How App Development Is Changing in the Age of AI
From an era when experts build for everyone to one where anyone can build and adapt software for themselves
How App Development Is Changing in the Age of AI
From an era when experts built for everyone to one where anyone can build and adapt software for themselves
On March 22, I hosted a YouTube livestream that ran for about three hours.
After the livestream ended, I uploaded two summary videos based on that session to my channel.
One was in Korean.
The other was in English.
You can see the results here first:
English summary video
Korean summary video
This was the second time I applied this workflow: capturing a long livestream in real time, organizing the output, and turning it into summary videos immediately after the broadcast.
It worked the first time.
It worked again this time.
And after seeing it succeed twice, I started thinking less about whether the technology works and more about what this kind of workflow suggests about the future of software, content creation, and app development itself.
Here is what I covered in the March 22 livestream
What makes these summary videos interesting is that they are not just technical demos.
The livestream itself included a wide range of real experiments, observations, and field notes from my week, including:
attending GitHub Copilot DevDays at the Bellevue GitHub office
reflections from a Seattle Korean IT community seminar on AI and PKM
the launch event for K-Initiative, organized with the Korean Consulate in Seattle and KSC
the official launch of Gobi Space
updates on VibeLearn AI v2.0 and continuous vibe learning
experiments in multi-topic learning
and even a very practical, everyday experiment involving AI and lawn care
So these videos are not only about “AI made a summary video.”
They are also records of how I am actually using AI in daily life, learning, workflow design, and content creation.
That is why I recommend watching at least one of the summary videos first. Even in a short format, they contain a surprising amount of context about where these experiments are going.
What stood out again was the power of real-time visual summarization
After using this workflow for the second time, one thing became even clearer to me:
real-time visual summarization during a live broadcast is powerful.
As I speak during the livestream, Gobi Desktop captures my voice, and within the AI4PKM workflow, the content is structured visually through Canvas.
That visual layer is useful because it lets me see the flow of the conversation, the major themes, and the relationships between ideas while the livestream is still happening.
At the same time, the post-livestream video production was possible because the workflow also generated a full transcript and a summary document in real time.
In other words:
real-time visual summarization helped me understand the live conversation as it unfolded, and the transcript plus summary documents made it possible to move quickly into video production right after the livestream ended.
That combination mattered a lot.
Without this workflow, making a summary video right after a 3-hour livestream would have been much heavier work
In the older way of doing things, I would have needed to go back through the long video, identify the key moments, build an outline, create slides, write or revise narration, and then edit the final video.
That is not just time-consuming. It also demands a lot of energy.
Right after a livestream, your attention is already drained. Turning a three-hour session into a second piece of content is possible, but it usually feels heavy.
This time, though, the full transcript and summary documents were already there by the time the livestream ended.
That made it possible to move directly into the next stage.
Of course, AI did not finish everything exactly the way I wanted in one shot.
I first generated a slides file from those materials, and then I revised parts of the narration myself so that the tone and wording felt closer to what I wanted.
I think this is an important point.
AI still does not usually deliver a perfect final result in one pass. But there is a huge difference between doing everything from scratch yourself and having AI produce a structured first draft that you then refine.
The second approach is faster, lighter, and often leads to better results.
That is how I was able to publish both the Korean and English summary videos so quickly after the livestream.
I am not using AI4PKM skills exactly as they are
There is another important detail here.
I am using the skills provided by AI4PKM, but I am not using them exactly as-is.
I am modifying them to fit my own workflow and needs.
For example, the original video-making skill was designed so that generated audio, image, and video files were created in a single folder using the same names, which meant they kept overwriting one another.
That structure did not fit my situation.
For one livestream, I needed to:
create a Korean version
create an English version
and keep all of those outputs for future reference
So I changed the skill to fit how I work.
This is not just a story about editing code.
It is also the starting point of the main insight I want to share in this post.
In the age of Vibe Coding, the meaning of general-purpose software may be changing
In the past, coding was more clearly the domain of specialists.
Because of that, apps and tools were usually expected to have a certain level of generality from the beginning. They needed to work for many kinds of users across many environments with minimal modification.
That expectation shaped the whole development process. It required more research, more planning, more design, and more coordination.
Users, in most cases, were simply users. Developers built, published, and maintained. Everyone else used what had been created.
But I think that assumption is starting to shift.
In the age of AI—and especially in the age of Vibe Coding—more people can now build, modify, or extend software for themselves.
And if that continues, then not every app needs to begin as a fully generalized product.
Someone can build and publish an app that mainly expresses an idea.
Then the people who download it can adapt it to their own environment, their own workflow, and their own needs.
That is the shift I keep thinking about:
from a world where experts built complete tools for everyone, to a world where anyone can build an idea into software and anyone else can adapt it for themselves.
This is exactly what happened in my case.
At first, I thought about the old pattern: clean up my changes, open a PR, and try to get them merged into the original repository.
But soon I realized that what I had added was not really a universal feature. It was something I needed.
I needed bilingual output.
I needed persistent storage of the generated assets.
I needed the workflow to match my own publishing process.
To merge that back into the original project properly, I would need to step back and redesign it for general use: think through settings, edge cases, compatibility, and how other users might want it to behave.
That is real work. Valuable work. But also expensive work.
And when people can increasingly use AI to add what they need on their own, the pressure to generalize every useful improvement may become smaller than it used to be.
That is why I think the software development process itself may be changing.
This does not mean general-purpose software is going away
I am not saying that enterprise-grade software, shared platforms, or broadly useful products no longer matter.
They absolutely do.
There are many domains where stability, security, consistency, compliance, and maintainability are essential. In those areas, general-purpose software remains critical, and it will remain critical.
What I am talking about is a different and growing space:
personal workflows, creative pipelines, one-person systems, small-team tooling, experimental utilities, and highly specific use cases.
In those areas, it may no longer make sense to spend so much time turning every good idea into a polished, fully generalized product before publishing it.
Instead, a new pattern may become more common:
build the idea, publish it, and let users extend it for themselves.
AI is making that pattern more realistic every day.
That is why I want to keep running these experiments
For about a year, I ran a weekly livestream series called Vibe Coding for Fun.
That series taught me a lot.
Now, with AI in Action — everyday experiments in applying AI to real life, I feel like I am entering the next phase.
I am no longer just asking how to learn AI.
I am asking:
how AI can be connected to everyday life
what kinds of workflows become possible when humans and AI work together
how our way of working changes in practice
and what role users begin to play when they are no longer just passive users
The March 22 livestream, and the Korean and English summary videos that followed right after it, gave me a very concrete example of that change.
And what makes it more meaningful to me is that this was not a one-time success.
This workflow has now worked twice.
That is why I wanted to write this.
If you have a few minutes, I hope you will watch one of the summary videos first:
I plan to keep running these small weekly experiments, and I will continue sharing the insights they produce here on Substack.
P.S. Original 3-hour livestreams