I think it's possible the amount of new software that will be written for an audience of 1-10 will be greater in 2026 than in any previous year, and then the same again for many years to come. I also think a lot of this software will be essentially 'hidden' - people just writing this stuff for themselves because the cost to say things to an agent is very low compared with the cost of actually planning out a software design and so forth.
Interoperability will probably be important in the next few years and I wonder if this is something solvable at the agent/LLM level (standing instructions like 'typically, use sqlite, use plaintext, use open standards' or whatever). I also think observability and ops will be pretty important - many people who want personal software but don't care for the maintenance and upkeep.
It's so strange to me that since the 1960s with BASIC then later on dozens of https://en.wikipedia.org/wiki/List_of_educational_programmin... including Logo by Feurzeig/Papert/Solomon there is effort to precisely help beginners program software.
The effort was not to onboard future professional software developers but rather to make the personal in personal computer, or PC, meaningful. It's YOUR computer, you can put YOUR software on it. In fact even pocket calculator do that.
We keep on re-discovering the foundations.
I now have tailor made apps with all kinds of bells and whistles that commercial products can’t offer easily ( I fall under non commercial usage which opens a lot of doors ), and that free software might offer, but later.
I have also learnt a lot technically in the process, since I’ve been able to venture into what was for me unknown territory but at controlled cost
I plan to create more such apps in the future. What is certain though is that my cooking app has immediately displaced all the others on the market , because none of the others cater to my requirements.
The production side is indeed of specific interest - most users don’t run production software so I had to think about that one. Tailscale and Cloudflare came in quite handy and there is indeed a market here
Basically, I am prepared to accept that there is a friction that LLMs lubricate away, but what is the source of the friction, and why am I (and a bunch of other colleagues) not feeling that friction daily in our practice?
[1]: And if so, where did we programmers and computer scientists go wrong? Were subroutines and macros not sufficient for automating all of that excess typing? Were Emacs and Vim simply not saving enough keystrokes? Did people forget how to touch-type?
You must be extremely talented and fast if LLMs make no difference for you.
For people like me though, it's another story: I've been doing this professionally for 25 years and of course, like many, I have been writing custom software for my own use all this time, on personal time. But with LLMs I get better results, faster and with very little effort. And that is the difference between another item in my list of unfinished software that consumed too much of my weekends and a cool utility/toy/useful thing I got after a few fun and interesting chat sessions.
> I find it hard to believe that there is a demographic of people that were yearning to write code, but simply could not because they lacked LLMs.
We didn't lack LLMs, we lacked time and energy.
Yes, because the price is measured in time.
With LLM tooling I’ve churned out idiosyncratic tools that fit my use cases quickly. Takes maybe a day instead of a week. A week instead of months. The fast turnaround changes the economics of writing custom tools for myself.
I also know that these days, for all kinds of reasons, I do not have the time to write the tools I’m writing now without AI. I don’t lack the ability, and I could - it will simply be multi months side projects that I can’t / won’t complete.
I'm a long time ops guy. I script, but I spend most of my time configuring, patch testing, and keeping the low level infra running much of which doesn't require "coding" per say. Infra as code is in the grand scheme relatively new and still not ubiquitous despite what silicon valley would have you believe. I never had a need to learn to code to a level to do many of the things I'd like to see happen and find useful. Now I can make those software desires a reality without having to alter my career, preferred hobbies, or much of anything else about my life.
I write plenty of code at my job, and generally don't have the desire to write more code as a hobby, except in rare cases when the mood really strikes.
I think the instinct that APIs, validation layers, and so on take on a much higher importance is right.. I have a few internal tools that made sense to make libraries out of, and once the first library is good, and a test suite is comprehensive, porting to a bunch of different languages is extremely simple.
Everting that, it's also going to be simple for someone to hook up to this library with custom tooling.
Really interesting period in computing, for sure.
What period were we for the past 50 years?
The ~20 years prior to that we were in a world where you chose to align with either Microsoft's tooling, IBM, or shops providing Unix tooling from proprietary vendors.
I elide a nearly infinite amount of detail, obviously.
What's new now is that you can get your own window manager written to spec in under a week, perhaps much more quickly, not just choose one of a few major window managers and configure it in accordance with the chosen configuration options delivered by the large developer team.
[1] https://daniel.lawrence.lu/blog/2025-10-22-sriv-simple-rust-...
More like Star Trek, we would just ask "computer" to do things, and its machinations (and "software") will be invisible to us. We would just have output to deal with.
I think this would mean a lot of things. I'm sure I can't fathom all of the implications, but it sure makes me feel old! Interesting times ahead.
Too bad this is all on the work computer and need to bring it to my personal one but can’t copy paste lol. It’s been thrilling building g and using them and the time from an ideating a small enhancement/ optimization to actually using it is like 5 to 15 minutes away. Soo cool.
So I think the same thesis holds for audiences of 10-100 and 100-1000.
A cambrian explosion of software.
Identifying a vulnerability that can be exploited against many thousands or millions of targets is perhaps more attractive than a single one of individually low value.
This of course would assume that vulnerabilities are in fact unique (which is admittedly questionable).
(Appreciate your counterpoint for its own sake. It’s an interesting idea.)
(Note: I’m not an LLM fan, don’t vibe code myself at all. But I would be unconcerned about security for the kind of things I would create if I did start doing so.)
My wm, shell, terminal, editor, file manager, pop-up menu (dmenu-like) are all pure ruby (including font rendering and X11 bindings). These all started before I started using Claude to improve them, so they're still mostly hand-written, but that is changing.
They're messy, they have bugs and "misfeatures" that works for me but likely would be painful for others.
Like OP, I don't really recommend anyone else use my code, at least not directly, and that is extremely liberating.
Overall, the projects covers the largest surface of what I use beyond the kernel, a browser, and Xorg (I'm so, so tempted, but I think an LLM will need to get a lot further first before I could fit it into my schedule).
It doesn't need to be polished because it's mostly for me. It's okay for them to have bugs as long as they work better for me than the alternatives.
I strongly believe more people should do this. It's both a great learning experience, and it gives you a system that has exactly the features you actually want and use.
And it's only going to get easier to do this.
Cheers :)
[1]: https://fortune.com/2026/04/28/nvidia-executive-cost-of-ai-i...
[2]: https://www.briefs.co/news/uber-torches-entire-2026-ai-budge...
As a hobby, normal rates don’t apply, but just not to be misleading on the equivalent cost.
I've spent two weeks with the cheapest tiers of Claude Pro + pi.dev+GPT-5.5 (+ some deepseek-v4 via openrouter recently) to create my own bespoke version.
I'm at 90% feature parity currently and surpassed on some levels. For ~20€-ish I've soon replaced a 60€/year subscription service.
I haven't spent a single second thinking about how someone else might run it, it doesn't have logins, security or anything - because I'm going to run it 100% behind a Tailscale node with no external access. The release and deployment processes are exactly how I like it, other people might not. I don't need to care, it's mine =)
A few months ago I did the same with Hazel[0]. That took maybe one evening to get an MVP and a week of casual updates to make it pretty. Now I have my own macOS application that does the exact things I needed Hazel for. It's mine forever and I can add or remove features as I feel like it.
I often continue a session on my phone, sometimes with voice. I have buttons for viewing files or following links the agent has referenced, extracted from the stream of text, and I have some buttons for exactly the git stuff I need. I have a button to toggle between yolo mode and normal.
Basically, very simple UI for everything I actually use, easy to use on a phone - and maybe more importantly, no UI for anything I don't personally use. Also all my machines have the repo for the uh, harness-harness, so I just open the tab for it if I need some changes and prompt them into existence and get the changes live.
All this is great, except it enables me to work every waking hour of my life. That part might be bad.
My goal is to have agents work independently from any of my devices (hence VM setup) and to access any session at any time.
My setup so far:
A simple VM on Google Cloud. Always on. Repo lives in there. Claude Code, Codex, Gemini installed as CLIs.
The agents run inside terminal sessions on the VM, managed by zmx (small Zig project, keeps a raw PTY around). Session keeps running regardless of what’s connected.
From the Mac, transport is Eternal Terminal on port 2022. ET is like SSH but reconnects silently. I never see the disconnect.
From the phone, transport is mosh. UDP based, roams gracefully (cellular to wifi and back). Client is Moshi (iOS), renders the terminal and manages tmux sessions well. Moshi needs alternate screen for scrollback, so on the phone path zmx is wrapped in tmux. The tmux is just an adapter for the phone client, not the persistence layer.
The point: both transports attach to the same zmx session. Same scrollback, same prompt, same agent state. Laptop, phone, both at once, all clients of the same long running shell on the VM.
The annoying part was scrollback: duplicated history, lack of native scroll, lagging. zmx avoids the alternate screen (tmux uses it by default, which makes scrolling agent output awkward). Combined with Claude’s --no-flicker flag, the terminal’s own scrollback ends up actually usable.
I wanted to avoid manually switching scroll modes and re-attaching. Moshi sees all my current sessions on the Mac, I just tap and swipe between them. Awesome.
A few shell functions hide the layering on the Mac with a few commands and keyboard shortcuts.
I use a whisper wrapper (also "built for one") for Windows and connect through SSH but it only works well enough because of the laptop's ARC graphics card. I wish I could do that when connecting to SSH from my smartphone.
They also do a pure play voice input keyboard, if that's your jam or you want to use it with other keuboards: https://voiceinput.futo.org/
But I also type into an input field rather than the terminal and can use the mic on the phone keyboard (iOS).
It points toward the parts of "reduce, reuse, recycle" that capitalism has demonstrated no interest in serving
A word of warning: a reliable lock tool for X11 is difficult. You should look at XSecureLock, which uses a multiprocess approach to avoid leaving the desktop unprotected in case of crash. It also implements a number of countermeasure to ensure the desktop stays locked and the locker stays in the front of the display. It's small too, so easy to audit (but written in C).
On this software itself: I’d like to know how this feels to use. It’s so very lightweight. Does it feel categorically different to what we are used to?
One of the things I miss about the 1980s home computers is that they booted into a usable command line in a handful of seconds, from a few KB in ROM. Imagine what today’s HW could do if we’d retained that level of efficiency.
Some of the folks who make things will go on to make things that suit not just their preferences but also those of a small audience.
Some of those audiences will go on to grow and grow and disrupt the big players.
The capital intensive part of software construction is melting away and being converted to opex (payg token costs and your time) and that will blast open the possibility space and lead to a massive new commons.
If the thing was so cheap to create why not open source it!
And if you like someone else’s open source thing but don’t want to take it wholesale why not give it to your agent and say “put the ideas from this onto my thing”!
It’s a new way of thinking about code too.
Another thing to watch for is how chatty the internet is about to become. A great many of these apps will hit APIs, ping each other, and so forth.
There is part of me that understands the appeal of the all-in on AI and personalized software approach. It's a bit cyberpunk! In terms of open-source software, the downsides outweigh the benefits in my opinion, though. Important principles like community ownership and commitment are absent, and this approach is even radically antisocial. And then there's the inevitable issues with maintainability, to say noting about dependence on big tech companies.
To each their own, but this is not for me.
those in the first camp are having a great time.
those in the second camp (which is how you're describing yourself, and how I'd describe myself) are wary and suspicious.
it is somewhat paradoxical, we've watched/read sci-fi/cyberpunk for years and dreamed of this kind of world. after all, when did you see any members of the Enterprise writing code? they just asked the computer to "write a subroutine" and that was that. what a world!
but here we are, with the craft in danger, not entirely impressed by the idea of "just ask and walk away".
i, too, fear for my loss of critical thinking, raw skills, and design sense, as do i think about being one of the few (in 2, 3, 5, 10 years) that didn't abdicate their cognition, their craft, to the tech overlords.
but i wonder if it will matter anyway. i wonder if "source code" will be a deep abstraction that nobody thinks about anyway, similar to how 99% of us don't care/need to care what the machine code we're eventually emitting does or looks like.
in any case, i'll keep my thinking for now.
Surely you read it more than once, because that has become a talking point. It’s a false dichotomy that, you’ll notice, is most often used by the people who put themselves in the first camp to steer the conversation. By framing it as “there are two camps, it’s just different, none of them is better”, it lends legitimacy to their position.
You don't have to pick one camp over the other. Good, high quality craft makes good products.
> after all, when did you see any members of the Enterprise writing code?
When did you see anyone in any media taking a dump, or sleeping, or doing any of the boring bits? Rarely, because if it’s not relevant to the story they don’t show it, but it doesn’t mean it didn’t happen.
I’m more of a DS9 fan, and I remember them having computer problems all the time. O’Brien, despite being highly competent and the chief of engineering with a team, was constantly overworked.
And their computers were infinitely superior to the LLMs we have now. When they gave you an answer, you could be confident it was correct. And if they didn’t know, they’d tell you!
Part of it's that the whole point of going into this industry is that I love coding and have been doing it since I was 8. Part of it is that I'm a control freak and it makes me uncomfortable to have to trust AI generated code. Sure, I already trust interpreters and compilers, but those are much more deterministic, and they don't generally do anything I have to be wary of. Part of it is that anytime I've used Claude to write stuff (using Opus 4.7 via an API key), I've had to handhold it when doing simple things (telling it repeatedly that a given column doesn't exist in Snowflake's task history table and eventually just giving up and taking it out by hand) and had to remove tons of completely pointless Python code it generates. The big difference is that the people in the first camp don't seem to care enough to check. Someone at my company used Claude to write 20k lines of code this past Friday. No way he read and scrutinized all of that in one day.
The other big thing I've noticed is that a lot of the people using it extensively seem to just be spitting out API endpoint after endpoint. Just doing endless CRUD with some light business logic. Yeah, it's not too hard to automate that with AI without any major issues. Hell, back when Ruby on Rails was hot, it was so fast to write those kinds of things with it that I could spin up things as fast as AI is doing now. Full websites or APIs in an hour or two because its syntactic sugar and scaffolding did what AI does with the FastAPI codebases I see these days. You could go from an ER diagram to a working app in minutes sometimes. I don't care that much if that kind of work is automated.
What we got are "statistical machines" that tend to do the right thing under the right conditions, but can go completely off the rail every now and then.
The former are more akin to a generalization of computers as we typically think of it, whereas the latter is something else. Maybe that something else is closer to human behavior in some ways, but also so very different - unlike humans, where you get to know people, build relationships, know who to trust in what ways, and so forth, you can never really trust an LLM with any critical tasks without close supervision.
I, in theory, can plane a piece of wood with a hand planer. But I'll never do it again, we did it at school in ye olden times before the millennium and it was boring then as it is boring now.
I know people who get satisfaction from it, they take one sliver off with the hand planer, feel the wood with their hand and figure out the perfect angle for the next tiny sliver of wood to come out off, repeating the process over and over again.
I, personally, will just feed the damn plank to a mechanical planer with the exact specs of the resulting board set up. I just want the board smooth so I can get to the next step of the process. I'm not doubting the "wood-slop" the machine produces, I can see and measure if it's good enough or not. I don't need to be involved in the process.
We're both making a table, mine will be done faster. It might not be hand-crafted to perfection, but it will hold the stuff I intend to put on it just fine. If I find out it sucks later on, I can make a new one that's slightly better or fix the existing one. My goal was a functional product, not a piece of handcrafted art.
More close is: if there was a table making machine, you just push a button and something like a table comes out, would you still be a woodworker? You haven't planed, nor measured, nor cut, nor jointed, you've only pressed on "make me a table"
Another example: I enjoy writing with a good pen. But whether I write by pen or on a keyboard, it's still me writing it.
However, AI does basically all the real work, only leaving you to guide it. Make a table? AI gives you one with 2 legs. More legs? Guess I can live with 5 legs.
And you wouldn't be making that table, AI is. You cannot have pride in something that you never made yourself. It's the same as 3D-printing something from Thingiverse and claiming you made it.
People who create AI blog posts are not writing. Those that prompt their way to a piece of software are not doing software engineering. The ones that generate AI images are not being artists.
There are big benefits to using a language that has good static analysis with LLMs.
rust can do that. You can run a hyper stripped down rust that was made for embedded devices specifically because those devices don't have room for a runtime.
This is... Not really true? Especially if you are writing just for yourself. These are week-long projects at most to get to a usable state, if you know what you're doing. This is why there are so many text editors and window managers in the first place.
I struggle to understand why, though.
"Computer. Run program Riker omega six." <things happen>
Also, reading it is probably not the intended use. It’s probably: “Hey Claude, give me a TLDR of this”
In particular I've found that if you have a good infrastructure layer available on which you can deploy then it's much easier to throw small purpose built webtools on there that solves personal problems. Infra here being fixed IP, mTLS reverse proxy, k3s/container, S3 etc. Basic building blocks like that - store data, run app & safe gateway to access it.
If you have that in place then most smaller apps (shopping list, notes etc) is a couple prompts away
Or a regularly backed up raspberry pi running services in lightweight docker containers exposed over tailscale.
Basically you have a repo with web-based tools and a way to deploy them on change. Then you can just quickly whip up (or ask an LLM to do so) any simple tool and it's live within minutes.
Would it be possible to share the jsonl files too, like how Mario Zechner shared his chats with the AI, while working on his Pi coding agent?
Fe2O3: https://isene.org/fe2o3/
Now it's perfectly possible to do a "good enough" solution in-house for less than what we pay for the SaaS monthly. And as a bonus we own the full solution and can add any features we want without the SaaS provider gatekeeping it.
This kind of summarizes the whole post for me. I struggle to see how, on a platform that I thought was passionate about engineering, this is gaining any kind of traction. Writing GUI tools in assembly, not to learn, but for whatever other reasons, is nightmarish levels of silly. I get the idea of making software TRULY yours. I get it. This isn't that. Letting an AI agent literally vibe code your entire desktop is not an idea that would come to anyone's mind as more than the punchline of a joke or a side note in a dystopian book.
You're not making software, Claude is. You're not learning anything, and the tools produced are (by design) not really editable.
The whole point of this sentiment is that the personal tools wouldn't EXIST due to the time sink needed.
The tradeoff makes sense for a lot of people even if it's not a good fit for you.
A cybersecurity research company can now spend a small fortune on finding zero days in iOS because of the amount of people that use it. It basically guarantees there will be clients like government agencies willing to pay through the nose for the exploits.
Software made for one might disrupt this business model.
With software that's deployed to millions of computers you have an abundance of targets, but trying to target some random LLM average todo list at scale is hard, isn't it?
Once millions of completely unskilled developers have "workflows" that consist of asking an LLM to make a thing, followed by those LLMs pulling in the same 100 (often outdated versions of) dependencies, you have a beautiful attack vector.
Yes, it's "easy" to attack something like Obsidian. It's probably easier to attack a couple hundred dependencies LLMs like to use, or to test what LLMs commonly do to implement things from scratch, and attack those weaknesses.
We are just lucky that enough real, smart, people engineered things that actually work, are well understood, and keep us safe, like firewalls.
Maybe, in fact, some group somewhere combines supply chain attacks with models and/or agents. A model or agent that’s compromised upstream and becomes designed to insert a backdoor is not beyond possibility.
At day 9, right as he is getting ready to deploy his beautifully crafted shell code, the clock hits midnight UTC. The website shuts down for maintenance.
"This is it" he thinks. "As soon as the backups finish I'm getting in. No problem."
Minutes tick by. He gets up, stretches, sits back down, watches the clock impatiently. Then, as he prepares to start refreshing the site he recollects, "I'm glad I begged so hard to get authorization to use this PHP 0day."
His partially obscured terminal window has the script ready to launch, all arguments pre-populated, waiting for the link and session token to be pasted in and executed.
The site comes back up. But the url of his launch point returns 404. Undaunted, he returns to a previous url. It is also 404. He curses aloud. Beginning to perspire, he goes to the homepage and prepares to navigate back to the launch point.
The link isn't there. Well, it's there, but it has changed.
"What the....!" The link is no longer a PHP url. He mouses over other links. NO links say PHP anymore. Starting to panic, he clicks on links at random. Not a single one appears to be PHP.
The following morning he schedules an urgent meeting with his supervisor.
"How's that project coming along. Got anything yet?"
"No. I, uh...I'm going to need a bit more time."
"Oh?"
"Yeah. Uh. The site. It got..." He mutes his microphone and, for the 22nd time since midnight, he screams in frustration. Unmuting, he continues:
"It got rewritten. Completely. In Nim."
"What??"
"Yeah. It's some esoteric language that just got a new web framework. I guess somebody decided they wanted to mess around with it. So they vibe coded a complete translation. The whole front end is nimlang now. None of the PHP attacks are going to work on it."
His supervisor expresses his disgust and ends the call.
11 days later the process repeats itself, this time with Rust.
The TAO engineer submits an application to change jobs to the DoD's procurement division, then requests an appointment with a mental health counselor.
Kind of funny seeing that at the bottom of this article. Especially given that static site generators are probably one of the biggest roll-your-own categories of software.
We're actually going to have hammers designed by blacksmiths, not by comittee.
Most software is done after the first or second version and the developers just keep working on it to justify their job; adding features no one needs and just get in the way or make the program worse. It'll be nice when the software I have does exactly what I need and doesn't change until I tell it to change for something I need.
The only feature Macos has shipped in the past 10 years that I actually like is air-drop. Everything else is a PITA annoyance, or as I've found out from upgrading, just bug ridden slop that doesn't work well anymore.
The other thing is that other people's applications are rarely useful. Their libraries are, the feature description READMEs are, but the software itself is full of attempts at generality that make them overly annoying for me to use. Instead I have extremely idiosyncratic software - anyone else would find it insufferable.
The wild thing, though, is that my software is outrageously useful for me. I can see why Anthropic and OpenAI are (or shortly will be) the trillion-dollar behemoths they are. They are enabling a personal productivity increase of epic proportions[2]. The highly specific functionality also means strange things performance wise. I don't need to use Electron or Tauri or whatever. Instead, my thing is Rust with objc2 and it starts instantaneously. On my M1 Max, it's the fastest text viewer I can start. 100s of megabytes of JSON and it's launching is imperceptible for my tool, pretty-printing is instantaneous, breadcrumbs are live.
Because I can make it do only the thing I want it to do. It can't do other things. I cannot edit or auto-complete or anything. And this is great. Useless to others and fantastic to me.
Likewise, my blog is on Mediawiki (which I like so anyone can edit) but the authoring flow is kind of annoying. Uploading images causes a break from writing, and requires a lot of form-filling that interrupts my thought. So I now have this software that does everything I want: link autocompletion, background image uploads, post-hoc publishing, previews and diffs, built-in Wikipedia search to interwiki link. Who would want this but me? It only brings me pleasure.
What a revolution in software.
0: https://wiki.roshangeorge.dev/w/Blog/2026-04-25/The_rise_of_...
1: https://wiki.roshangeorge.dev/w/Blog/2026-04-30/Personal_Sof...
2: Predictably, I have chosen to use the spare time on leisure
Brother mine, you will learn that the future you is ignorant of all the things, and every bit of documentation goes a long way
Its fun, and a lot more rewarding than replacing tool X with tool Y, realizing you actually hate it.
Not sure I can use it as a daily driver yet, but it would be pretty cool!
ok seems a lot of fun (for those like-minded), but who seriously want to be dealing with maintenance of everything they use in the long term, in pure assembly all the more?
I do find it curious how even after replacing all of your software, but are still using Claude Code instead of building your own coding agent.
However, gpt5.5 does not make me miss Opus.
And before you say "oh but these are just stochastic parrots and this is really stolen code" - well the jury is stil out, but currently companies don't have these kind of problems, so I expect that anyone will be able to go "my claude did it so I declare it to be Free software".
Programmers turning into mindless slop feeders.
