A few weeks ago, I started building a small AI video project.
My assumption was pretty simple: if video generation becomes fast enough, there should be interesting products you can build around it.
So I connected a video generation API, built a basic interface, and got the first version working.
Prompt in. Video out.
At first, that felt like progress.
Then I actually tried using it as a real workflow.
And I realized the API was probably the easiest part of the whole product.
Generating one video is easy
Most AI video demos follow the same pattern:
Prompt → Generate → Wait → Video
That's fine when you're testing a model.
But imagine you want to create short-form content regularly.
Now the workflow becomes something like:
Idea → Prompt → Generate → Wait → Review → Download → Organize → Upload → Publish
Then you do it again.
And again.
If you want to create 10 videos, you don't really have a video generation problem anymore.
You have a workflow problem.
That distinction wasn't obvious to me when I started building.
Waiting becomes a UX problem
Another thing I underestimated was generation time.
With many normal web applications, a user clicks something and expects an almost immediate response.
AI video doesn't always work like that.
A generation can take long enough that keeping someone on a loading screen feels wrong.
My first version basically did this:
click → wait → result
It worked technically, but it didn't feel like a good product.
I started thinking about each generation as a job instead.
submit → queue → process → result
That small change affects a lot of things.
Now multiple generations can be submitted without forcing the user to wait for each one individually.
It also makes batch generation much easier to think about.
The next problem appears after the video is generated
Once I had generation working, another question appeared:
What happens after the video exists?
If the user still needs to download every video, rename files, open another platform, upload them manually, and repeat the process, I've only automated one part of the job.
This is where my original idea started changing.
Instead of building another page with a prompt box and a Generate button, I became more interested in the entire pipeline:
Create → Queue → Review → Publish
I'm currently testing this idea through a small project called ReelsLaunch.
I'm using it to experiment with short AI video generation, queues, batch workflows, and reducing the manual steps between creating a video and publishing it.
It's still early, and I'm not sure yet which part of that workflow will turn out to be the most important.
But building it has already changed how I think about AI products.
Three things I've learned so far
- Don't build around the model
It's tempting to see a new model and immediately think:
I should build a website for this model.
I did something close to that at the beginning.
But models change very quickly.
Today's impressive model can become a standard feature somewhere else a few months later.
The model is infrastructure.
The more interesting question is what workflow becomes possible because that infrastructure now exists.
- Long-running AI tasks need different UX
A traditional request/response interface isn't always a good fit for video generation.
Users shouldn't have to babysit a generation.
For longer AI tasks, queues, background processing, status tracking, retries, and notifications can become just as important as the generation itself.
This sounds obvious in hindsight.
It wasn't obvious to me when I started.
- Look at what happens before and after the AI
This is probably the biggest lesson for me.
When we build AI products, it's easy to focus almost entirely on the AI step.
But users usually have a larger job they are trying to complete.
For short-form video, generation is only one piece.
There is planning before it.
There is reviewing, organizing, and publishing after it.
The opportunity may not be in making the Generate button better.
It may be in removing the need to click five other buttons afterward.
I'm still figuring this out
I don't know yet whether the best version of this product is primarily a video generator, a publishing tool, or something closer to an automated content pipeline.
That's part of why I'm building it.
But one question has become much more useful to me than asking which AI model I should use:
What repetitive work can disappear because this model exists?
If you're building something on top of an AI API, I'd be interested to hear how you think about this.
Are you building around the model itself, or around the workflow that happens before and after it?
Top comments (1)
Treating each generation as a job is the right call, and it pays off twice: once the queue exists, retries, progress and "email me when it's done" become small features instead of rewrites. One thing worth adding early: store the prompt and settings with every job. In a workflow tool, "make another one like this, but shorter" is the most common request, and it only works if the inputs outlive the output. Are you keeping the videos on your side, or handing them off to the user's storage?