Create Image + Video Prompts from Lyrics with Muzak
Muzak turns lyrics or reference images into detailed, sequenced AI prompts for music videos through simple commands, replacing dozens of pages of manual prompt-writing.
Writing a full prompt sheet for an AI-generated music video by hand can run twenty or thirty pages once you account for sequencing, style, and shot-to-shot continuity. This video introduces Muzak, a command-based tool built to compress that work into a single instruction, turning lyrics, images, or both into ready-to-use image and video prompts.
What Muzak is built to do
Given lyrics, reference images, or both, Muzak generates the initial image prompts for a music video and then the sequenced video prompts that follow from those images. Like every GPT-style tool built this way, it's backed by real research, in this case research into how professionals actually sequence music videos, condensed into five to ten pages of rules behind each command so the user doesn't have to write that detail out manually every time.
Commands instead of manual prompt-writing
Muzak is entirely command-based. A command like "dance" pulls in the tool's researched rules on scripting, sequencing, and music video structure to produce a full set of prompts. Commands can be combined with style modifiers, dance bluegrass, hip-hop, folk, for example, to shift the tone toward a specific genre feel. Users can start from lyrics alone, or upload reference images of the look they're going for and let the tool build around that starting point.
Why sequencing matters
The important property of Muzak's output is that the prompts are sequenced, not generated in isolation. When the goal is to tell a story, image one needs to connect to image two, which needs to connect to image three, and so on. Because the output is entirely text, the results are legible even before generating anything: you can read the prompts, get a sense of where the sequence is heading, and decide whether to run with it, regenerate it, or hand it to another tool entirely.
First drafts, not final drafts
Like every creative AI workflow, Muzak's prompts are explicitly framed as starting points rather than finished work. They get you close, and from there the normal creative loop takes over: refine, refine, and refine again until the result matches what you actually want. The tool is also actively evolving, with new commands and functions added over time based on what does and doesn't work in practice, while existing functionality is generally kept rather than removed.
Key takeaways
- Muzak converts lyrics, images, or both into sequenced image and video prompts for AI-generated music videos.
- Each command, like "dance," draws on five to ten pages of researched rules on music video sequencing and structure.
- Style modifiers can be layered onto a base command, such as combining "dance" with "bluegrass" or "hip-hop."
- Sequenced prompts keep a visual story connected from one image to the next, rather than producing disconnected results.
- Muzak's output is a strong first draft meant to be refined repeatedly, not a finished prompt.
- The tool is continuously expanded with new commands based on real use, while existing functions are rarely removed.
Try it yourself
Muzak is available at humanitarians.ai/muzak for anyone looking to speed up ideation on AI-generated music videos, whether working in a vertical short-form style or landscape format. It won't hand you a perfect prompt, but it will get you dozens of visual directions to explore in the time it would otherwise take to write one detailed prompt sheet by hand.
Full transcript(auto-generated, with timestamps)
[0:00]Okay, bear here. Um, just going to talk about this little music tool. What musicac tool helps you do is you have lyrics and images or just lyrics or just images. It'll help you create the prompts for both the initial images for a music video and then the video prompts. the the prompts once you have a um set of images a video prompt for those images like all GPTs that I create basically what we do is we do a lot of research on how to do that and so there's a lot of data here each one of these is five 10 pages of rules of how
[0:42]To for example sequence for a music video based on doing research on how professionals sequence for music videos. All this is a lot of research. So basically what these tools do, what this tool does, I'll go back to the tool here. This there's always a help command here. And what help will do is just explain how to use it. And so what this is my explanation here, it's very simple. It just it helps you do your initial prompts as with everything in AI. These initial prompts are initial prompts. They're rarely your final prompts, but they get you close. But as they get you something which is nice,
[1:29]Which then you refine and refine and refine and refine and refine and refine until you get it the way you want it. Um, it's all command-based like all the other little tools we have. And basically what it does for example is let's say I want to create a dance video from uh on the housetop Christmas song. I just put in the command dance. I can also put in other things like hip-hop version or folk version or bluegrass version or whatever. So I can say dance bluegrass and it'll make it more bluegrassy if the music is bluegrassy. You can also up old images. Typically, I will do that. I'll
[2:13]Take actual images from, you know, what I think I want the look to look like and put it up there. But you don't need, you can start with just your lyrics and go generate images and then go back and forward and back. And you just hit go and it gives you suggestions. The important thing about these suggestions is that they're sequenced. So generally when we want to make a video we want to tell a story which means image one is connected to image two which is connected to image three image 4 etc etc. You can then paste these in your favorite uh image generator and see what they do and then
[2:51]If you don't like them just change them and do something else. But because they're all words you sort of know what it's trying to get to you. In fact, you may not like this at all based on the words in which case you can regenerate it or try it something else or have a tool like GBT or whatever write that. But that's it. This is you know like all the other tools. The basic key here is I do not want to write 30 pages of prompts to do something but I want the detail that writing 30 pages of prompts will give me. And so each command command
[3:27]Like dance uses all of this information at least the relevant information for that particular command in order to write very very detailed prompts. And with all of the AI sort of garbage in garbage out the the better information you have more detail you have the better your results are. So this is it. I mean these are just convenience tools. You could do this by writing 30 pages of prompts every time you have posters on the inbt or you could just use command like dance and it'll do that for you. So that's it. Um hope it's a useful tool. As with all these tools, we constantly improve them, refine them,
[4:02]Improve them, refine them. For the rest of this video, I'm just going to show sort of side by side because I mostly do Tik Tok style and this will be landscape style. I will just have some videos that we created using this tool just running just so you can see what it looks like. You can also just go to the website and watch them there. Okay, thank you. Okay, here's some examples of things with the help of the tool. Remember that it won't magically give you a perfect prompt. Stab it gives you ideas. Thus, the whole sort of workflow process with AI. The I it's very quick for ideiation.
[4:41]You can try 40 different Popeyes and see which one you like. This is based on some historical images of the resistance to fascism in World War II. This is uh Jingle by elf character. Um and so you can play with things and see, you know, do you like them? This is more experiments in different trying to get dance styles, dance things. So, with all of these tools, what these tools are powerful at is very, very quickly, here's a Night Before Christmas. Basically, the idea behind that is people capturing Santa's um Christmas Eve light on various iPhones around the world. So, you can play with a lot a lot of
[5:29]Things. So, this was an example currently what it looks like. There's Wednesday Adams there. Um, currently what it looks like. We change these things all the time. We add typically add new functions. We don't typically remove functions but add new functions. And we do that by playing with it, seeing if it does what we wanted to do. And if it doesn't, we add some more functionality. But still, all of this has been helped with. It's not created by it's an assistant. It helps you think about how you might want to approach things. So, um, that's mostly it. Uh, this is the Mayfield King song. Great song by
[6:08]The way if you love sort of soul and Christmas music. Um I also work on the lettering. So the lettering for there's Marge my March character. So that's about it. Basically the tool just helps you. It doesn't create for you. It helps you and it helps you with ideation. Helps you give ideas. So that's it. I hope this was take care.
More videos
2:08Bridging the Pixel Gap in Browser Automation.
2:23How One Narrow Safety Rule Can Make an AI Less Safe Everywhere Else.
2:04Why splitting a chunk from its document makes it retrieve for the wrong question
4:20Three You Can Take Back. One You Can't.
2:21Why a 50-turn agent pays for the same screenshot 35 times unless it caches the pixels
1:53