Show the full transcript (3,994 words)
Now, since Opus 5.5 has been released, people have been generating insane videos with this model. We have product launch videos like, for example, this one by Chain on X, where he also creates this little avatar with a voice to introduce this new product that he is promoting with this video. And Opus 5.5 does a great job at animating the UI of this app. Or you have Adrian on X, who is also animating his Figma designs using Opus 5.5. And this one, which is one of my favorite ones by DevTeamDrew on X, which has a tremendous amount of detail, and if you would have told me earlier this year that this was entirely generated by AI, I would call you crazy. Like, this is truly unbelievable. But I also see content on X where people are saying, "This took me 15 minutes to generate," or "This took me just one prompt to generate." And then some people see that and they try it out on their own, and what they get is some sloppy AI video with a centered text and a gradient, and kind of the whole design looks very generic. But the thing is, the prompt is only 10% of the outcome of this video, and the rest of the 90% is the harness.
So in today's video, I want to show you how we can get the best of that 90% and generate stunning videos with Opus 5.5. And by the way, guys, if you like these types of workflows, I do have a free Discord community where we do live co-working sessions every single week. Feel free to join that; link is down below. And also, before we get started, it's important to understand how these models generate videos. Now, Opus 5.5 takes images and text and actually gives you text in return. It can't actually create an MP4. So you start off with a brief where you give Claude a text plus different references. They can be images, they can be links, they can be videos. Claude, using Opus 5.5, then writes code based off of your brief and then chooses one of these four directions. So the first route is it could draw the code out with, for example, SVGs and a headless Chrome session, or it can use a framework like Remotion, HyperFrames, Diffusion Studio. It can be a mixed pipeline with image-video models plus Opus composites on top of that, or you have existing videos, and Opus 5.5 can cut those videos, add captions, add B-roll, etc., etc.
And by the way, if you want to learn more about Opus 5.5 and how it works and a really good deep analysis of Opus 5.5 with video generation, be sure to check out this X article. Big shout-outs to Moves or Movez for for putting this together. This article played a huge role in the research for today's video. But anyways, going back to this model, this is an example video by someone named Tommy Rossi, where he said, "One shot, it put all the code in a single index.html file and rendered it using Playwright frame by frame in a headless window," then FFMPEG to generate the MP4. And he explicitly said the model seems to prefer doing everything with zero dependencies from scratch instead of using tools like Remotion, Igaki, HyperFrames, Diffusion Studio, whatever. So if you don't really mention a framework to Opus 5.5, it's just going to automatically take this first route.
All right, so one of the first things that we want to do here is we want to, obviously, start generating a video. And to do that, you need to open up the Claude Desktop app, create a new folder on your computer, which would be your working directory, choose the Opus 5.5 model, and then choose high effort. And then what I'm going to do is I'm just going to give it a random prompt to generate a 10-second video of a weather app forecast. The sun rises, clouds roll in, it starts raining, and the temperature ticks down. And so we end up getting this video over here, which, to be honest, it's not bad. It's really kind of exactly what I expected. But I gave Codex the exact same prompt, and we get something way less impressive, which was automatically built in HyperFrames. And I actually have HyperFrames installed in Codex and in ClaudeCode, but like Tommy Rossi said in his tweet, ClaudeCode actually defaults to generating the code rather than using a specific framework.
And if you don't have HyperFrames or Remotion connected, what you can do is you can go over here to the connectors and just search for HyperFrames, right? You install that, you authenticate your account, and you can do the same thing with Remotion, install this Claude code plugin. And so let's start from scratch in a new project. I'm actually going to go back to this article and I'm going to copy this text over here, which are the Motion Studio rules, which consists of four different categories. The first one is the render contract, which helps you with sustainable rendering of your video. Then we have the look, which bans certain generic output of your video. And then you have the sound and then you have this feedback loop stage. And by the way, this and everything that you see in this video have it linked down below. You can literally just download it, plug it in, and start generating really nice videos.
So I'm just going to paste this in here, hit submit, and so the next thing that I'm going to do is I'm going to test it out by giving it a one-liner prompt. So there's this video here, generated by Stephan Livera: "Make a dynamic 15-second motion graphics video that shows what an incredible motion designer you are, like it's your showreel for a resume, go all out." So showreel for a resume sets a genre with known rules: fast cuts, a new technique every shot, the best work first. Opus already knows what a reel looks like. "What an incredible motion designer you are" makes the model the subject, so it shows techniques instead of explaining a product. No content to get wrong. "15-second" is short enough to finish in one pass and long enough for six to eight shots. "Go all out" works as an effort multiplier on top of your xHigh or Max. And then this is the result that we get. We get some really interesting movements, a lot of transitions, a lot of different art directions, but all in this black and orange style because it's it's based on Claude, which takes us to the next step, which is actually adding a real brand or another brand, maybe it can be your brand or your client's brand, instead of this generic output.
So for today's use case, I'm going to be using MagicPath.ai, which is this awesome infinite canvas plugin that you can use for Codex and for ClaudeCode. And I'm just going to give it the same prompt as before, but now I'm going to say about MagicPath, and see what result it gives me. And then this is the result that we get. So we have the different steps with showing the interface of MagicPath in this infinite canvas view, and it's showing in different, you know, dark and light mode, everything that it can do. And then we have the final animation to the logo, which for a first step is not bad. It's taking all of that harness from the previous output, but now changing the entire animation and the branding to match MagicPath. Now this marketing video by Tony Dinh is actually very impressive because he says that a year ago he paid around $1,000 to get a video like this, where it gets all of the components from his UI. It does like an entire product explanation, and he basically told Claude to gather the brand assets by itself by saying, "Use actual product screenshots, logo, assets, two: must have music and motion must match the music. Do it like a real professional production video, not like a demo or prototype. Try again, ultracode." And so this article on X that I was referring to before has a great prompt that you can use, where it can do this research for you and then generate the video based off of that research. All you have to do is like replace the placeholders over here.
So we can copy this, and then we can just paste that prompt and replace the placeholders with whatever brand that you're working with, and we get this 20-second clip, also in different formats. But it basically explains the whole product. It shows the different UI components, gets the screenshots from the website, I can tell, and and generate some of the animations from scratch. And then in another session, I actually wanted to test this out, and I used this prompt, and it said, "Make a 15-second motion graphics launch video for MagicPath," and then I gave it these two sentences. And we actually got something like this, which I don't know, it might be a little bit more cleaner. You tell me, but it also looks good, shows everything that needs to be shown. Pretty much a very similar style. But like I showed you before, we can also add audio to these videos, and not just audio, like we can add like an actual person speaking while they're showing the product. And I actually partnered with a great company called Fish Audio who can help you with this, and they're also today's sponsor.
Now, quick detour, because this build needs a voiceover. And most AI voiceovers sound pretty flat and generic. And there are some good voice APIs out there that sound much better, but they charge per character, which is pretty crazy, so the bill kind of climbs up very fast. So I want a real voice that my coding agent can use where I don't really have to worry about the pricing, and that's where Fish Audio comes in. And they have this S2.1 Pro free text-to-speech API with no usage cap all the way through November, which is pretty much the same model as their paid tier and also much, much better when it comes to their competitors, which comes with 83 languages and you can also do voice cloning. So the first thing that you want to do is you want to go to fish.audio, and you want to sign up for an account. And once you're in, this is how it's going to look like. You have all these different products on the left, and you have these tools here in the middle, where you can generate text-to-speech, voice design, create a voice from scratch, etc., etc. But in this case, I'm going to be using the Fish Audio MCP, where you can basically talk to all of these tools inside of ClaudeCode or Codex. So in this page, which I'll link down below, you want to go down here and click on Claude.ai and select and copy this link.
And then in your ClaudeCode settings, you want to scroll down here and click on Connectors, click on Add, Custom Connector. You can add the name Fish Audio, and then paste in that URL. And then once that's connected, what I like to do is I like to head over to Tool permissions and set this to always allow. And then you can write out a simple prompt like, "Hey, use Fish Audio to generate a voiceover for a YouTube intro hook about coding agents." So I can then go ahead and send this, and then we start getting different options of voices. It ends up giving us this one: "Last night, I gave an AI agent one sentence. By morning, it had written the code, run the tests, fix its own bugs..." Now that's pretty good, but let's say that we want to clone my original voice and maybe add a little bit more emotion. So I can say, "Keep the same script, but use my voice (clone from my YouTube channel) and also add some excitement." And the crazy thing is that it generates these things in just a few seconds, and you don't have to pay a single dime. "Last night, I gave an AI agent one sentence. By morning, it had written the code, run the test, fixed its own bugs, and opened a pull request. I didn't touch the keyboard once..." And honestly, the thing that surprised me the most was not only the speed and the price, but the actual quality that you got for this speed and price.
So yeah, guys, try Fish Audio's S2.1 Pro API, I'll have it down below. And so I gave ClaudeCode this prompt template again, just to kind of try out a different variant, and I said to use the Fish Audio MCP, and I also asked it to use my voice. And this is the result that we end up getting. So we have a nice design, showing MagicPath's UI, showing the features with my voice. And now the next step here is to try to give Opus 5.5 references to build on, and I've done this in some of my other videos in my channel, if you want to check that out as well. But there's this great website that I found, it's called whatships.com, and basically this is like a library of different videos generated by Opus 5.5 and now listed on this website. And so you can really pick any of these, and it takes you to that original post on X. And I, for example, really like this one, which was for this company called MTI-OON or M-Tune, and I like the playful style, I like the different colors, I like the fonts, I like the movement, so why not just take this one as a reference?
So we can start by saying, "Hey, this is the reference," download the video, then extract every frame every half of a second, write a style guide markdown file that includes the palette, the type, the shot lengths, the transition types, the camera movements, the texture grain, how text enters and exits, then write a shortlist markdown for a 15-second launch video about MagicPath in that style and use real screenshots, take the grammar of the reference, never its content, logos, or characters. And then what we get is this, right? It's like, literally a MagicPath video in the style of that other reference video that we found on Twitter. But instead of showing like a timeline interface, it shows the actual MagicPath interface, and I think that Opus 5.5 does an amazing job when you give it a reference and when you give it this specific prompt.
And now, when you want to start making things be consistent, just like with what you saw in that previous example, you want Claude to write down these markdown files so that it basically can come back to those and memorize your brand, memorize the style, the the grammar, the the tone of voice. And what's also important, apart from that, is like a spec list. So for example, this can consist of different buttons, loaders, checkmarks, charts. And so in order to set this up, you want to start by giving it several inputs, like what the product is, what it needs from this product. This can be the the logo, the brand colors, the fonts, UI screenshots. But then we also need the direction, which can obviously vary on the type of video that you're trying to generate. In this case, we want something more like this, it's more like animated UI components. So we have Product Film UI Motion, one container never cuts, every state is the same element changing size, radius, and fill while its content swaps behind a short blur. And you can obviously ask ClaudeCode to help you write this out for you by giving it a reference.
Then you want to give it the structure, the build, then we have the gotchas, which are things the model shouldn't do or things that it should keep in mind. And then finally, I wanted to show me the state list on the beat grid before writing code. And what does that mean? Well, we have a beat grid in a table view, where we basically get a blueprint of how this animation would look like frame for frame. And then once satisfied or if not satisfied, you can always tweak a few things, but if you're satisfied, you can then render the video, and you get something like this, which is basically based on that same style that we wanted. Looking very nice, um, you know, showing how MagicPath works, you can select different things, turn it into code, send it to to ClaudeCode for example, make it real, and by the way, this is in that article that I was telling you about earlier if you just want to copy this prompt. But everything that I'm doing in today's video I have in one zip folder as well, and you can just download it, I'll link it down below, and then add it to your ClaudeCode session. So you don't have to go back and forth to this article.
And now there's this other example by Onur Ozcan, and it's basically a video that he generated with Opus 5 that tells the story of Steve Jobs. And if you listen to the video carefully, you can tell that every single movement is in sync with the sound. So we have the music layer, you have the sound effects layer, and then you have the picture layer, which is what you see. And then everything moves in sync every 120 BPM. And so you can write something like score the MagicPath film to the beat, synthesize an original 120 BPM track in code on the same timeline as the picture, measure the track and output beats.json with BPM, beats, downbeats, and hits, synthesize UI sounds in node, a click on every cursor press, a whoosh on every container morph, a thump on the logo, and then retime the animation to this JSON file. And so now in this new video, we get a lot more sound than in the previous version. Every time a text pops in, there's a sound. Every time there's a transition in the picture, there's a sound as well. And I don't know, maybe it's better than the older one, maybe it's not, but I just think it's a really interesting thing to experiment on.
There's actually this example by Vox on X, and he's using a very similar system as we just did. And for this particular video, it actually, I would say, fits. Now, this video example that you hear, I have to say, is one of my favorite use cases of someone generating a video with Opus 5.5. I just discovered it today actually. It's by Donald on X. "I made this with one prompt using Opus 5.5. I spoke to my computer for 5 minutes, Claude worked for 12 hours, and I woke up to this." And you can see his prompt that consists of like thousands of different characters, multiple paragraphs, but it's basically based on this original video over here. And if you read this prompt that he wrote in the comments, you'll kind of understand the skeleton here, and that it consists of these 12 different points, which you can argue can be sort of like a director's brief.
So you have film in one line, you have the references, you have tools and keys, which consists of different different skills to load, APIs available like C-Dance for example, character bible, which is a Bible for each character in this video, so that Bible consists of different perspectives of the character, different proportions, different emotions and expressions of this specific character. So you could have like hundreds of images and docs of this specific character in this Bible, a beat sheet, text on screen, workflow gates, critique loop, which which we're going to cover soon, and deliverables. And Donald's video actually inspired other people to remix this, his initial video that he was remixing, and we get insanely impressive works by other creators like Pleiometric in this like retro anime style. And so in this article by Moves, he gives us this prompt which I find very helpful and very interesting. So you are the director, animator, sound designer, and render engineer for space-holder film made in code. Then you have the film in one line. So he basically goes through all of these different steps that we just that we just saw recently. And so what you can do is you can copy this, and then you can just paste it into ClaudeCode and replace the information for whatever your whatever it is that you're trying to build or reference on.
I used another reference video from before and we got this one, another video from MagicPath. It's actually much cleaner than the other uh, videos that we got generated from before, I'd say. So, as you can see, everything, all the UI is being generated, we have the text floating, we have the agent actually showing how the UI is working in action, and it's like one camera perspective, then we open ClaudeCode terminal, then it starts generating in ClaudeCode, they build right here. So those designs turn into actual code. So as you can see, now that the prompt is longer and has this this structure to it, the actual output is getting much better. And then finally, we have this critique loop, where you want the model to basically analyze what it's done, score the result, then fix those bad scores, and then repeat until it has a specific score or higher.
So this video that I showed you before, you can really tell that it's that that it's, uh, he bought a lot of work into this one, but there were actually 163 model calls and nearly seven hours, obviously not one shot, to get this video right. So again, we can just copy this prompt whenever you get a specific result and you want to re-iterate on that, and so again once you do that, it goes round by round and improves these different scores, and tries to keep everything seven or eight plus. So after this third round, you can see that, you know, this went from 7 to 8, and this one, depth, went from 6 to 8. And then the final part of this video is actually shipping this video, formatting it, or turning it into a skill, or actually turning this into a service. So for example, the guy that made this video, you can actually click on this link, and for around $70 you can generate a video just like this but for your own brand, which I think is a really cool type of service that you can offer in today's age.
But basically, what you want to do is you want to package this as a skill, where it has a specific name. It'll, then, next time you use it, ask for specific information like the URL, or the duration, or the format. And then it goes through this pipeline, and it has to abide by these hard rules. But anyways, guys, that was it for today's video. If you have any comments or questions, please let me know down below. And like always, thank you all so much for watching. Hope to see you next time. Goodbye.
Video
Video
