Asked it to write a TODO app with FYNE. <1k LOC it needed to get around some sort of tool call it couldn't do, wrote the bash file and executed it.
I purposely had it on YOLO. Everything was backed up idc really. Its been a bug with claude on cursor getting around permissions. It was the day claude 4 came out and I wanted to see what it could do with YOLO. It was kinda funnny but annoying, could be catastrophic for someone who doesn't back up everything obsessively.
I am ai skeptical and have seen it do some dangerous things, such as trying to configure my iptables in a way that will definitely not work at all, let alone the way I asked it to, but I have major doubts that it would do something this stupid.
How recently did this occur? What did you ask it in your prompt? What did it say as the explanation for why it was doing that?
See my comment below. It is a known bug with claude 4 where it tries to bypass permissions. I did have it on full permissions to test claude 4, it wasn't a big deal as I have plenty of backup.
Aside, I tried to move my ~/Code folder (where all my programming projects and work live) into the macOS iCloud folder so that everything would auto-sync.
But the constant Git churn created a lot of traffic and collisions in that folder. And when I googled it, apparently it's just not a good idea to sync Git projects in iCloud.
I really want to easily sync my folders so that I can just get a new macbook and recreate my environment.
When moving someone from one macbook to another, i just rsync desktop/documents/downloads over.
I've found migrating ~/library over is problematic even with built in migration assistant, especially if this is their third or later machine being transfered between.
Git is designed with a remote server in mind, just do it that way. Get github or git lab or setup your own remote host.
TL;DR: Move your .git folders out of ~/Code into ~/Git/[reponame]. Use ln -s or plain text .git file[1].
I have my code folders in Dropbox on OSX but I believe my process should work for you. I work in node and have a number of monorepos with nested node_modules folders. I run my idempotent dev-files.command script [2] with or without "--nono" in any monorepo and it handles the entire git aliasing and dropbox folder ignore for node_modules process. I use --nono if I ever need to clear out all of the node_modules and do a fresh pnpm install.
Interesting, so it's as easy as putting `gitdir:{path}` in a .git file.
The downside is that git history is pretty useful, so it would suck to not sync it.
But you gave a real solution to the problem.
Edit: I remember now that one problem with iCloud was that doing something like switching Git branches or checking out commits causes a catastrophic sync cascade in the code folder itself.
Maybe Dropbox is better at dealing with that kind of thing in general. But that's pretty annoying for any syncing system to deal with, heh.
I guess the simplest solution is to just get into the habit of pushing every little project to github.
I use claude code in a staging env also with full bash permissions via the open source github actions (I think this is much safer and also what you're supposed to do). The github issues and PR integrations make it really organized
tbh that one example isn’t the greatest, though yes, just talking about a topic without editor open is a refreshing change in how I usually write, so wanted to give it a go.
Exactly what I was thinking. Another Shill on the FP of HN in a week. This is getting old.
1. List of uses that are nothing but claims with zero tangible repeatable steps.
2. One vague statement about using it to bootstrap a machine... yeah lets see it.
I don't see any use cases for Claude code that I have not already tried with Goose or Dagger Agents. And I don't have to let those loose on my OS to use them.But hey the author made a zero proof claim that they can re-bootstrap with claude anyway...
Then at the end you are challenged to
"Got a crazier Claude workflow? Ping me @steipete."
IDK You didn't share any details beyond some vague use cases. Pretty much the same half working stuff from 2 years ago.
Yo, Steipete. Since claude is such a banger why don't you have it generate the config that you can share as proof. I'm not gonna wait though since its all hot air.
I don’t see how the OP is paying enough attention to know that terrible things are not happening. Sure, he can roll back, but that only works well for obvious problems that he catches in time.
I haven’t heard that it is inherently dangerous to use popular tools of the normal sort without carefully scrutinizes every output. But there are many horror stories about the unreliability of LLMs.
Backups should solve most of the issues. I've had some reservations, but if you think a bit about it, not that much can happen - if you have good backups! On macOS, your passwords are in the keychain and not accessible without putting your thumb on TouchID. So it cannot transfer all your money offshore, or order a thousand shoes.
Do make sure your AWS/fly/etc accounts have budget caps (a wise idea anyway).
> Automate Content: Like this very post. I use Wispr Flow to talk with Claude, explain the topic and tell it to read my past blog posts to write in my style.
This post is generated. Meaning, it wasn't written with the same aim towards truth and relevance that we assume writers have. It looks like writing, and fools readers into thinking it's writing.
But it's just text. What's the purpose in reading it as opposed to any other generated text?
If using LLMs for writing, you should provide your prompts up front so we can see your actual thinking, and then ignore the rest of the content. Or better yet, synthesize those prompts in a writing style that we like more!
I wanted to try Claude code but it's 200$ per month which is INR17000 per month. This was expensive even when I had a job. Now it feels like all this power is accessible only to the rich. When Geoffrey suggested [1] AI will increase wealth for rich he was not wrong.
You can try it with a regular dev account and it charges you per token. It's still costly, e.g. around $20 for an active hour of experimentation, but you can just run `/cost` at any point to see how much you've paid so far, and decide when to stop if it's not worth your money.
Yeah, $200/month makes sense if you're a contractor passing the cost to your client, but if you're a hobbyist or employee it's a massive chunk of money.
>> “commit everything in logical chunks” and Claude handles the entire flow—staging changes, writing meaningful commit messages, pushing, opening PRs, watching CI, and fixing any CI failures. When builds break, it analyzes the errors and patches them automatically. It’s also extremely good at resolving merge conflicts.
Is this marketing hype for the post or does it actually do all this? I know it's relatively new, but I haven't been able to find a resources to do such a setup? Only posts glorifying that it's capable of such tasks. If anyone else has had better luck with such complex tasks, would love to find out more.
It can do all this and more, you just have to nudge/instruct it to.
It's also amazing at using the CLI (I've been a terminal nerd for years, and yet watching claude 4 pipe those unix commands to work around the fact it can't use interactive TUI's yet, is pretty amazing)
You can give it tools either via funciton call, MCP, or just terminal unix shell.
Prompt it right, and the agentic stuff is like magic imho
Looks miserable. Prompt, tweak, review, test, tweak, prompt, review, test, review, tweak, prompt, etc. Fiddling with settings windows in some black box software that belongs to someone else.
Programming: a powerful skillset, one of the few things that can give you autonomy and agency besides money. Doable on basic computer hardware with no internet connection. Self-reinforcing, investing in yourself.
Prompting: requires expensive subscriptions to someone else's computer, someone else's software. They can turn it off, they control it. Spending your days QAing software and reviewing code and describing in worse language what you want the computer to do, over and over, until you test your way to something approximating what you wanted. Not to mention incinerating the environment.
All those “AI works great for me” blogs and HN comments are suspiciously light on details. They could just be marketing hype for all we know. I for one don't trust them one bit.
Have been watching him lean more into this over the past few weeks. It inspired me to try something similar with fantastic results. Watch what he does, try it with an open mind. You'll be pleasantly surprised
(1) That doesn't affect everybody. (2) If it does affect you there's a simple fix. (3) The existence of one minor migration bug doesn't really justify avoiding the entire approach (consider: how many issue exist in your ad hoc migration?)
> Automate Content: Like this very post. I use Wispr Flow to talk with Claude, explain the topic and tell it to read my past blog posts to write in my style.
If you had any respect for your readers you’d put this as a disclaimer right at the top. I can’t be bothered to read anything you can’t be bothered to write.
Then why not just write it? You mention not having to write Markdown in the post, but c’mon dude, you’re a software developer.
Presumably the fact you have a personal blog means you must enjoy some aspect of writing, so what’s the end game here?
I mean you do you I guess—it’s your blog and I’m nobody. But if I’m reading someone’s thoughts I want _their_ words, not something filtered through a machine.
Assuming this can write emails and access the internet, I would suggest also backing up all relationship states and account balances periodically as well.
If you're interested in seeing the process behind this piece of writing you can read through a lot of the details in the 71 commits that went into creating the story in the PR: https://github.com/steipete/steipete.me/pull/106/commits
I don't think that's true. What matters to me is the human editorial touch: I don't want to wade through 50 prompts and responses, I want a human author to have resolved that process into a final output that they think is worth sharing with me.
Try reading a manuscript copy of a book before it’s been edited. Yes I know some people do this out of interest but for most people it’s not the type of writing they are interested in reading or would get the most out of.
Taking the title literally, it also means that computer is not yours anymore, it's a rented appliance. And only works for you as long as you pay the rent. No doubt a wet dream come true for those who rent it out to you.
What matters to me is that a human writer has verified the content and is ready to stake their reputation on it being worth my time to read.
It sounds like that's what you've done here, in which case I don't feel that you are wasting my time by having me read something that you haven't even reviewed yourself.
It does seem weird for someone to expect others to spend their time fully reading something… when they also have access to the same tools and can just tell it to summarize it back.
That would be true for writing where the author typed a sentence and the LLM expanded it to multiple paragraphs.
That is not what happened here: the author provided a lot more input than the finished article, and used the LLM to help crunch that down to as good a version as possible of the points there wanted to make.
I've been trying to organize my thoughts of how I feel about consuming AI generated content. This comment really encapsulates how I feel.
As long as a human put time and effort into making something, then I'm willing to consider putting my effort toward reading/watching. If someone just spends 5 seconds to throw a prompt out there, that's when I get annoyed.
> As long as a human put time and effort into making something, then I'm willing to consider putting my effort toward reading/watching. If someone just spends 5 seconds to throw a prompt out there, that's when I get annoyed.
Why? Do you care more about the origin than the quality?
> Why? Do you care more about the origin than the quality?
Because they are linked. AI content can be generated so frivolously and at such volume you easily be overwhelmed by low quality garbage. Humans can also generate crap, but a much slower pace and I think that AI being so good at crap generation that it will push out any humans in the space that used to meek out any work here. So, what we are left with is AI content that is mostly low-effort crap, with maybe some rigorously reviewed bits that are good here and there, and the human-content which will mostly be people who care enough to make quality content otherwise they would be already posting AI schlock.
The end is that using AI as a proxy indicator for garbage will be right more than its wrong. So if I see something is AI generated, I should give it a pass and not waste my limited time resource on it.
My brother-in-law just accidentally stepped on my laptop and it literally broke into pieces. Fortunately, I have an extra laptop that I can use instead. Please recreate my environment on the new laptop.
This is pretty fascinating, but I am not brave (or foolhardy) enough to try it myself.
Having solid backups is good protection against errors that delete data, but I'm more worried about problems that leak my private data - especially to malicious attackers.
Think malicious instructions hidden on websites or in emails that cause Claude Code to steal API keys or account credentials and exfiltrate them to an attacker.
Right now, it is unlikely that such prompt injection attacks are being targeted against specific individuals. I know they are theoretically possible though, and that's enough to prevent me from trying this out.
I am quite honestly baffled by all these posts. I'm not a hardcore skeptic, at least I don't think so, I have tried most of the vibe coding tools, and I regularly try new ones and continue using many; and I do NOT have the experience described in all these posts.
So two possibilities come to mind:
1. I, somehow, use every single one of those tools wrong.
2. Or; you and the other people raving about Claude code are solving trivial problems.
Of course, I'm human so I'm inclined to believe I do not use all tools wrong. But reading your experience and those similar to yours I can't help but think two things:
- you are solving incredibly rote, already solved, problem that the LLMs are just able to spit out by heart. I do too, and yet I don't have the same AI successes so damn your problems must be especially trivial.
- You are so wildly overpaid that paying hundreds (I even saw people "boasting" about spending thousands a day !) of dollar a month is perfectly justifiable.
I don't think I'm wrong on the last point, and if I can see it then employers will be able to see it too. Why on earth would anyone pay 600$/h (your rates according to one of your post), when apparently I can just pay 200$/mo and get the same thing?
And don't tell me "because it takes skill to use Claude Code", you just wrote a blog post telling it mostly doesn't (apparently).
Don't take this whole post as a mean criticism (and if it reads like it, I'm sorry), I am just truly flabergasted.
Every time I see a post like yours, I see the reply saying "yeah you're using it wrong". I can't say either 1 or 2 is true, but if you saw the recent post[0] about cloudflare making an API entirely with Claude, it could be more of a case of not using the tools to their potential.
Sometimes you have to learn how to frame the problem in a way to get the results that you want. These tools need lots of context, not just about the rest of the code base but the problem itself. You can think of it a bit like how the early adopters of high level programming languages had to fight against compilers to get the assembly output that they wanted.
For example, if I tell an LLM to generate a python script that finds the square of a number I might want:
def square(x):
return x * x
but it may give me:
print("Enter a number:")
x = int(input())
print("The square is", x * x)
This is a very very simple example but I think it illustrates my point. If you provide enough context to the exact problem you want to solve the results are astronomically better.
I feel like I’m beginning to see what countless others have doubtless discovered before me, which is that much of the effort and ingenuity of a modern life is directed toward figuring out how to live as little as possible.
108 comments
[ 5.0 ms ] story [ 194 ms ] threadI purposely had it on YOLO. Everything was backed up idc really. Its been a bug with claude on cursor getting around permissions. It was the day claude 4 came out and I wanted to see what it could do with YOLO. It was kinda funnny but annoying, could be catastrophic for someone who doesn't back up everything obsessively.
https://news.ycombinator.com/item?id=44097108
How recently did this occur? What did you ask it in your prompt? What did it say as the explanation for why it was doing that?
Spoke like a man who really, really takes that Roko thing seriously.
I'm a little envious of how all that must feel. Must be a real trip.
If you have reasonable guard rails for your AI, you can throw much more at it.
Let it write tests before implementing something, ensure the code is in version control before promoting a change, etc.
But the constant Git churn created a lot of traffic and collisions in that folder. And when I googled it, apparently it's just not a good idea to sync Git projects in iCloud.
I really want to easily sync my folders so that I can just get a new macbook and recreate my environment.
Anyone have any tips?
you might copy the code working dir in read mode only or whatever idk\",
I've found migrating ~/library over is problematic even with built in migration assistant, especially if this is their third or later machine being transfered between.
Git is designed with a remote server in mind, just do it that way. Get github or git lab or setup your own remote host.
I have my code folders in Dropbox on OSX but I believe my process should work for you. I work in node and have a number of monorepos with nested node_modules folders. I run my idempotent dev-files.command script [2] with or without "--nono" in any monorepo and it handles the entire git aliasing and dropbox folder ignore for node_modules process. I use --nono if I ever need to clear out all of the node_modules and do a fresh pnpm install.
1. https://git-scm.com/docs/gitrepository-layout
2. https://gist.github.com/chiragmehta/a41bd33356b6a2f84075d23d...
The downside is that git history is pretty useful, so it would suck to not sync it.
But you gave a real solution to the problem.
Edit: I remember now that one problem with iCloud was that doing something like switching Git branches or checking out commits causes a catastrophic sync cascade in the code folder itself.
Maybe Dropbox is better at dealing with that kind of thing in general. But that's pretty annoying for any syncing system to deal with, heh.
I guess the simplest solution is to just get into the habit of pushing every little project to github.
I tried the same with Google Drive at work and had lots of issues, but recently it seems to have stabilized.
These are serious costs.
Are... are you serious? Markdown formatting? The formatting standard that's designed to be essentially plain text?
1. List of uses that are nothing but claims with zero tangible repeatable steps.
2. One vague statement about using it to bootstrap a machine... yeah lets see it.
I don't see any use cases for Claude code that I have not already tried with Goose or Dagger Agents. And I don't have to let those loose on my OS to use them.But hey the author made a zero proof claim that they can re-bootstrap with claude anyway...
Then at the end you are challenged to
"Got a crazier Claude workflow? Ping me @steipete."
IDK You didn't share any details beyond some vague use cases. Pretty much the same half working stuff from 2 years ago.
Yo, Steipete. Since claude is such a banger why don't you have it generate the config that you can share as proof. I'm not gonna wait though since its all hot air.
I don’t see how the OP is paying enough attention to know that terrible things are not happening. Sure, he can roll back, but that only works well for obvious problems that he catches in time.
I haven’t heard that it is inherently dangerous to use popular tools of the normal sort without carefully scrutinizes every output. But there are many horror stories about the unreliability of LLMs.
Do make sure your AWS/fly/etc accounts have budget caps (a wise idea anyway).
This post is generated. Meaning, it wasn't written with the same aim towards truth and relevance that we assume writers have. It looks like writing, and fools readers into thinking it's writing.
But it's just text. What's the purpose in reading it as opposed to any other generated text?
If it was dictated or transcribed, it's somehow lesser and unworthy of attention?
No. I don't know where you got that from.
https://www.indiatoday.in/technology/news/story/ai-will-incr...
At the end of the day all do the same. a cli tool that can call tools and use them as needed.
https://github.com/openai/codex http://aider.ai/
Is this marketing hype for the post or does it actually do all this? I know it's relatively new, but I haven't been able to find a resources to do such a setup? Only posts glorifying that it's capable of such tasks. If anyone else has had better luck with such complex tasks, would love to find out more.
It's also amazing at using the CLI (I've been a terminal nerd for years, and yet watching claude 4 pipe those unix commands to work around the fact it can't use interactive TUI's yet, is pretty amazing)
You can give it tools either via funciton call, MCP, or just terminal unix shell.
Prompt it right, and the agentic stuff is like magic imho
Here you go
Programming: a powerful skillset, one of the few things that can give you autonomy and agency besides money. Doable on basic computer hardware with no internet connection. Self-reinforcing, investing in yourself.
Prompting: requires expensive subscriptions to someone else's computer, someone else's software. They can turn it off, they control it. Spending your days QAing software and reviewing code and describing in worse language what you want the computer to do, over and over, until you test your way to something approximating what you wanted. Not to mention incinerating the environment.
I know which of the two I prefer.
its nit really choosing one ot the other imho
they supplement each other
Or just use Migration Assistant?
If you had any respect for your readers you’d put this as a disclaimer right at the top. I can’t be bothered to read anything you can’t be bothered to write.
Presumably the fact you have a personal blog means you must enjoy some aspect of writing, so what’s the end game here?
I mean you do you I guess—it’s your blog and I’m nobody. But if I’m reading someone’s thoughts I want _their_ words, not something filtered through a machine.
It sounds like that's what you've done here, in which case I don't feel that you are wasting my time by having me read something that you haven't even reviewed yourself.
That is not what happened here: the author provided a lot more input than the finished article, and used the LLM to help crunch that down to as good a version as possible of the points there wanted to make.
As long as a human put time and effort into making something, then I'm willing to consider putting my effort toward reading/watching. If someone just spends 5 seconds to throw a prompt out there, that's when I get annoyed.
Why? Do you care more about the origin than the quality?
Because they are linked. AI content can be generated so frivolously and at such volume you easily be overwhelmed by low quality garbage. Humans can also generate crap, but a much slower pace and I think that AI being so good at crap generation that it will push out any humans in the space that used to meek out any work here. So, what we are left with is AI content that is mostly low-effort crap, with maybe some rigorously reviewed bits that are good here and there, and the human-content which will mostly be people who care enough to make quality content otherwise they would be already posting AI schlock.
The end is that using AI as a proxy indicator for garbage will be right more than its wrong. So if I see something is AI generated, I should give it a pass and not waste my limited time resource on it.
My brother-in-law just accidentally stepped on my laptop and it literally broke into pieces. Fortunately, I have an extra laptop that I can use instead. Please recreate my environment on the new laptop.
Having solid backups is good protection against errors that delete data, but I'm more worried about problems that leak my private data - especially to malicious attackers.
Think malicious instructions hidden on websites or in emails that cause Claude Code to steal API keys or account credentials and exfiltrate them to an attacker.
Right now, it is unlikely that such prompt injection attacks are being targeted against specific individuals. I know they are theoretically possible though, and that's enough to prevent me from trying this out.
So two possibilities come to mind:
1. I, somehow, use every single one of those tools wrong.
2. Or; you and the other people raving about Claude code are solving trivial problems.
Of course, I'm human so I'm inclined to believe I do not use all tools wrong. But reading your experience and those similar to yours I can't help but think two things:
- you are solving incredibly rote, already solved, problem that the LLMs are just able to spit out by heart. I do too, and yet I don't have the same AI successes so damn your problems must be especially trivial.
- You are so wildly overpaid that paying hundreds (I even saw people "boasting" about spending thousands a day !) of dollar a month is perfectly justifiable.
I don't think I'm wrong on the last point, and if I can see it then employers will be able to see it too. Why on earth would anyone pay 600$/h (your rates according to one of your post), when apparently I can just pay 200$/mo and get the same thing?
And don't tell me "because it takes skill to use Claude Code", you just wrote a blog post telling it mostly doesn't (apparently).
Don't take this whole post as a mean criticism (and if it reads like it, I'm sorry), I am just truly flabergasted.
Every time I see a post like yours, I see the reply saying "yeah you're using it wrong". I can't say either 1 or 2 is true, but if you saw the recent post[0] about cloudflare making an API entirely with Claude, it could be more of a case of not using the tools to their potential.
[0] https://news.ycombinator.com/item?id=44159166
I use these tools a lot and I would say they fail 50% of the time for my use cases and I baby them every step
”Is it just me” is such a flawed mindset. The reality is staring you in the face
Sometimes you have to learn how to frame the problem in a way to get the results that you want. These tools need lots of context, not just about the rest of the code base but the problem itself. You can think of it a bit like how the early adopters of high level programming languages had to fight against compilers to get the assembly output that they wanted.
For example, if I tell an LLM to generate a python script that finds the square of a number I might want: def square(x): return x * x
but it may give me: print("Enter a number:") x = int(input()) print("The square is", x * x)
This is a very very simple example but I think it illustrates my point. If you provide enough context to the exact problem you want to solve the results are astronomically better.