drltaidev

AI is now used to estimate the times until completion.

Yeah I'm a bit cheeky here, but it's not even fully wrong probably! I definitely do this when working on my solo dev stuff, but well, there it's much lower impact (and I feel like Claude sometimes estimates too high, not realising that itself is going to be doing all the work so it'll be faster).

For my solo dev work, I'm using Claude Code exclusively, nowadays almost always with Opus 5.5. I'm combining it extensively with Wispr Flow, I can give much more detailed strategic and tactical advice and requirements, trying not to micro-manage Claude because it actually does a great job implementing what I need once it know the requirements. I also chat with it first to set up those requirements first.

I prefer using Claude in terminal, it works fast for me. I used to use the "agents view" to set multiple agents to work, each in their own worktree, if I needed work on multiple features simultaneously. But I've actually gotten into just drafting a large list of features first in advance, using Claude of course, and then setting Claude to work to implement them all (so it takes the results of already implemented features into account when implementing the next).

For organising myself, I let Claude write a lot of Markdown files (so I'm not completely locked in into Claude Code, in case I do decide to use something else). And of course a proper Claude.md.

I also made myself a kind of Vision.md > Features.md > Tasks.md workflow, where I make sure Claude is aware of my bigger vision for a project, and uses that to distill features that should be implemented from that, which then leads to very specific tasks to execute. I can then just tell Claude "execute the next batch of tasks" and review them once they're done. My apps are rather visual, so I also made Claude build pages where I can easily review new functionality, so I don't have to dive into the code where not needed (I would, I am a computer scientist, but honestly I'm doing it less and less nowadays...).

I did a PhD in Machine Learning before, and then worked as an AI Engineer / Data Scientist (still manually doing the programming, but basically only python).

Now I'm realising that I love building apps, also because AI makes it so much easier to do... but also devalues it, anyone can do it. So I'm still seeing where my technical background is helpful and where not, though nowadays app building seems way more about marketing than about building. Still, I'm discovering that the building and owning my own projects is just a lot of fun. For now it's not paying enough though, so I'm also taking a job as an AI Engineer / Consultant, might as well keep up to date with AI advancements and their applications in business while it keeps being the big thing.

Yeah you probably had Claude Code make an artifact for you. Which is indeed: only for you.

That would be an interesting development, but makes sense. As humans move towards simpler GUI apps to set their AI to work, AIs start making their own terminal apps to get stuff done (because the primary writer of software is no longer human either...).

I still stick to Claude Code in terminal though, it just seems faster and smoother and I keep some more view of what is roughly going on under the hood, even if it's just a little)

Indeed. If current frontier models are already so disruptive now, we may have vastly better models in a few years but also the current models will have become significantly cheaper. Exponential growth (or shrinkage) is definitely at play here, and should not be underestimated.

In a way I wish this would become true, and to some extent I believe it will (excitement will go down, and some part of the bubble will burst). But I can't fully see it happening this way, these tools are just so disruptive in so many different areas... perhaps not all, but it can speed up the development of almost anything so I don't see it disappearing. I hope it'll help improving its own efficiency as well so we can still get rid of some of the datacenters (or at least stop building more of them), but I don't see it going away as "just another tool", it really sits on top of other tools I think.

What makes this hard to predict is that it's a question about human behaviour as much as about technology... if we do it well, I don't think AI should take everyone's job but it should certainly make our jobs more fun and give us more free time. But when did humanity really do things well...

On the other hand, jobs are about money in the end, who spends money? Humans, not AI (I mean, AI does but still ultimately because humans wanted it to). Products & services sell because humans want or value them. Since the technology/building side becomes so cheap (especially in software), who will actually earn money is determined by who can produce stuff/services that humans actually want to pay for (not what they actually need, but what they think they need/want). I'd say that should keep a large human component in decision making, at least on a strategic level. On a tactical and detailed implementation level, yeah AI is gonna be doing most of the work.

"polarization, explained"

I like the idea, I definitely think there is potential in apps helping with teaching music, of course this isn't anything new, but the direct feedback possibilities have improved a lot with recent ML / AI advancements. As others said, singing is so much more than just hitting the right notes of course, so understanding the scope of the tool is important. But still, it's a neat idea. It's awesome that AI coding tools have made it so much easier to work out these ideas... but also tricky because everyone seems to make their own now, and it's not even per se the best ones that survive.

I've thought about these kinds of apps a lot recently, and am also working on some (not for singing though), I have a programming background (and a PhD in Machine Learning) but still I am impressed by what Claude Code can do so I'm hardly writing a single line of code myself any more. Asking for the right things is still important of course.