Building a Mac app with Claude code
indragie.com
indragie.com
It feels like I've got superpowers. I love it. I feel productive, fast, creative. But at night, there's this strange feeling of sadness. My profession, my passion, all the things I worked so hard to learn, all the time and sacrifices, a machine can now do most of it. And the companies building these tools are just getting started.
What does this mean for the next generation of engineers? Where's it all heading? Do you feel the same?
The thing LLMs replace is the need for understanding all of the trivia required for each platform.
I don't know how to write a for loop in Go (without looking it up), but I can write useful Go code now without spinning back up on Go first.
I still need to conceptually understand for loops, and what Go is, and structured programming, and compilers, and build and test scripts, and all kinds of other base level skills that people without an existing background in programming are completely missing.
I see LLMs as an amplifier and accelerant. I've accumulated a huge amount of fuzzy knowledge across my career - with an LLM I can now apply that fuzzy knowledge to concrete problems in a huge array of languages and platforms.
Previously I would stay in my lane: I was fluent in Python, JavaScript and SQL so I used those to solve every problem because I didn't want to take the time to spin up on the trivia for a new language or platform.
Now? I'll happily use things like Go and Bash and AppleScript and jq and ffmpeg and I'm considering picking up a Swift project.
With minor exceptions, moving from one language to another isn’t a matter of simple syntax, trivia, or swapping standard libraries. Certainly, expert beginners do espouse that all the time, but languages often have fundamental concepts that they’re built that and need to be understood in order to be effective with them. For example: have you ever seen a team move from Java to Scala, js to Java, or C# to Python - all of which I’ve seen - where the developers didn’t try to understand language they were moving to? Non-fail, they tried to force the concepts that were important to their prior language, onto the new one, to abysmal results.
If you’re writing trivial scripts, or one-off utils, it probably doesn’t build up enough to matter, and feels great, but you don’t know what you don’t know, and you don’t know what to look for. Offloading the understanding of the concepts that are important for a language to an LLM is a recipe for a bad time.
I completely agree. That's another reason I don't feel threatened by non-programmers using LLMs: to actually write useful Go code you need to figure out goroutines, for React you need to understand the state model and hooks, for AppleScript you need to understand how apps expose their features, etc etc etc.
All of these are things you need to figure out, and I would argue they are conceptually more complex than for loops etc.
But... don't need to memorize the details. I find understanding concepts like goroutines to be a very different mental activity to memorizing the syntax for a Go for loop.
I can come back to some Go code a year later and remind myself how goroutines work very quickly, because I'm an experienced software engineer with a wealth of related knowledge about concurrency primitives to help me out.
I’ve been encouraging the new developers I work with to ensure they read the docs and learn the language to ensure the LLM doesn’t become a crutch, but rather a bicycle.
But it occurred to me recently that I probably learned most of what I know from examples of code written by others.
I’m certain my less experienced colleagues are doing the same, but from Claude rather than Stack Overflow…
Code are practical realizations of concepts that exists outside of code. Let's take concurrency as an example. It's something that is common across many domains where 2 (or more) independent actors suddenly need to share a single resource that can't be used by both at the same time. In many cases, it can be resolved (or not) in an ad-hoc manner. And sometimes there's a basic agreement that takes place.
But for computers, you need the solution to be precise, eliminating all error cases. So we go one to define primitives and how these interacts with each other and the sequence of the actions. These are still not code. Once we are precise enough, we translate it to code.
So if you're learning from code, you are tracing back to the original concept. And we can do so because we are good at grasping patterns.
But generated code are often a factor of several patterns at once, and some are not even relevant. Just that in the training phases, the algorithm detected similarity (and we know things can be similar but are actually very different).
So I'm wary of learning from LLMs, because it's a land of mirages.
LLMs are like a map of everything, even it's a medieval map with distorted country sizes and dragons in every sea. Still, it can be used to get a general idea of where to go.
In software these days that means learning from many different sources: documentation and tutorials and LLM-generated examples and YouTube talks and conference sessions and pairing with coworkers and more.
Sticking to a single source will greatly reduce the effectiveness of how you learn your craft. New developers need to understand that.
That's one of the argument that seems to never make any sense to me. Were people actually memorizing all these stuff and are now happy that they don't have too? Because that's what books, manuals, references, documentations, wiki,... are here for.
I do agree with you both that you need to understand concepts. But by the time I need to write code, I already have a good understanding of what the solution is like and unless I'm creating the scaffolding of the project, I rarely needs to write more than ten lines at time. If I need to write more, that's when I reach for generators, snippets, code examples from the docs, or switch to a DSL.
Also by the time I need to code, I need factual information, not generated examples which can be wrong.
The reason I used to stick to just a very small set of languages that I knew inside out is that because I was using them on a daily basis I was extremely productive at writing code in them.
Why write code in Go where I have to stop and look things up every few minutes when I could instead of use Python where I have to look things up about once an hour?
LLMs have fixed that for me.
> Also by the time I need to code, I need factual information, not generated examples which can be wrong.
If those generated examples are wrong I find out about 30 seconds later when I run the code: https://simonwillison.net/2025/Mar/2/hallucinations-in-code/
My workflow is different. Whenever I can't remember something, I look it up. And for something that I'm not familiar with, that usually means a lot of browser tabs. But I try to retrieve it from my memory first (and it usually exists in my short-term memory). And after some days, it has become part of my long-term memory and I don't need to look it up again. If I don't touch the language and its ecosystem for a while, I forgot some details, but they are refreshed with a quick read.
> If those generated examples are wrong I find out about 30 seconds later when I run the code
I would be grateful if those were the only kind of wrong I encounter. Writing code that does the stuff is usually easy. The difficult part is to ensure that the data fits the definition of correct at every stage of the process. And there's not some corrupting action somewhere. Doing so exhaustively is formal verification and it is costly (mostly in terms of time). So what you do is other kind of testing which only covers the most likely places and dimensions where thing can go wrong.
The stressful part is when you have to change some part to fit a new use case. You then have to carefully revise both the change and the other parts so that their definition of correct is consistent. And sometimes there's some decoupling mechanism which makes the link non-obvious.
That's the definition of wrong that I'm fearful of. Where two part are internally correct, but their integration is not consistent.
I run Python workshops on weekends, and honestly, I'm speechless at the things students are already doing. But is it realistic to encourage them to study computer science in a year or two, when it's still a 4-year degree? By the time they start uni, LLMs will be 100 times more powerful than they are today. And in 6 years?
Any advice?
I think this is not true for most programming languages. Most of the ones we use today have converged around an imperative model with fairly similar functionality. Converting from one to another can often be done programmatically!
You can’t tell me that precisely the same paradigms, mechanics, patterns and practices will be used to develop software in both Java/Spring and Ruby/Rails, for example. The way you need to (be able to) think about how you write the software using those two languages will be non-trivially different.
Like, languages aren't just different ways to express the same underlying concepts? It's clearly not the case that you can just transliterate a program in language X to language Y and expect to get the same behavior, functionality, etc. in the end?
You most definitely can; if you have Turing complete languages A and B, you can implement an interpreter for A in B and B in A; you can implement a simple X86/aarch64 interpreter in either A or B and use that to run compiled/interpreted programs in the other; you can also write higher-level routines that map to language constructs in one or the other to convert between the two.
some elements are necessities to the language while others are more for the human / team.
That's from 6 months ago and the tooling today is almost night and day in terms of improvements.
Also:
> This wasn't something that existed; it wasn't regurgitating knowledge from Stackoverflow. It was inventing/creating something new.
Didn't he expressly ask the LLM to copy an established, documented library?
It’s odd the post describes what it’s doing as creating something new - that’s only true in the most literal (intellectually dishonest) sense.
I find it kind of funny (or maybe sad?) that you say that, then your examples are all languages that basically offer the same features and semantics, but with slightly different syntax. They're all Algol/C-style languages.
I'd understand moving from Java to/from Haskell can be a bit tricky since they're actually very different from each other, but C# to/from Java? Those languages are more similar than they are different, and the biggest changes is just names and trivia basically, not the concepts themselves.
LLMs aren't good engineers. They're mediocre, overengineering, genius, dumb, too simplistic, clean coding, overthinking, underthinking, great but terrible engineers rolled into one. You can get them to fight for any viewpoint you want with the right prompt. They're literally everything at once
To get good code out of them, you have to understand what good code is in the first place and lead them to it
My experience is iterating through 5 potential options being like "nope, that wont work, do it this way.. nope adjust that.. nope add more of this.. oh yeah that's great!"
The "that's great!" made me nervous that an LLM was capable of producing code as good as me until I realized that I threw out most of what it gave.. and someone less experienced would've taken the first suggestion without realizing its drawbacks
That's a great way of putting it. Completely agree.
It's pretty wild that you're characterizing the understanding of a language and its details as "trivia" that you don't need to "take the time to spin up on" before you write programs in that language.
I mean, I get this perspective, but it's the position of a product manager, not a software engineer...
That's not to say it's trivial, or to disparage that knowledge. But it's not at the same level as understanding how eg Unix processes can be piped together.
If I'm interviewing a candidate and they can't remember the syntax for a Bash loop, I don't care. If they can't explain what happens when you pipe output from one process to another (at least at a high level), that's a problem.
Being technical still is an advantage, in my opinion.
However, they also likely had little education, 1-2 of their children probably died of trivial illnesses before the age of 10, and they lived without electricity, indoor plumbing, running water or refrigeration.
Yes, blacksmithing your own tools is romantic (just as hand writing python code), but somehow I’m guessing our ancestors will be better off living at higher layers of abstraction. Nobody will stop them from writing python code themselves, and I’m sure they’ll do it as a fun hobby like blacksmithing is today.
Can it? It doesn't have experience, it doesn't have foresight nor hindsight, it cannot plan and it just follows the instructions, regardless if the instructions are good or bad.
But you, as a human, have taste, ideas, creativity and a goal. You can convince others of your good idea, and can take things into account that an LLM can only think about if someone tells it to think about it.
I'm not worried about programming as a profession disappearing, but it is changing, as we're moving up the ladder of abstractions.
When I first got started professionally with development, you could get a career in it without knowing the difference between bits/bytes, and without writing a line of assembly, or any other low-level details. Decades before that, those were base-level requirements for being able to program things, but we've moved up the abstraction ladder.
Now we're getting so far up this ladder, that you don't even need to know a programming language to make programs, if you're good enough at English and good enough at knowing what the program needs to do.
Still, the people who understand memory layout, and assembly, and bits and bobs will always understand more of what's happening underneath than me, and probably be able to do a "better" job than me when that's needed. But it doesn't mean the rest of the layers above that are useless or will even go away.
But the stuff I enjoy doing is something I don’t delegate to the agent. I just wonder how long the stuff I enjoy can be delegated to an agent who will do it cheaper and faster.
I feel since the worst criticisms of llms in this thread boil down to down to ‘llm might be offline, might include nsfw stuff and might lead to non experts coding with poor results’ which are such weak criticisms that it’s is a damn good compliment to just how good they’ve gotten isn’t it?
I’ve been (carefully) using AI on some AWS infrastructure stuff which I don’t know about. Whenever it proposes something, I research and learn what it’s actually trying to do. I’ve been learning quickly with this approach because it helps point me in the right direction, so I know what to search for. Of course, this includes reviews and discussions with my colleagues who know more, and I frequently completely rewrite what AI has done if it’s getting too complex and hard to understand.
The important thing is not allowing AI to abstract away your thought process. It can help, but on my terms
You lose some, you gain some.
Although I'm not happy with code quality/style produced in most cases, or duplications, lack of refactorings to make the code look better/smaller.
One post that fits with this feeling is https://silicon-dialectic.jostylr.com/2025/07/02/impostor-sy... The techno-apocalyptic lullaby at the end that the AIs came up with is kind of chilling.
no need to be sad though as you can tell from your own creations you haven't really created anything of any value. its all AI slop.
Yes, IDEs of the future will not look like those we have today. I too started with Cursor and similar VS Code enhanced IDEs. And ended up using Claude Code. And realised my terminal is more important for me now. So I migrated from iTerm to Ghostty (faster, lighter, more modern), Tmux and Tmuxinator, and NeoVim! Because I just cat/bat files and occasionally edit a file, if at all. Claude Code does the heavy lifting, I write specs and prompts, in NeoVim or Emacs. And I am loving my workflow.
I too use Claude Code for more then just generating code. Whenever I need to rewrite a config file (for zsh, neovim, ghostty, etc), I start Claude code and assign the task to do it. It will do the changes and even refactor my settings file in a few minute.
Lastly, I use it to ask questions about my codebase, refactor code, let it document my code, even make meaningful commits.
Pure awesomeness.
I've found that CC writes really good commits; I added info about Conventional Commits [1] to the CLAUDE.md file with a few examples.
I think I would revise it now to allow that LLMs can be really useful teachers for unskilled developers, however I don’t get the sense that that’s how they’re being used in many cases. It seems that it’s more common that unskilled developers using LLMs just vibe code their way through issues that they run into, never really gaining insight into why the thing they’re working on failed, and just push through until it seems like it isn’t failing anymore.
And that more or less reinforces my idea that they’re dangerous assistants in those cases where the developer is unskilled. It’s pretty much inevitable that this will introduce problems that make it through the process of creation not only unnoticed, but problems that the developer is incapable of understanding even if they are noticed.
This is my conclusion to date also, however, I believe the reason behind that is the increasing adoption of agentic tools like Claude Code and Gemini CLI etc, that encourage users to disregard the actual code generated.
For instance, I am adding bluetooth capabilities to my app, and the bluetooth manager can be quite large, to the point where the LLM context window starts to become an issue.
When it gets to that point, the LLM can start making suggestions that don't make sense; it's starts forgetting what you had already done, it can start making architectural suggestions that contradict what you had been doing before. So if you follow them, you end up in this weird state mired in conflicting decisions. Almost like design by committee without the committee.
It's important to recognize this when it's happening rather than just blindly following what it suggests, so some expertise at engineering is still necessary even though the LLM can write okay Swift code.
I never really enjoyed using full blown ide's anyway, setting them up was a chore, configuring QT to cross compile for various platforms was a project in itself, I always found that using a text editor and terminal the most logical combination for editing. Adding another terminal window with claude code to perform my requests while i debug, edit, test, review, i feel like i have graduated from coder-editor to project leader and there are no employee associated problems to deal with. All or my shelved side projects are complete as i wanted them and all in just a few months since March when Claude went big.
Ok, so they knew where Claude went wrong and could correct for it.
So unless you have 15+ yoe better add more reps. You can always switch to llm code assist in a blink, there is no barrier to entry at all.
Agreed, but I wish I had it as a teacher while learning. The amount of help my interns need from me has reduced by at least 50%, and what remains is the non-trivial stuff which is completely worth my time to coach and mentor them
You learn far far faster from reading code and writing tests than you do just writing code alone.
I've long suspected for years that the bottleneck to software development is in code generation not keeping up with idea generation.
This is not about syntax but about learning how to create solutions.
When you read solution you merely memorise existing ones. You don’t learn how to come up with your own.
You can't solve programming problems without reading code.
Period.
The more code you read the better you get at solving problems because your internal knowledgebase grows.
I think they add to expertise honestly.
I also haven't had much luck with getting llms to generate useful code. I'm sure part of that is the stack I am using is much less popular (elixir) than many others, but I have tried everything even the new phoenix.new , and it is only about an 80 to 90% solution, but that remaining percentage is full of bugs or terrible design patterns that will absolutely bite in the future. In nearly everything I've tried to do, it introduces bugs and bug hunting those down is worse to me than if I just did the work manually in the first place. I have spent hours trying to coach the AI through a particular task, only to have the end solution need to be thrown away and started from scratch.
Speaking personally, My skills are atrophying the more I use the AI tools. It still feels like a worthwhile trade-off in many situations, but a trade-off it is
Where I’ve found them best is for generating highly focused examples of specific APIs or concepts. They’re much better at that, though hallucinations still show up from time to time.
If they diminish your skills, you're letting them do that. Learn to apply these tools in a way that helps you learn more things and do better work.
In the real world, away from those whose salary depends on marketing these agentic tools, an LLM is a context shredder. It provides plausible code snippets that are globally incoherent and don't fit style. CONVENTIONS and RULES files are a kludge, a sloppy hack.
These tools flatten the deep, interconnected knowledge required to work on complex systems into a series of shallow, transactional loops that pretend to satisfy the user.
The skill being diminished is not the ability to write a single-page utility or single-purpose script. It is the ability to build and maintain a mental model of a complex machine. The ability to churn out a hundred disparate toy tools is not evidence of a superior learning method, it is evidence of a tool that excels at tasks with no deep interconnected context.
I wrote about my process for non-vibe-coded projects here: https://simonwillison.net/2025/Mar/11/using-llms-for-code/
> The skill being diminished is not the ability to write a single-page utility or single-purpose script. It is the ability to build and maintain a mental model of a complex machine.
That's the thing that LLMs help me with 90% of the time. It's also why I don't think non-programmers armed with LLMs are a threat to my career.
And none of these AI companies are profitable. Imagine how much it will cost or how much it will be enshittified when the investors come looking for their returns.
If everyone whose code illegally trained these models won a copyright lawsuit against Claude it would suddenly not be so good at writing swift code.
Do we really want to bet on Disney losing their AI lawsuit?
Honestly I realize my comment is not adding much to the discussion but the AI fatigue is real. At this point I think HN and other tech forums would do well to ban the topic for posts like this.
Imagine if we were upvoting stories about how people are getting lots of coding done easier with Google and StackOverflow. It would be rightfully ridiculed as vapid content. Ultimately that’s what this blog post is. I really don’t care to hear how yet another dingus “programmer” is using AI to do their hobby/job for them.
For people who do not need thinking they can use flash lite which is cheaper as well.
If AI is truly progressing at the amazing rate that everyone is claiming then we should be continually getting better capability at the same price.
It then does a bad job, which, like, makes sense if it's using the flash model.
Also i code all day and have yet to hit the $10/month cap on claude that the jetbrains library offer.
But I’m also not enjoying the idea of software engineering requiring a subscription to a service provider in a similar way to how creative professionals get backed in to adobe creative cloud, because up until this point most developer tools have been free or very cheap.
I find paying for Claude significantly less offensive than paying for e.g. Creative Cloud, because I'm largely paying for the compute resources.
Heck, I know a web developer who was using a raspberry pi as their main workstation.
Prior to AI the person with a used ThinkPad wasn’t really at a disadvantage when it came to most development tasks.
So, firstly, the limits on the $100/month plan are reasonably high in my experience. I do hit them, but it takes quite a bit.
> I have trouble convincing myself to give Autodesk $50/month and I need that software for my primary hobby.
Before I used Claude Code, I absolutely would have agreed with you, $100–$200 every month is just a ridiculous amount for any software. After using Claude Code... yeah, that's just how darn good it is!
> Imagine if we were upvoting stories about how people are getting lots of coding done easier with Google and StackOverflow.
You know, something I've thought about before (as in, before LLMs) is just how hard (impossible?) it would be for me to program without an internet connection. I'm constantly Googling stuff.
I can absolutely imagine that if I was a programmer before widespread internet use, and the internet launched with things like Google and lots of resources out of the gate... yeah, that would be revelatory! I'm not saying AI is necessarily the same, but it's something to think about.
No we are going to have a haves and haves not situation where the poor kid who wants to get a better job and get out of poverty is competing with the dude who has a $100-200/month AI subscription writing code for him.
I do feel AI impeding my own learning in this way and it's probably the aspect I'm most concerned about.
E.g., a job interviewer in the world of AI will be asking you to perform more complex example work under the assumption that you have access to AI.
I started with the pay-as-you go plan; I'm currently using the Claude Pro plan at $20/month which is great for my use case.
> And none of these AI companies are profitable. Imagine how much it will cost or how much it will be enshittified when the investors come looking for their returns.
I suspect investors will give AI companies a lot of runway. OpenAI went from $0 to over $10 billion in revenue in less than 3 years. I know that's not a profit but it bodes well for the future.
(As an aside, it took Microsoft 22 years to reach $10 billion in revenue.)
Anthropic went from $0 in 2021 to over $4 billion in about 3 years.
In comparison, it took Twitter about eleven years after its founding in 2006 to become profitable in 2017. And for much of that time, Twitter didn't have a viable business model.
I don't think investors are concerned.
Regarding lawsuits, I'm sure the AI companies will win some and lose some. Either way, after all is said and done, there will be settlements and agreements. Nothing is going to stop this train.
But—chalk one up for Anthropic for winning their case, Meta getting a favorable ruling and all the rest [1].
> won a copyright lawsuit against Claude it would suddenly not be so good at writing swift code.
I suspect Claude is going to get really good at writing Swift--Anthropic is working with Apple [2].
> …but the AI fatigue is real
You might want to get off the internet if you're already suffering from AI fatigue; things are just getting started.
[1]: "AI companies start winning the copyright fight" -- https://www.theguardian.com/technology/2025/jun/30/ai-techsc...
[2]: "Apple Partners With Anthropic for Claude-Powered AI Coding Platform" -- https://www.macrumors.com/2025/05/02/apple-anthropic-ai-codi...
You need to adjust this figure for inflation. Microsoft became huge decades ago when money was worth more.
It’s not being on the Internet that’s giving me AI fatigue. My employer is forcing me to use it. It’s being used at the drive-thru window even. “Touch grass” isn’t a valid argument here.
This idea that AI will improve after the AI companies’ inevitable exits is one that isn’t backed by history. I’d like to have one exited unicorn company named that has a better value and/or lower cost now than its pre-exit state. YouTube? Netflix? Facebook? Uber? MongoDB? Atlassian? Slack?
It is legitimately hard to find a tech company that is a more desirable company to patronize post-exit than pre-exit.
We are so obviously in the customer acquisition phase of AI…people in this thread are talking about spending hundreds per month on AI services and I think all of those will double in price soon after large private company players like OpenAI go public or get acquired. It happened with Netflix, it happened with YouTube (more ads, YouTube Premium), it happened with Uber, the list goes on and on.
I guess it remains to be seen, but the cost for tokens is only going to decrease, not increase. Also, remember that OpenAI and Anthropic get a lot of revenue from large companies licensing and otherwise paying to use these models.
“It’ll only get cheaper” until it doesn’t.
Moore’s law is already dead in the context of the underlying hardware manufacturing technology that powers AI. Once neural processor design reaches an end state we can’t count on any significant lithography breakthroughs powering a reduction in cost.
For example, graphics cards aren’t getting cheaper, and if they are, certainly not very quickly. A lot of consumer electronics categories key to prior tech booms like laptops and workstations aren’t getting cheaper or aren’t getting cheaper/faster very quickly.
Claude Code costs $200/month really isn't something that needs to be "revealed."
Yes, I know of https://xkcd.com/1053/
Indeed but one potential saving grace is the open sourcing of good models (eg. Meta’s Llama). If they continue to open source competitive models, we might be able to stave that off.
> The most exciting thing about this entire journey for me is not the app I built, but that I am now able to scratch my coding itch and ship polished side projects again. It's like I found an extra 5 hours every day, and all it cost me was $200 a month.
I created one utility to show my launchctl/launchd tasks with status indicators for what is loaded/unloaded/running/failed (very much like the menu icon you get for OrbStack) and it works great. With claude it only took a couple of hours to get workig and dial it in the way I wanted.
So ... "how" does one use Claude to make a Mac App? The blog post left me not wiser here.
I have a collection of tools I've built in this way here: https://tools.simonwillison.net/python/
#!/usr/bin/env -S uv run --script
I don't think requiring all CLI tools to be "native binaries" makes sense. Plenty of popular CLI tools are not compiled binaries. The Python and Node.js ecosystems run on those.