18:01Dan Kieft · Tutorial
My AI Avatar Clone is So Realistic It Replaced Me
An 18-minute walkthrough of building a photorealistic AI video clone from a selfie, an audio clip, and a structured timeline prompt.
May 19thThe argument in one line.
AI video quality is decided before you hit generate — five fixable prompting habits separate creators who get cinematic results from those who burn credits on randomized output.
Read if. Skip if.
READ IF YOU ARE…
- You have been generating AI videos but the results feel generic or inconsistent and you cannot pinpoint why.
- You copy prompts from ChatGPT or Claude and paste them directly into video generators without editing them.
- You want a repeatable prompting system rather than starting from scratch every generation.
- You are choosing between Seedance, Kling, and Veo and want a practical breakdown of what each model does well.
SKIP IF…
- You are an experienced AI video creator who already uses structured prompt frameworks and reads model-specific documentation.
- You are looking for advanced post-production techniques beyond basic cuts and upscaling.
The full version, fast.
Most AI video failures happen before generation starts. The host walks through five root causes: prompting without a concept, blindly copy-pasting LLM output, having no prompt structure, picking the wrong model, and treating the first generation as final. The fix for each follows a clear system: a five-field concept worksheet, a Claude storyboarding template, four named prompt structures (simple, Seedance, timeline, JSON), a model comparison benchmark, and a post-generation checklist that includes multi-take culling, deliberate cuts, separate sound design, and optional upscaling.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 00:27
01 · Cold open — the slot machine hook
AI-generated slot machine footage, host introduces the five-mistakes premise.
00:28 – 01:47
02 · Sponsor + setup
Higgsfield sponsor read, video roadmap overview, five mistakes teased.
01:47 – 04:24
03 · Mistake 1 — Prompting before thinking
Bad prompt demo (teenager at arcade), concept worksheet introduced with five fields.
04:24 – 06:50
04 · Storyboarding with Claude and GPT Image 2
MD template uploaded to Claude, six-panel storyboard generated, converted to Seedance timeline prompt, final generation shown.
06:50 – 08:45
05 · Mistake 2 — The copy-paste trap
ChatGPT verbose prompt demo, model skips actions, headphones change every shot. Fix: back-and-forth dialogue with LLM as adviser.
08:45 – 14:10
06 · Mistake 3 — Pros use systems (four structures)
Simple structure, Seedance structure, timeline prompting, and JSON prompting all shown with live demos and a detailed Claude back-and-forth refinement session.
14:10 – 16:20
07 · Mistake 4 — Using the wrong model
Kling 3.0 vs. Seedance 2.0 vs. Veo 3 compared across realism, action, dialogue, emotion, cinematics, and text-on-screen. Veo called outdated for action.
16:20 – 20:16
08 · Mistake 5 — Not going the extra mile
Multi-take culling, deliberate cuts in Premiere, separate sound design, upscaling with Higgsfield or Topaz. Final CTA to Skool community.
Lines worth screenshotting.
- Bad input equals bad output — AI video models are good enough to make movies, but most creators are prompting like they are pulling a slot machine lever.
- Visualize before you generate: storyboarding in Claude and rendering with GPT Image 2 saves more credits than any other single habit.
- LLM-generated prompts are verbose, unoriginal, and confuse video models — the model skips half the requested actions because there are too many.
- A three-field prompt (style plus action plus camera) beats a two-hundred-word ChatGPT block for single-action shots every time.
- The Seedance structure was built from the official documentation — reading tool docs and feeding them to an LLM produces a custom prompting guide for any model.
- Timeline prompting assigns one action per time segment; most AI tools support up to 15 seconds, which is enough to tell a complete micro-story with precise control.
- Veo is currently outdated for action scenes; Kling and Seedance dominate cinematics — knowing this before generating saves significant credits.
- Every generation costs money — treating it like a gamble is the real expense, not the per-second pricing.
- Generating the same clip multiple times and cutting the best sections together is standard practice among experienced AI video creators.
- Do not trust native AI audio — generate sound separately using dedicated tools or stock libraries for reliable results.
- Skipping vague words like cinematic and replacing them with specific references (shot on Arri Alexa, Blade Runner 2049 color grade) produces more consistent model responses.
- The LLM works best as a strategic adviser that asks clarifying questions, not as a one-shot prompt writer.
Five decisions that happen before you hit generate.
WHAT TO LEARN
The quality gap between generic AI video and cinematic AI video is almost entirely a prompting problem — and every failure traces back to one of five skipped steps.
- Define a concept before touching the generator: a five-field worksheet (subject, action, environment, style, audio) forces the visual clarity that vague text prompts outsource to the model.
- Storyboard first using an LLM and an image generator — visualizing scenes before generation catches bad concepts early and saves credits on shots you would have thrown away anyway.
- LLM-generated prompts copied directly into video models produce verbose, inconsistent output; treat the LLM as a back-and-forth adviser that refines your idea, not a one-shot writer.
- Consistent results require a consistent structure — the Seedance five-field format, timeline prompting, or JSON prompting each solve a different generation problem and are worth learning as distinct tools.
- Read the official documentation for every video model you use and feed it to an LLM to extract a custom prompting guide; models respond better to language that matches their own training vocabulary.
- Specific visual references (shot on Arri Alexa, Blade Runner 2049 color grade) give models an internal reference point that vague words like cinematic do not.
- Model selection matters as much as prompt quality — knowing that Veo lags on action while Kling and Seedance lead prevents wasting credits on the wrong tool for the job.
- The generation is not finished when the clip appears: multi-take culling, deliberate cuts, separately produced audio, and optional upscaling are what turn a passable clip into a polished one.
Terms worth knowing.
- Seedance structure
- A five-part prompt format (Subject, Action, Environment, Style, Audio) derived from Seedance official documentation, designed to give video models exactly the context they need without redundancy.
- Timeline prompting
- A prompt technique that breaks a generation into time-stamped segments and assigns one action per segment, giving the model a precise shot-by-shot plan.
- JSON prompting
- A structured prompt format written as JSON with model-recognized keys that makes instructions machine-readable and easier for LLMs to help refine iteratively.
- Character sheet
- A reference image set showing a character from multiple angles (front, side, back) used to lock consistent appearance across multiple video generations.
- Concept worksheet
- A five-field planning document (Subject, Action, Environment, Style, Audio) completed before writing any prompt, ensuring the creator knows what they want to see before spending credits.
- Higgsfield
- An AI video platform that hosts multiple models including Seedance 2.0, Kling 3.0, and Cinema Studio 3.5 under one interface, along with upscaling and editing tools.
Things they pointed at.
05:10toolGPT Image 2
11:00productSeedance 2.0
14:10productKling 3.0
14:10productVeo 3
Lines you could clip.
00:23
“Bad input equals bad output. It's as simple as that.”
08:12
“Just because the prompt is long doesn't mean the prompt's actually useful.”
13:05
“Each time you hit that generate button, essentially, you're spending money.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
metaphoranalogy
00:00Your AI videos look fake, and it's not just the model's fault. It's these five mistake that set apart the pros from the average Joe's. And once you see them, you cannot unsee them.
00:12AI video models today are good enough to make entire movies, but most creators are stuck . And I've been saying this for a long time, but bad input equals bad output. It's as simple as that.
00:28So that's why in today's video, we will cover five of the most common mistakes people make when it comes to making AI videos, and I will show you how to fix them. Make sure to watch the entire video because even if you're an expert, I'm sure you will find something that will be useful to you. I'll also lay out all the assets, all the prompt files, all the information you need in the school community.
00:49So if you want to go through this at your own pace, then check out the links in the description. This video was sponsored by Hicksfield, and we will be using them throughout this video because we're using various different models. If you wanna try them out yourself, I will leave a link in the description.
01:03K. Let's start with mistake number one. You started prompting before you started thinking.
01:08Let that sink in because that's what a lot of people are doing right now. They just type something in. They have no idea what they actually wanna see.
01:16They just say something completely random, then they run out of credits, and then they start complaining online about how expensive AI is. I mean, I get it about how expensive it can be, but if you're prompting incorrectly, then it's just a massive waste of money. Before you write a single prompt, you should have two things.
01:33First, your concept, and second, your storyboards. I will show you how to make that in a second. But first, let me give you an example of what a bad prompt is.
01:41This is what I call a bad prompt that a slot machine type of user generates. So it's just without any thinking. It's just like, oh, let me do a prompt about a teenager in a neon windbreaker skateboarding past a retro arcade at sunset.
01:57You might think it's it's good enough, but then you get something like this.
02:16At first glance, this might look like a good video, although it's still very boring. And that's because we have let AI do all of the guessing work because we didn't quite know what we want to make. We didn't quite know what kind of camera angle.
02:28We didn't quite know what type of specific action. We didn't quite know what the surrounding should look like. So the AI is making all of that up itself.
02:37And if that's how you prompt, that's fine, but it will cost you a lot of credits. The better way to prompt is using a concept. And for that, I got this worksheet for you that you can copy, you can take screenshots, or you can find it in this free school community.
02:52So I'm using this worksheet quite a lot. It is about adding in a subject, who exactly is the main focus. In this case, we're going with a teenager in a bright neon jacket and vintage sneakers.
03:04Then the action. What exactly are they doing? So here, they're describing that he's skateboarding past an old eighties mall.
03:11Then the environment. What is special in the background? Like, what can we see?
03:15So we're going with nineteen eighties sunset street with parked cars. Then the style. Is the lighting, the mood, or the visual tone?
03:23This is a extreme close-up, a tracking shot, handheld camcorder, fisheye lens, golden hour lighting. This also goes in terms of, like, the camera style. So, lastly, this is optional, but you can do the audio.
03:34For example, with C Dance, it natively generates audio with that. So I wanna give it a few instructions of what I want it to sound like. So I'm going with skateboard wheels on the pavement, very detailed.
03:45Now, if we combine all of our sentences, we have a prompt. We have put a bit more thought and effort. It basically comes down to putting in effort into writing something.
03:54So then you get something like this.
04:08So this already looks a lot better. We have a clear visual understanding of what we want to see, and we have a bit more of a structure and a concept.
04:17So that prompt generated this one shot. But what if you wanna generate an entire scene? That's where storyboarding comes in.
04:24So this next method, storyboarding, helps you to develop your concept. For example, I've created this character using Soul Cinema. Looks quite cool.
04:33Right? We're rocking the 80 vibes right here. Then I created a character sheet using g t image two that we're gonna use as a reference.
04:41So now the next step is to use AI to get a better understanding of what type of concept we wanna make. What I did is I went over to Claude. I uploaded the character sheets, then I also uploaded this MD file that I specifically made for you guys that you can use and just copy into your text.
04:59You don't have to use this as a skill. You can just copy it right here as a, uh, file that you add into your chat. It's designed to be as easy as possible.
05:07What it does is you can literally enter in your ideas. So I said, create a scene where this guy enters a eighties arcades on a skate park by jumping through the window. He does a three point landing then says something cool.
05:18Make it feel like an action movie entrance. So now it is gonna make a six panel storyboard for us. It's basically describing, like, everything that we wanna have happen.
05:28And now we can take this prompt so you can copy this entire prompt and put this into GPT image two. So now I'm going over to GPT image two. I've uploaded this entire prompt in there, then I've also added in the character reference sheet again.
05:42And then make sure you put this on high, and then you generate your image. This gets you a sketch like this. So now it has visualized a bit of our storyboard.
05:51It has visualized our concept. It has made the concept look cool and it's made it into a reference that we can use. You can do this for all of the different scenes that you have to create like a short movie.
06:02This is also the best way to save credits because this way, you visualize first, you basically storyboard first. And once you have everything in images, you can then start generating. That is how the pros are doing it where they visualize their concept first.
06:16That's how you get the result that you wanna see, and that's how you're more in control. What you can then do is you can literally ask Claude to help you convert this six panel storyboard into a fifteen second CDense two point o timeline prompt. I will explain more about different prompt structures later, but here we're asking it to provide time codes and one action per shot, character and style block at the top.
06:39So then it gave us this prompt. Of course, you always wanna go through it and make sure it's correct and to your liking. Then you can put this together with your references in CDNs two or in Cinema Studio 3.5.
06:50So here we have the prompt. You add in the two references and then you can hit generate. That will get you something like this.
07:11Game over. The second mistake most beginners make is using the copy and paste trap. And by that, I mean, literally copy and pasting from Chechippity or Cloth.
07:22But I wanna give you a few more things to consider before you actually start up straight up copying from, like, Cloth or ChatGeePete. A lot of these AI tools can handle more context, and most people think the more context, the better, which is right in some kind of case. But you don't wanna overwhelm it with a bunch of unnecessary context.
07:41So you gotta be careful about the context that you put in. And oftentimes, when you prompt something like this in Chatchip T where you ask it to write you a highly descriptive fifteen seconds action movie prompt for heavy duty foam headphones, you will get a large sum of text.
07:58It's giving you a lot of words you wouldn't use yourself. It is adding in unnecessary detail.
08:03It is making unoriginal ideas. And that's why I want to explain to you that you should not blindly trust these LLMs to make the best prompt for you. You should still use it as a tool, and it's only as good as the inputs that you give it.
08:16So, for example, if I put this massive prompt into an AI video model, I will get this result.
08:37If we then compare the outputs together with the prompt, then we notice that the AI has skipped a lot of actions we asked for, like the boat, the helicopter, and the too many details confuse the AI. Like, the headphones are changing in every single shot.
08:51Just because the prompt is long doesn't mean the prompt's actually useful. There's a better way to do this. So here's how I would do it instead, and I'm not banning LLMs completely.
09:00Not at all. The main differentiator that I would give you is make sure also when you're writing your prompt, you give it an image reference. So here again, I'm using the image reference So we at least give the AI a reference that it can hold on to, that we know it needs to keep that consistent.
09:15And then I'm using a prompting structure or a format. More on that later, but for example, this is the format that we used in the beginning at the first mistakes, which is the subject plus the action plus the environment plus the style plus the audio, and then we have our finished results. So I actually did the thinking myself here.
09:34Now don't worry if you have no idea still how to prompt in that format. I will teach you in a second. That brings us to mistake number three.
09:41The pros don't write prompts. They use systems. This is the single and most biggest time saver that I've come across when generating any type of video that I do.
09:52I tend to use four different systems, and I will share all of them with you. So in terms of systems, I'm also talking about structures, prompt formats. That's what I mean by systems, And not having to write prompts from scratch every time.
10:05The first structure is the simple structure. This is one that I use quite often. It is style plus action plus camera.
10:13To give you an example of how I do this, I first use a character sheet. So that's the first step that I use when I have a character that I want to put into a video. Then I prompt it as follows.
10:24So I start with the style, which is gonna be gritty 35 millimeter film. Then the action. This goes kinda like hand in hand with the subject.
10:33So I have a man in a leather jacket, and the action is that he's walking through a rain soaked alley. Then for the camera, I'm asking for a slow tracking shot from behind. Super simple prompt.
10:44It's just, like, three or maybe four if we count subject, uh, different elements that go into it. When we hit generate, you will get a result like this one.
11:00So use this for short single action scenes. It's fast to write and easy to generate. The second structure, I call that the c dance structure.
11:09This is one of my favorite because I use c dance quite a lot, and it's very easy to review and reiterate change. For this, we're using subject plus action plus environment plus style plus audio.
11:22I built this based on the official CDENSE documentation, and that is the tip that I wanna give you. Each tool, regardless of when you're watching this, there might already have been a new tool being released that is better than Cdance currently is, but each tool has their own, like, guidelines and documentation. You can go and review those documentations.
11:40You can literally put them into Chetchip Tea or Cloth, and you can have it spit out some kind of, like, directions, instructions that you can use for each and every generation. That's how easy it is. Either way, the CDN structure works generally well across multiple different AI video generators that have audio natively integrated into them.
11:58So for the subject, again, who or what is in the shot? Then for the action, what exactly are they doing? We can do with one action only.
12:05Then for the environment, where is the scene? And here, you wanna be specific. Then for the style, this is, I would say, the hardest part where you really need to do some thinking.
12:14You can ask any type of LLM to help you with this, but you can go from the camera shot type, lens, camera movements, then the visual stone. You can skip any vague words like cinematic. You need to use specific references.
12:26Like, for example, you can point out that on which camera you want it to be shot on. For example, an Arri Alexa, or you can do specific colors, like, let's do a late runner twenty forty nine color grade. The AI responds well to that because it has visual references it.
12:42It has understanding of what the image looks like from that. It has references of what the color from that looks like. So being specific there matters.
12:50And I know it takes out the guesswork by doing this, and it may be not as fun because you're not having a gamble if your video is gonna be good or bad. But doing it this way will save you credits, and it will also allow you to think more seriously about what you actually wanna create. Because you need to think about this.
13:06Each time you hit that generate button, essentially, you're spending money. And when I'm spending money, I'd like to be not reckless with it. Like, maybe in some cases, but the majority of the time, I wanna know what I'm doing and I wanna be in control, and that's what I'm trying to teach you here.
13:20For this style, I've built this structure that you can use. So this CDN structure helps you come up with these prompts. And to give you an example, I started off with a short prompt.
13:30So I did, like, pretty thirty five millimeter man in leather jacket walks through a rain soaked alley, slow tracking shots from behind. It told me, like, hey, good start, but not good enough. Who is this guy?
13:40Can you give more details? Can you do this, that, that? Um, so then I gave it a bit more details.
13:44Then I really thought about, like, what do I actually wanna see? What do I wanna have happen? And all of that.
13:48So now it identified that the subject is clear, the action is clean, but the environment is a bit too vague. Like, with C Dance, you can use more. So then it asked us a few questions about that.
13:58What is the world looking like that he's walking through? I still, like, sometimes refuse to think about it for quite a bit, so I just say something really standard. Then it will tell me, like, hey.
14:09We need a bit more context. So it's a bit of fighting between you and LLM to give it a context that it wants. And then it says, like, okay.
14:16Now we finally got it. Here is your prompt. Again, this is different than how you use the bad example from the LLM where you just simply copy and paste it.
14:25This is you going back and forth, back and forth. And the that helps you come up with a good concept that you've thought about yourself. It just asks the right questions.
14:37So, if you put that example prompt into the cinema studio, then we get something like this.
14:49Again, this is not a hard example, but it does follow-up with the type of world that we're building, the location, the character, and all of that. The next structure that I use quite a lot is timeline prompting, and you can make this as simple or as complex as you want to.
15:04What timeline prompting essentially is is basically giving the AI a structure of what happens at each second or at each moment. So, for example, you can do so per second or per two or per three seconds. You're basically describing what you want to see during that segment.
15:20For example, with most AI tools nowadays, you can generate up to fifteen seconds. In the future, that will definitely be longer, but that allows us to be more specific what we wanna see during each second of our video. So, for example, you can have the first three seconds where you have a wide shot of an alley entrance, then you have the third to the eight seconds where a man is walking into the frame, then from the eight to the twelve seconds, we have a close-up on his face as he stops, and then from the twelve to the fifteen seconds, he looks over his shoulder.
15:47This is really simple. I tend to use more complex prompts where I also include things the action, the style, and all of that, but this is just to explain you what this structure does.
16:05Now the last structure that I wanna share with you, which I don't use that often anymore, but still is being used by some people quite a lot, which is called JSON prompting. This is, I would say, the more technical version of time line prompting. I I think I prefer time line prompting more.
16:21But this is an advanced structure that is using JSON, which is easily understood by LLMs and easily understood by AI models. So this helps you to essentially structure your prompt in such a way just like a time line prompt or just like that advanced time line prompt that I shared with you to add in all kind of different details and to separate each and all of them so it's easy to read for the AI and easy to understand.
16:44If you are still confused where to get started, try all four of them. I've provided a MD file for each of them that you can use. You can put that into ChachiPeet or Cloth, and it will help you write a prompt based on your ID slash concept.
16:58Mistake number four that I see a lot of people making, which, to be honest pisses me off, is using the wrong model. Every experienced creator right now knows what model is best for what. That is maybe also why you watch my videos.
17:11That's why you watch other people about this topic. Um, you wanna understand what a model like Cling is good for. You wanna understand what Cdance is good for.
17:18You wanna understand what the newest Google model is good for. If you don't know that, you might be using the wrong model because, for example, if you're making an action scene and you're using something like Vio 3.1, currently, it's outdated and a model like Cdance is better. But for some generations, a model like Cling might be better than Cdance at other parts.
17:37To give you a better understanding which tool you should use, I've put together this test where you can go through it yourself. It is a test where I compare realism, action, dialogue, emotion, cinematics, and text on screen.
17:50I've put in the prompt so you can go through it yourself at your own pace, and you can compare which one you think looks best. In my opinion, for most of these things, you wanna be using Cdance or Cling. Vio is kinda outdated right now, but is yet to be replaced soon, so that's gonna change up the whole test.
18:09As soon as this, I will start and update it so you can go through it again and see which one you prefer. Mistake number five is that you don't go the extra mile. Most creators are done when their generation is finished, and that is exactly where you can go wrong.
18:22The first thing that I would say that you should try out is generating the same clip multiple different times and then cutting out the best bits of each separate clip and then stitching that together. If you, for example, have a timeline and the beginning of your video sucks, you can just trim out that bit, regenerate that, and then stitch it together with the rest of the video, and then make it look seamless or have a cut in there.
18:47It doesn't matter. I just use that technique a lot of times. The second suggestion I wanna give you is make cuts.
18:52Don't just generate, like, five second shots, ten second shots, and just place them together like that. You wanna add cuts in there. You wanna trim off the bad pieces.
19:01You wanna be more in control about what you're showing on the screen. Then we have sound design. Don't trust the AI to generate the AI sounds for you that come with the clip.
19:12Generate them yourself separately using other AI tools or use stock library AI sounds. Lastly, a super powerful but very time consuming thing that you can do is upscaling your video. You can do this through Higgs Field itself through Topaz Labs, and it is helping you to add in some grain, to add in some texture, um, but it will also upscale the quality of your video.
19:32You wanna be careful with it though, because sometimes it's a bit too much. So if you go over your own process right now and think about all the mistakes that I've named in this video, you can analyze and see where it went wrong, and you can try implementing the fixes that I gave you in this video. If you need more info, I will have all of the MD files, all the prompting structures, and more in-depth guides in my free school community, which you can join at any time.
19:58You can leave at any time. It's just I'm just trying to build a place where you can hang out and ask your questions, be seen, and communicate with other AI enthusiasts. Also, make sure to check out Higgs Field.
20:09And if you wanna see more interesting videos, then click the video that's on the screen right now where I show you how you can master CDENS.
The Hook
The bait, then the rug-pull.
The slot machine metaphor lands in the first sentence: most AI video creators are not directing, they are gambling. The hook positions bad output as a prompting failure, not a model failure — a reframe that converts viewer frustration into fixable technique.
Named ideas worth stealing.
02:45list
Five-Field Concept Worksheet
- Subject
- Action
- Environment
- Style
- Audio
A pre-prompting planning sheet that forces the creator to define every visual element before writing a single word in a video generator.
Steal forAny AI video workflow as a required first step before opening the generator
11:00list
Seedance Structure
- Subject (who or what)
- Action (one action only)
- Environment (specific)
- Style (camera type, lens, movement, then visual tone with specific references)
- Audio
Built from the official Seedance documentation. Core tip: read docs for every AI tool and feed them to Claude to extract a custom prompting guide.
Steal forAny video generator with native audio or needing precise visual style control
14:55model
Timeline Prompting
- 0-3s: wide shot
- 3-8s: character enters
- 8-12s: close-up
- 12-15s: reaction
Breaks a 15-second generation into per-second or per-segment shots with one action each.
Steal forAny multi-beat scene that needs precise narrative control
16:05model
JSON Prompting
Advanced structured prompt format using JSON with keys like model, aspect_ratio, duration, and prompt. Easier for LLMs to parse and refine iteratively.
Steal forPower users who want machine-readable prompts they can version-control
How they asked for the click.
VERBAL ASK
19:10link
“Join my FREE community where you can find my prompts and chat about AI”
Soft close — community pitched as a resource hub, not a paid upsell. Skool link displayed. Sponsor (Higgsfield) also prompted at end.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
AFFILIATECommission earned if you click.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
