, and an end. To do this, you need beat structure within your prompt, which means that you divide the clip into distinct moments and describe what happens in each one, a build, [music] a turn, and a release. The model then has a pacing map to follow rather than an

10:02 open-ended brief to fill, and the output has actual shape to it. SeaDance 2.0 is particularly good at this because it was built for longer clips and can handle multi-shot generation from a single prompt, which means you can get a complete three-beat sequence without having to generate each shot separately.

10:18 This prompt breaks 15 seconds into three roughly equal beats. The first is calm, the second is action-packed, and the third [music] resolution. That way you get a complete emotional arc in 15 seconds. Make sure you set your duration to 15 seconds before generating this one.

10:49 This looks like a clip from a short film. This way gets you a complete emotional arc. That's entirely because the prompt gave the model a structure to follow rather than just a scene to fill. For any clip over 10 seconds, beat structure is something I now consider non-negotiable. We've covered six

11:05 techniques now, but none of them have touched on what happens when your character actually needs to speak. And that last one changes everything about how dialogue lands in a generation. When you want your character to talk, writing the line in the prompt is the bare minimum. But this way you'll get back a

11:20 character speaking in a monotone, unrealistic way with no real emotion. I waited. The delivery is flat, the timing is off, and it ends up looking like a text-to-speech demo rather than an actual performance. The reason that happens is that a line of dialogue on its own gives the model almost no

11:44 information about how it should be delivered. Is it said quietly or loudly? [music] Does the voice break? Is there a pause before the last word? All of those choices are what give a line its emotional meaning, and without them, the model just generates speech in the most neutral way it can. Delivery cues are

12:00 how you fix that. They're descriptions of tone, volume, physical reactions, and timing that you place into the prompt around the line itself. They tell the model not just what the character says, but what the character feels while saying it. I waited. I really did. Wait. And the difference in output is

12:22 significant. This prompt gives the character one line, but it also picks up on the quiet steadiness in her voice, the way it catches just before the final word. That's a complete emotional performance brief, and what comes back executes it. Every single thing she does to express her emotions isn't just the

12:38 model improvising. If the prompt had just said she says I waited, the output would have been completely flat. Delivery cues are the difference between a character reciting a line and a character actually feeling it. These seven prompt techniques will elevate your outputs no matter which model you

12:53 use, and they'll help you generate results that truly feel professional. And with Higgsfield, you can run all of these under a single subscription. So, if you want to start generating your own cinematic AI videos using these [music] exact techniques, I've left a link for Higgsfield in the description below.

13:09 Thanks for watching, and I'll see you in the next one.