Opus 5.5 is good at explainer videos
launchvideo.io
launchvideo.io
Made entirely with Opus 5.5 + $3.21 of OpenRouter API usageClaude Code Workflow
https://www.reddit.com/r/ClaudeAI/comments/1wogab3/
Jaw literally dropped. I ran the prompt from the "Made entirely with Opus 5.5" post on my own project. Here's what Claude Code made on its own for about $4.Claude Code Workflow
Fable was a huge leap in terms of model persistence and raw intelligence, but it still had terrible taste for human writing and code architecture. It would constantly keep making decisions which would achieve the desired objective (and make the code correct), but would bite you n years from now, and n years from now isn't RLVR checkable.
Opus 5.5 has a very different "feel" than anything else I've seen in this generation, though GPT-6 does seem to be moving in a similar direction. They have finally solved the writing part, and architectural taste also seems to have improved significantly.
I did a review of some GPT 6 Sol's code with Opus 5.5 yesterday, and it went "the code is correct, but there's a bunch of things here that could be simplified, and the split of responsibilities doesn't follow your established architectural layers" (which was true and exactly what I've noticed myself when reading the diff). I don't think I've ever seen a model do this before and actually be on-point.
Does the js stuff render real-time?
Did you get script/dialog and voice settings (again, for making other films in same series/theme) - or just the rendered audio?
I don't care how "good" the graphics become - it will always just be slop. This isn't the part of our lives we should be trying to replace with technology.
I mean, ever since AI was a research program, it was assumed it will some day achieve its goal -- in particular, make computers be creative.
I rather live in a world where society decides the nature of technological change, for better or worse, in a democratic fashion than the current iteration we have which is mostly about lighting trillions of dollars on fire while never providing basic needs for civilians (free school lunches, medicare for all, universal childcare, free community college) only to be force fed on what to use by a group of people that do not care about me or my community that always seem to espouse anti-human beliefs.
It's a sick anti-human ideology and it's not shocking when you look at how anti-human tech leaders truly are.
Right now, looking like an AI did it (even if it was actually a human) is a sign of being un-professional. To a limited degree, one can prompt an AI better and get something that doesn't look as cliché as the default settings, but even then there's often someone who can spot a tell. "You can fool all of the people some of the time, some of the people all of the time, but not all of the people all of the time" applies here.
Art is at least two different things, for at least two different groups: nice to look at etc. is one of them; the human equivalent of a peacock's tail (i.e. the effort is the point and cheating is worse than having nothing) is the other.
Making frozen pizza doesn't get close to eating pizza at an authentic Italian restaurant, but frozen pizza is still pretty good. So if someone says "introducing frozen pizzas will enable anyone to eat a good pizza at home, no need to go to a restaurant" then it's a half truth. The true restaurant experience will still have its role, but the frozen product is still pretty good compared to not having that option at all. But your friends will not be at awe at your culinary skills if you make frozen pizza. But they'll still likely eat it and appreciate it when hungry and chilling at your house.
Who are they gonna sell those to, though, if everyone is unemployed because they're not needed?
Exactly who, do you think, will remain employed, and what will they do?
How far away are we from “build a startup that generates $1M MRR, my openrouter key is in my .env”?
What is it that makes it "better" versus simply asking for an animated video? Is it the ability to later do some adjustments programmatically?
If it's unusually skilled with one tool, but it's not the default tool for a job, then you have to put it in their hands before they reach for something else.
The Opus 5.5 javascript-art + art direction definitely seems to be one of these surprise capabilities jumps. Maybe strong enough that providers will start to nudge in that direction in the system prompt, so that the user request doesn't need to specify it.
It was good enough that our CEO and marketing team are in a room trying different ideas with it after I showed it to them an hour ago.
I did one for our company and showed it to the CEO and now I am talking him out of just yeeting it directly onto social media.
The one shot video was brilliant and polished. It would need minor refinement before presenting to a customer, because it used some of that internal vocabulary Claude makes up to describe parts of a system it's working on.
However: I found the explainer astoundingly useful to me. Now, finally, I understand some of that internal vocabulary Claude's been burying me with. The visuals really helped round out my understanding of what it's been talking about.
I love LLMs in research, programming, email cleanup—the list goes on. I want the word guessing program to give me lessons I can test and verify. But I’ll ask humans to give me opinions on unanswerable questions. I don’t want a confident dead bot telling me about life. FO. I guess it was only a matter of time before someone created something with LLM that rankles. /rant
Instead we've worked out how to get the LLM to make videos.
Neither gives any guarantee. Both static text and videos can be done well and badly. But I usually find it useful to have both. A video is like being guided through their vision of what the thing is (a "push" message). This can be annoying to many nerd types who want to cut through all that and just want to "pull" the info they need, for themselves, at their own fast pace, quickly identifying it and not being spoonfed. But the typical user is not like this, and prefers the spoonfed/edutained approach.
Yes that's the trick. And much more. If you have any sort of content, you should express it as a slideshow, a white paper, a blog post, a video, a podcast, etc. Get it out in as many forms as you can. Maybe even performative dance or song if you've got that energy.
Needless to say, usually if my choices are video or nothing, I'll choose nothing and figure it out myself.
I love using notebookLM for orientation stuff.
It's particularly good for those frustrating recipe videos where they don't tell you quantities of ingredients to use - you can have Gemini hallucinate the quantities and it normally gets them about right.
Plus having the video file means I can extract frames as images at specific timestamps.
Homely though the ask feature is probably fit for purpose, at least on YouTube.
Video that cover different parts of an outline are a great use of videos. Videos that try to cover too much all in one are a lesser use. Yes, you can use scrub to find in a video, but showing a text outline and letting users click to reveal videos for specifics topics works very well.
If you don't care so much about adding something significant to the world, and your LLM-buttons-wrapper is a novelty, chug along and see where it goes if you're having fun. Looking for what value mean at this moment is a much harder effort, and probably a luck-based endeavour.
I don't have fun creating LLM-wrappers or at least haven't thought of a fun idea for it that could be even a fun novelty to work on so personally I'm not doing it but I'm using LLMs for other fun stuff that required much more of my free time before.
Very few things can't be reduced to insignificance. MacOS was just cribbing PARC, Facebook is just a glorified PHP forum, Dropbox is just SFTP, etc. etc.
Even in deep-tech, Zipline is just wrapping from deeper-tech (batteries motors etc), GLP-1s were a VA throwaway that dusted off, etc. etc.
LLM wrapper doesn't mean anything, it's an implementation detail. Besides being an LLM wrapper what is a given thing?
The beer app wasn't just a beer app, it was the intersection of the first time accelerometers were doing something that the average consumer could interact with in their pocket, the first time there was something to spend money on for your phone besides a wallpaper, a ton of things.
If anything now when something seems trivial or like a toy, but it has a lot of traction, I want to spend energy figuring out why is it more meaningful than it appears.
I don't get what exactly you are talking about as a reply to my comment, what exercise exactly? Deciding if something is worth spending time if you're having fun with it?
> If anything now when something seems trivial or like a toy, but it has a lot of traction, I want to spend energy figuring out why is it more meaningful than it appears.
Sure, if you're having fun (and in fun I include a sense of accomplishment, curiosity, whatever tickles you) with that, go for it.
It's a strangely arrogant reply to my comment, perhaps you read something into it that I didn't say and replied to that?
Also I honestly thought we were two arrogant people conversing in good faith: you don't find unarrogant people saying things like "If you don't care so much about adding something significant to the world..."
I think text distorts what I meant with that, I personally don't care much about adding something significant to the world, not anymore at least.
I found my peace by pursuing what interests me personally even though I know it might not be of much significance to anyone else, and that I won't be pursuing the path of "adding something significant" just for the prestige of it in my life.
In a sense it's the opposite of arrogance since I care much more that the person is enjoying the process of whatever they are doing rather than the results of it. I've seen too many bright friends end up miserable chasing it, I've felt too miserable myself when putting too much focus on that, I'd rather have people enjoying doing what they are interested at if it's not something harmful.
- Video devlogs were long a shitty way to make a community (I can explain this one in more detail, but in short, the tradeoff between opportunity cost and video quality is a Pareto curve that is bad at all points on the curve, if your goals include both “make game” and “build community for players with devlogs”, the way out is to drop one of those two requirements, either “make YouTube/Twitch/TikTok channel” replaces “make game”, or you build some other audience for your devlogs)
- Good devlogs don’t have that level of detail, to let people easily recreate games
- Most games can’t easily be replicated by LLMs, in short, the people who are good at steering LLMs like that are making their own games.
This is not the first time I heard this story. Sometimes it comes with the lamentation “back in the day people could build communities with devlogs” and no, that was generally not a good way to build communities. It was mostly good streams from YouTubers who were making the game in order to make YouTube content or it was mediocre streams from random devs. Throw in a few people who are famous game devs who choose to stream and get a big audience because they already have a community.
The whole problem with LLMs is that you don't need the detail.
An LLM "recreation" doesn't have to be complete or very good. It just has to steal enough thunder to be profitable.
It's it possible for an LLM-created game to be profitable but "unsuccessful?"
For instance, if (on average) if it costs $100 in tokens and time to make a crappy LLM-game clone, but you can clear $200 in sales per game on average, you're ahead.
And the economics of slop mean there are a lot of people in 3rd world countries who will make all that effort for a $100 payout (or less).
My point is they don't actually have to do a good job, they need to make a pile of shit that looks just good enough to trick a few people into buying it.
Scenario one: they’re competing with my game and stealing my thunder. In this scenario my game wasn’t very good to begin with.
Scenario two: we’re going after different customers. They get the customers who want a cheaper game, I’m getting the customers who want a better game.
And someone trying to vibe copy another person's game is definitely the type who has no idea of value to add. So many of the slop games the whole concepts are so generic that they definitely just asked chatgpt for everything.
In the modding scenes for some games I play I've seen vibe coded mods where the gameplay additions make no sense and have no sense of balance or fun, with these completely new to the community devs having ko-fi links set up from the start.
AFAICT this is a fear that some game developers have, that someone will steal their game and run with it, and the stories told are more repeated based on this shared fear than based on realistic scenarios. People repeat it because it’s a good story.
And yes, I’m aware of some situations like 0x10c and the like.
The idea is not valuable, ideas are a dime a dozen. The actual value is in your implementation.
There's no moat in an idea either so I guess it's down to marketing budget.
It requires discernment, taste, on the part of the user to be able to point the LLM at the weird parts that look wrong.
Even with that, the LLM can only fix most of, not all of, those rough edges.
More specifically to AI... A blog on a nice website is no longer a signal for quality so there's just no point in devs investing in that.
It really makes me appreciate how important good game design is. Claude is doing a fine job coding everything I describe, but it doesn't really understand fun, so I need to.
(It's not a great game)
And that's why we can't have good things...
(This is just gonna keep happening more and more until eventually we'll need something like a patent system for ideas)
So, a patent?
Some SaaS products have network effects, but that doesn't appy to new software.
Starting a SaaS business in 2026 seems a bit silly to me. Not saying people won't make money, but you'd have to be pretty lucky.
The LLM only does average returns from average inputs.
I've been playing with an LLM backed reference checker for my partner who is a lecturer. For each reference in some work she's marking an agent is dispatched to read the referenced source and check that the reference is correct / accurate / not hallucinated.
It's useful to her, but it feels like it would be a waste of effort to turn into a product, because pasting the same document into Claude/ChatGPT with a prompt like "download and read each reference etc" would work pretty much just as well.
I had similar concerns about an AI backed training generation company I interviewed with. I asked how they saw themselves competing with increasingly capable generic harnesses. I don't recall exactly what they said but the impression I got was they were so focused on competing with the big LMS players that they hadn't really considered it.
Also for creatives, some sort of aesthetic control? Most of the examples on this page look like templates you’d see in an office suite that appear swish, but are ultimately bland.
I built https://gallotails.com/ because I am a cocktail enthusiast. First commit, March 6th 2019. One of the features I am proud is which cocktails you can build based on your ingredients, and I was smart, the bar ignores garnishes and can suggest cocktails you're one ingredient away. But you have to add the ingredients. And interpret the bar screen.
And here's what I did with ChatGPT a few days ago: https://chatgpt.com/share/6ab59877-4b74-83e8-8da5-966d9f3a38...
Clicked the microphone icon and rambled for a few seconds (could have just taken a picture too I guess). Nudged ChatGPT to only give me simple cocktails I can stir instead of using a mixer. Said I could buy a lime, sure. Then asked how to keep the ginger beer fresh.
There is NO WAY I can code all of that in my little gallotails.com - a single chat window has entirely replaced my site. And that's my pet little project I do it on weekends, mostly to keep my tech skills sharp. I can only imagine what the biggest cocktails website, actual companies, are going / will go through once more and more people just realize the chat tab is enough. And I am not talking about purely content, ChatGPT can do everything my site does, and more, into any direction, in an instant.
Which left me with the question, then what should my site be? Community? To be taken over by agents?
Currently munching on what sites like mine should do, and what they should be.
Oh yeah, and I am not mourning or anything like that. And I do want the human touch, because AI can't. But humans, at least over the internet, can send their agents, or companies can run agents and I won't know from my site if it's an actual human, or some AI fabricated interaction (can't even trust an image it uploads).
When AI first showed up a few years ago, I thought about opening an actual physical cocktail bar. Hypothesizing humans will be tired after a long day of their agents talking to other agents, they would crave for actual humans. I would ban electronics inside - no distractions. Maybe one day I will go for that...
This is not sustainable in the long term, but irrational longer than solvent yadda yadda.
Chatgpt would shit on my great grandma's waffle recipe but it's the one I'll keep making till the day I die.
Eg from your share, the model output “I prefer it Scotch-heavy rather than 50/50 because otherwise it gets very sweet.”
It’s basically cribbed this from someone; it literally can’t taste.
The frontier labs are not yet able to do rlvr over mixology/meatspace. So maybe think on how to capitalize on that.
My software development process feels exactly like every debugging session in the holodeck that seemed terribly unrealistic to me.
I'm in SRE so systems reasoning is something of the value. But me writing code or even configuring stuff? That's dead. It is a total waste of time - Claude can do it better, faster and it works.
Then it’s sufficiently non-obvious. I believe there’s still a pretty big field of things that fall into this category. Making proper things is a lot more than writing some code.
Nope. If all your doing is prompting an agent to build stuff (especially if what it is building is based on LLMs), then you might as well take it back to the whiteboard and rethink how you're spending your time/tokens.
> That's why I created isitsaas.io: your platform for determining whether your llm wrapper has product potential, or whether people would rather just use claude directly instead. Simply pop in your business idea, and our award winning, proprietary technology will ask claude if you can make money off of it.
> Trial memberships start at $10/month
Start from thinking about product value add, and user behaviors first.
You can also host a chat interface that connects to said MCP server as a convenience. But serious users probably already have their own inference and it's already hooked into other resources as well.
Back in the days, when I made websites or home pages for clients, my first question was: "What's your message you want to tell with your website?" Sometimes the client couldn't answer this, because he simply wanted a website for the sake of having a website. But a website without a message is meaningless more or less.
That said, there was an entire industry out there cranking out these """fun""" explainer videos, so there must be some market for it out there.
Not anymore, looks like.
When I visit a website, I'm usually looking for information and not for a message.
A good consultant / contractor understands the implicit requirements and is able to ask the right questions that the client is able to answer instead or planting traps, gotchas, and acting like a mean teacher examining a naughty student.
Yes, likely the client wanted a website because it was trendy and competitors were starting to have websites and he didn't want to be left behind. Now, that you sneered enough at this business wanting to stay in business, you can actually advise them. Maybe they don't have a message. Maybe they never had to think about things that way, because they never made this type of branding decisions and operated in the old offline word, based on word of mouth, old types of ads, phone calling people whatnot.
And a website can be quite meaningful without branding messaging. For example a valid message is: this is our address, this is our phone number, here is our team, our product descriptions and so on. Or for that type of business: the opening hours, etc.
I think this is part of the reason users like working with AI better, because it actually helps them get to the thing they actually want, instead of gatekeeping and testing whether the client is worthy and deserving enough to have a website made for them.
meh, I don't care about the message of your business, I just want the ADDRESS and OPENING HOURS. Why is that so hard to put up front and center and instead I've got to scroll past videos and hero images and chatbots and dickovers and embedded interactive maps without the actual address showing argh </rant>
sorry that website was probably not your fault
Some of the stuff I've seen is mindblowingly impressive in how they've captured quality creative decisions, and how the models can reason about visually appealing designs.
It understands things like continuity, themes, facial expressions, cultural memes/references, symbolism, etc. And it generates things using a lot of multi-modal behavior by using 3d, images, etc.
Some really incredibly impressive stuff.
The drawing/animation style feels suitably off from xkcd - more naive stick figure, than what xkcd looks like?
Some speed/timing issues with the walk cycles.
For all that - mind-blowing that it's generated so quickly.
Say more about the process and / or drop some link pointers.
Then start Claude. From that point on, it’s mostly prompting and evaluating the results. Like any good engineer, you shouldn’t simply start with “Create a video from this comic.” Instead, take a step-by-step approach: discuss the visual style, animations, ask for audio examples, create a script, and then develop a storyboard (ask for a .html document).
The important part is to break the whole generation process down into small, manageable steps. I also used the brainstorming skill from Superpowers for the discussions. It makes the whole process much easier when the AI asks the questions and I just have to answer them.
The result is 4500 line python code generated by Opus 5.5 and a lot of video and audio files.
> After all, I undertook to tell several trillion years of human history in the space of a short story and I leave it to you as to how well I succeeded. I also undertook another task, but I won't tell you what that was lest l spoil the story for you.
I say Bummer if more people adopt this style, its gonna become less and less authentic feeling as the months go on now.
But: I think as LLMs fully generate more and more complicated things, we’ll start discovering ways to make them generate with good user control.
Now that said, it is wild that in just a few years we've got AI making videos like this!
I'm running Arch Linux and would rather not install DaVinci Resolve, but Blender would be fine for the non-linear video editor.
My plan is to use ChatCut to trim and split into the different sections. After I have the five segments, what would you suggest for swapping the background (environment) and attire? There are two simultaneous camera shots (front and side) that I'd like to stitch together as well.
Any suggestions? (Paying someone a couple of hundred $CAD to take this task off my plate would also work.)
One should be skeptical of the demos they initially see at model release, as there have been instances of some being called out as AI generated video or work that took days and millions of tokens and not a one shot as claimed.
The first few days Astra was out, I attempted to reproduce a few of the demos I saw on twitter, and it was clear Astra had a distinctive style and certain limitations when working in short sessions with tools like Blender that could distinguish genuine demos from bs for retweets.
People putting out these demos should share their sessions to really show what was going on, and I invite people to test models and try to replicate what they see and draw their own conclusions.
- Voiceovers: aligns animations to the voiceover, can generate voiceover with elevenlabs, or will transcribe and timestamp a real voiceover
- can reorder scenes both in code, and using ffmpeg for audio.
- interactive controls during authoring, can ask for micro edits or re-builds
- MP4 export/encoder
- Generates the video with agent of your choice (obviously)
Fast forward a few months, and Opus 5.5 is now running my entire video production pipeline. This model is leaps and bounds ahead of the previous model.
I can generate a complete 5-minute explainer video on my MacBook Air M4 in about five minutes from a single prompt, using the Gemini TTS API and MiniMax H3 Max for b-roll. It creates thumbnails and titles, handles the entire process, and automatically uploads the result via the YouTube API—all powered by a Claude Code plugin. (youtube.com/@ctrlaltexplain for those interested).
I asked it to make a site for my TI4 reference project: https://axiomvortix.com
It produced a video about an AI startup with the vague goal of centralizing data.
I would really like to take a playwright driven script, then use a similar pipeline to create explainer videos. Then I can add my own voiceover. It will really reduce production time but keep it flexible so I can tweak the script but not need to record over and over.
Does something like this exist? Would be a great open source project.
Above all, the model’s sense of what feels and what looks good is the most important. So many times the models have missed clearly wrong layouts etc. I guess it just shows the limitations of LLMs when they should generate something they have not been trained on?
If opus 5.5 is now better capable of that, that would be brilliant!
Few times I asked Claude to edit a photo it says it can't, or it produced soemthing awful by running some python library.
I don't get it. Can Claude do any image gen at all? Does it just delegate tasks to other tools to make the video or can it actually produce video?
They will do well provided tools and ways to verify the work and iterating on that.
All these impressive things are by making these tools available by default, rather than having to know about them and integrating them manually with skills and mpcs.
https://gzvxcspoxhhgoeog.public.blob.vercel-storage.com/vide...
Instead of paying designers and editors the big bucks, enterprises can use such models to be self-reliant.
Good explainer videos need a human teacher to craft the narrative and exposition.
But made a local fork that works just as well with 'claude-code -p' and a local copy of ffmpeg
No way to download or watch it in the webapp of opencomputer. No instructions on what to do? Do I need to run it locally? It seems it did use usage.
I wish more people optimized for learning than "how fancy can i make it look"
The code for the infamous "I'm upping my p(doom)" video (https://www.youtube.com/watch?v=8j-hR4fJywU) is open-source and uses Processing: https://github.com/JohnHeibel/PDoomVideo
About 6 months ago I spent 12 hours creating a video for our Edge.js [1] announcement that, in the end... people were not very inspired by (to say the least). See the video here [2].
Then, about a month ago, I spent 6 hours creating a video for our Wasmer SDK announcement. It was better, but still didn't go as viral as I wanted [3]. I always thought we would need a big budget to do them.
But then, yesterday I tried Opus 5.5... and man, I'm impressed. The video was one-shotted with this prompt:
Make a modern slick and punchy video for this announcement:
[content of the blogpost in markdown]
Exactly how? What does ffmpeg record?
For 80% AI does the work in seconds and 20% is manual work in minutes to get exactly what they want.
Kurzgesagt makes some of the best explainer videos, and they spend most of their time writing the script, not making the animation: https://youtu.be/uFk0mgljtns?si=NCMxIYGUYY-BbQgB&t=75
If you're the author congrats, great work!
(not that they were great before, but ...)
I am not super sure what makes it so good, but it seems like it's a good option to try if you have access and want to see how the frontier models are doing.
Maybe this is why Apple chose that family to help train Siri AI. (Siri AI is not a fine-tuned Gemini)
So predictable.
Why does opus get all the credit?