I am having trouble relating to this post as someone outside of the larger techno-blob. I have a 9-5-type job which for many reasons I would not be using Claude for. In my home life my hobbies only involve computers for gaming. What would I use Claude Code for?
imagine a Linux nerd who is highly skilled at using computers from the command line (though still has enough graphical access to visually view documents/pictures to some extent).
for a normal person, the GUI is better for many tasks. however, for almost anything you can do in a GUI, there's a way for the Linux nerd to get the same job done from the terminal.
what jobs do you want to get done on your computer, and what could this person do for you if you sat them down at your computer and asked?
Yeah, I mean I don't have any jobs in mind that I'd want to get done on my computer. But the sense I get from reading posts like this is that there's actually a ton of value I could unlock doing new kinds of jobs on my computer that I wouldn't have done before, and I'm having trouble thinking of any specific ways that Claude Code would make my life better. But it also seems crazy that such a powerful tool would actually be completely valueless to me such that I wouldn't even bother to set it up. So I feel a bit confused and would love some help trying to think through this confusion.
Yeah I don't really do any of those things. Outside of work I mostly cook and clean, lift weights, play sports, watch sports, manage my kids, play videogames, read books, watch movies, and try to get together with my friends and extended family as often as I can.
You're the most human human right now. My wife is the same way. It brings me joy to see others find joy in the normal things in life.
Honestly, most of the people you're reading here have a constant drive for learning and improvement that often derails or leads us to undervalue that most normal and human way of living.
Or maybe I'm just projecting.
Anyway, I'm trying to say I am happy for you. Keep living the good life!
The biggest thing for me has been to realize when I have an idea but immediately dismiss it as 'not worth the time,' I should instead take 30 seconds and see if claude code can just do it for me. Like, I would not know how to set up alerts to notify my if there was an unusually good deal on DDR5 ram for a new build, but maybe I could just ask claude to do it for me?
An example of this is that I was looking through a list of people seeking collaborators to work on projects, which had >150 people/projects. I had the thought that it would be nice to search this specifically for the type of thing I wanted to do. I asked claude to scrape it (I wouldn't have known how to do this, and copy/paste wasn't an option), and make an ordered list of 10 projects that were the best fit based on my CV and a short description of what I was looking for. From this list it captured all of the projects I'd found from brief skimming, plus like 5 additional ones I'd missed but seemed great. I also agreed with it's priorities.
It's not like I couldn't have found those one my own, I just wouldn't have because I didn't want to spend 2 hours reading all the projects. Instead it took 10 minutes. My suspicion is that there are maybe 1-5 ideas like this every day that would improve my life, but I'm still learning how to notice them and pretty novice at claude code.
I've said exactly the same thing several times, and everytime I look into it I find the "EVERYONE SHOULD USE CLAUDE CODE" clan completely full of hype.
I'm using Claude as my biographer. Dump whatever I want into Obsidian, let it clean up and organize and prioritize/ask questions so I can keep building the vault out.
Although I've started putting sensitive information in there so recently had Claude set up gpt-oss 120B. Only the local model reads my information but Claude manages the system.
Claude Code is great for a lot of things, but like other AIs today it is very jagged. Images in particular are a huge tripping point, in my experience. It's shocking to me since Opus 4.5 almost never hallucinates in text, but in images and ESPECIALLY series of related images, it will "see" things that are not there, reference the wrong image to support a claim, or just straight-up make things up that are not in any of the images.
I do think a lot of the twitter hype is about what's essentially "markdown doc todo-list workslop" and making a nice UI for your todo list is baby-tier easy for Claude Code, so it's easy to see why many people are so hyped. But for real software that needs to be correct, I still have to do a lot of hand-holding and cross-checking.
But hey, remind me where we were a year ago? Buckle up, 2026 is going to be a wild ride.
I don’t know how credible they are. It basically looks similar to the existing “Code” feature in the desktop app, just repurposed for general computer tasks. I hope it’s real because I’d like to use it.
"Claude Code built in an hour what took a Google team a year." feels like the original, _un_stretched, ai 2027 timeline. We seem to have agent-1, and we seem to be at, or nearly at, having the supercoder.
Would you still post your articles to Twitter if they hadn't throttled Substack links? It seems so weird conceptually to have 20k word effortposts on the wobsite that used to be about 140/280 characters of lossy compression. Instrumental convergence of media platforms...
Non-standard precision is part of what makes me tentative about getting back into coding (although the field as a whole seems far easier to learn today than when I was bashing my head against Ruby On Rails bootcamp decade+ ago). The most "labour-intensive" computing I do is editing music for personal taste...and that's rather hard to define in machine-interpretable terms. Even a Heuristic That Almost Always Works like "normalize the volume so no section is too quiet or too loud" isn't really correct, since different kinds of sound are perceived as softer or louder than the waveforms strictly suggest, and in some pieces the contrast is critical to the integrity of the art. Like I'm just imagining Claude struggling to remaster a Boris album...not literally outside p-space, just a very narrow and moving target for each listener (and audio setup).
Claude Code with Opus 4.5 is amazing. Codex 5.2 is also fantastic. I find that Codex 5.2 is just straightforwardly better for solving hard problems, but Claude Code is much nicer to use and easier to synergize with.
> So far I’ve been working on two coding projects. I’ve been using the terminal over the web interface on the ‘skill up at doing this before you reject it’ theory and it’s been mostly fine although I find editing my prompts annoying.
VSCode accessing the CLI through its GUI integration extension largely fixes the edit stuff (I hate vi with passion, so share the annoyance). It will force you to use the bare CLI occasionally, I've mostly seen this when messing with MCP config.
The big unsolved issue is sandboxing. I really do not want AI agents (Claude or others) to have full access to my desktop right now and letting it access personal cloud services is dicey, too.
Francis Bacon had this to say about reading 400 odd years ago: Some books are to be tasted, others to be swallowed, and some few to be chewed and digested; that is some books are to be read only in parts; others to be read but not curiously; and some few to be read wholly, and with diligence and attention. Some books may also be read by deputy, and extracts made of them by others.
In these claims of hyper AI productivity, I detect a whiff of those notes experts whose systems for categorization were going to unlock the secrets of the universe. But in the end these guys only seemed to write articles about taking notes.
Today, I believe AI allows everyone, even the non-programmer (especially the non-programmer?) to jump one or two rungs up the IT ladder. But serious, quality software still takes time. Less time now, certainly, but time nonetheless. When anyone claims to be as efficient as one thousand Google engineers, I think it fair to ask where his equivalent of Google is.
I've been increasingly tempted to take the plunge in the past few weeks. I fairly routinely use LLMs as StackOverflow substitutes these days, but this is very targeted non-agentic work. As a scientific software developer, it makes me very uneasy to push code that I haven't fully read or understood, and therefore may not be able to update or maintain. And reading someone else's code is much less enjoyable than writing one's own code.
Entirely Claude Code phenomenon? The hype maybe, but not the substance. Early adopters did it with Codex too, since GPT-5.1 probably. Especially given the more generous OpenAI quotas. Big users burn lots of tokens - with Anthropic it quickly gets both expensive and restrictive. True that Claude Code won the popularity contest a majority agree on. Doesn't mean it's the only one, or even the 1st one.
I think the main benefit that CC currently enjoys is that there is a lot of experimentation resulting in tons of plugins and skills - most of which would be applicable to Codex too but not always easily installed. I am not convinced the harness itself is on a fundamentally different level or the LLM, for that matter. Compared like for like, the OpenAI models (codex or regular gpt) are definitely trading blows with Anthropic (see https://swe-rebench.com/) and Gemini is not all that far off, either.
However, the explosion of skills makes it tricky (impossible?) to actually converge on best practices right now.
True that. Yeah I agree it's the mind share that matters. Even people in AI and writing s/w for living - I was surprised to find them not that interested in 'agent writing code independently' back in Oct. For many Fill-In-the-Middle in Cursor sufficed.
I tend to think that 1/2 of the smarts is in Claude Code, and the other 1/2 is in the Opus model. For C/C++ jobs I actually found Codex 5.2-high do better for me. When paused on CC, I used alternative Zai API. Their glm-4.7 seemed between Sonnet and Opus. While the quota was multiples more generous. (switching forth/back is 3 vars/lines add/delete in ~/.claude/setting.json or shell env)
Since I've seen OpenRouter offering PAYG models under CC too - openrouter .ai/docs/guides/guides/claude-code-integration, but have not tried it. PAYG is more expensive than fixed price plans, but sometimes there are freebies to try. Or to try-before-you-buy in the exact agent (CC) one will use the model in.
IDK how much in keeping with Anthropic CC T&C is this underlying model/api back and forth. I presume they will crack down if it becomes widely used.
Lately I get the impression that opencode maybe as good as CC. I try it from time to time and seems to work, but have not needed to use it fr on a real project. Opencode is api/model agnostic, looks to me as long as the model is trained with 'interleaved thinking and tools use' (latest such being minimax-m2.1), it should 'just work' comparable to CC.
Opencode bonus point - one can plugin local llm too. Without any special gymnastics with configs and proxies, local is fully supported. Big local model is too still too slow for me (on aging m2 mbp ultra-quantised 55gb minimax-m2.1 is 10x slower than their api). But good to have local llm an option in the background for dependency risk reduction.
I've been using GLM a fair bit - seems to work best on CC. It sometimes get stuck, and isn't very good at debugging but usually one of the frontier LLMs can quickly fix it. Using multiple ones is generally helpful anyhow, they all eventually get stuck.
That would be a strong point toward opencode (or maybe codex-cli) as switching is easier than the .env hoops with CC. Local LLM is not really a consideration for me, the hardware needed for a decent one is a fundamentally unsound investment and the speed remains bad. But multi provider for open weights is a definite plus.
FWIW, if you want model selection on a plan, there is always chutes or nano-gpt....
Saved and will review thanks. I used chatgpt to help me create a matlab-like interface to a Raspberry Pi so I can drive a 3-axis mill using L-Sysytem scripts (fractal grammars) to execute CNC commands. I am pulling that back up and will take a look at Claude to help write the remaining code.
“When you hit auto-compact, Claude Code does its best to condense the prior conversation and keep going, but you will lose important context. Daniel San disabled auto-compaction for this reason, instead choosing to restart sessions if and when limits get hit.”
it’s a frustrating thing either way. context drift is real, and i appreciate anthropics approach over openai’s. chatgpt just pretends like it remember and it makes hallucinations worse.
I am having trouble relating to this post as someone outside of the larger techno-blob. I have a 9-5-type job which for many reasons I would not be using Claude for. In my home life my hobbies only involve computers for gaming. What would I use Claude Code for?
imagine a Linux nerd who is highly skilled at using computers from the command line (though still has enough graphical access to visually view documents/pictures to some extent).
for a normal person, the GUI is better for many tasks. however, for almost anything you can do in a GUI, there's a way for the Linux nerd to get the same job done from the terminal.
what jobs do you want to get done on your computer, and what could this person do for you if you sat them down at your computer and asked?
Yeah, I mean I don't have any jobs in mind that I'd want to get done on my computer. But the sense I get from reading posts like this is that there's actually a ton of value I could unlock doing new kinds of jobs on my computer that I wouldn't have done before, and I'm having trouble thinking of any specific ways that Claude Code would make my life better. But it also seems crazy that such a powerful tool would actually be completely valueless to me such that I wouldn't even bother to set it up. So I feel a bit confused and would love some help trying to think through this confusion.
What about jobs to do on your phone?
Do you ever mod games or tweak performance settings etc. in complex ways?
Do you have any “smart home” devices?
Do you ever find yourself dealing with spreadsheets and documents for non-work purposes?
Possibly you just have a very healthy life where work stays at work and home life is mostly leisurely?
Yeah I don't really do any of those things. Outside of work I mostly cook and clean, lift weights, play sports, watch sports, manage my kids, play videogames, read books, watch movies, and try to get together with my friends and extended family as often as I can.
Sounds like you're luckier than most guys hustling over Claude code heh. Stay blessed my friend
You're the most human human right now. My wife is the same way. It brings me joy to see others find joy in the normal things in life.
Honestly, most of the people you're reading here have a constant drive for learning and improvement that often derails or leads us to undervalue that most normal and human way of living.
Or maybe I'm just projecting.
Anyway, I'm trying to say I am happy for you. Keep living the good life!
The biggest thing for me has been to realize when I have an idea but immediately dismiss it as 'not worth the time,' I should instead take 30 seconds and see if claude code can just do it for me. Like, I would not know how to set up alerts to notify my if there was an unusually good deal on DDR5 ram for a new build, but maybe I could just ask claude to do it for me?
An example of this is that I was looking through a list of people seeking collaborators to work on projects, which had >150 people/projects. I had the thought that it would be nice to search this specifically for the type of thing I wanted to do. I asked claude to scrape it (I wouldn't have known how to do this, and copy/paste wasn't an option), and make an ordered list of 10 projects that were the best fit based on my CV and a short description of what I was looking for. From this list it captured all of the projects I'd found from brief skimming, plus like 5 additional ones I'd missed but seemed great. I also agreed with it's priorities.
It's not like I couldn't have found those one my own, I just wouldn't have because I didn't want to spend 2 hours reading all the projects. Instead it took 10 minutes. My suspicion is that there are maybe 1-5 ideas like this every day that would improve my life, but I'm still learning how to notice them and pretty novice at claude code.
I've said exactly the same thing several times, and everytime I look into it I find the "EVERYONE SHOULD USE CLAUDE CODE" clan completely full of hype.
I run a tech company, and I see no use for it.
I was taught programming by a guy named Shannon Code so I can't help but see Claude Code ironically
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/claude-codes
I'm using Claude as my biographer. Dump whatever I want into Obsidian, let it clean up and organize and prioritize/ask questions so I can keep building the vault out.
Although I've started putting sensitive information in there so recently had Claude set up gpt-oss 120B. Only the local model reads my information but Claude manages the system.
Claude Code is great for a lot of things, but like other AIs today it is very jagged. Images in particular are a huge tripping point, in my experience. It's shocking to me since Opus 4.5 almost never hallucinates in text, but in images and ESPECIALLY series of related images, it will "see" things that are not there, reference the wrong image to support a claim, or just straight-up make things up that are not in any of the images.
I do think a lot of the twitter hype is about what's essentially "markdown doc todo-list workslop" and making a nice UI for your todo list is baby-tier easy for Claude Code, so it's easy to see why many people are so hyped. But for real software that needs to be correct, I still have to do a lot of hand-holding and cross-checking.
But hey, remind me where we were a year ago? Buckle up, 2026 is going to be a wild ride.
There are rumors based on digging around in the Claude app of something called “Task Mode”.
https://www.testingcatalog.com/exclusive-early-look-at-claude-tasks-mode-agent-from-anthropic/
I don’t know how credible they are. It basically looks similar to the existing “Code” feature in the desktop app, just repurposed for general computer tasks. I hope it’s real because I’d like to use it.
I’ve been enjoying my Claude Code workflows (“Qs”) for the last two months.
I particularly like my $0/month omnipotent virtual assistant…🤩
Download them all from the Q-Marketplace (if you’re a cool kid, that is…)
the-ai-masters.com/q-marketplace
I built a Claude Code version of Nate B. Jones' Second Brain system.
Instead of Zapier + Notion, it uses:
- Slack as your inbox (capture from anywhere, even your phone)
- Claude Code as the sorter/filer (runs when you're working)
- Local YAML files instead of Notion
Same principles: one inbox, zero decisions at capture time, AI classification, fix button for corrections.
It's session-based (runs when you open Claude Code) rather than always-on, but the core ideas translate perfectly.
DM me if you want the template - happy to share.
"Claude Code built in an hour what took a Google team a year." feels like the original, _un_stretched, ai 2027 timeline. We seem to have agent-1, and we seem to be at, or nearly at, having the supercoder.
Would you still post your articles to Twitter if they hadn't throttled Substack links? It seems so weird conceptually to have 20k word effortposts on the wobsite that used to be about 140/280 characters of lossy compression. Instrumental convergence of media platforms...
Non-standard precision is part of what makes me tentative about getting back into coding (although the field as a whole seems far easier to learn today than when I was bashing my head against Ruby On Rails bootcamp decade+ ago). The most "labour-intensive" computing I do is editing music for personal taste...and that's rather hard to define in machine-interpretable terms. Even a Heuristic That Almost Always Works like "normalize the volume so no section is too quiet or too loud" isn't really correct, since different kinds of sound are perceived as softer or louder than the waveforms strictly suggest, and in some pieces the contrast is critical to the integrity of the art. Like I'm just imagining Claude struggling to remaster a Boris album...not literally outside p-space, just a very narrow and moving target for each listener (and audio setup).
Claude Code with Opus 4.5 is amazing. Codex 5.2 is also fantastic. I find that Codex 5.2 is just straightforwardly better for solving hard problems, but Claude Code is much nicer to use and easier to synergize with.
> So far I’ve been working on two coding projects. I’ve been using the terminal over the web interface on the ‘skill up at doing this before you reject it’ theory and it’s been mostly fine although I find editing my prompts annoying.
VSCode accessing the CLI through its GUI integration extension largely fixes the edit stuff (I hate vi with passion, so share the annoyance). It will force you to use the bare CLI occasionally, I've mostly seen this when messing with MCP config.
The big unsolved issue is sandboxing. I really do not want AI agents (Claude or others) to have full access to my desktop right now and letting it access personal cloud services is dicey, too.
Francis Bacon had this to say about reading 400 odd years ago: Some books are to be tasted, others to be swallowed, and some few to be chewed and digested; that is some books are to be read only in parts; others to be read but not curiously; and some few to be read wholly, and with diligence and attention. Some books may also be read by deputy, and extracts made of them by others.
In these claims of hyper AI productivity, I detect a whiff of those notes experts whose systems for categorization were going to unlock the secrets of the universe. But in the end these guys only seemed to write articles about taking notes.
Today, I believe AI allows everyone, even the non-programmer (especially the non-programmer?) to jump one or two rungs up the IT ladder. But serious, quality software still takes time. Less time now, certainly, but time nonetheless. When anyone claims to be as efficient as one thousand Google engineers, I think it fair to ask where his equivalent of Google is.
I've been increasingly tempted to take the plunge in the past few weeks. I fairly routinely use LLMs as StackOverflow substitutes these days, but this is very targeted non-agentic work. As a scientific software developer, it makes me very uneasy to push code that I haven't fully read or understood, and therefore may not be able to update or maintain. And reading someone else's code is much less enjoyable than writing one's own code.
Entirely Claude Code phenomenon? The hype maybe, but not the substance. Early adopters did it with Codex too, since GPT-5.1 probably. Especially given the more generous OpenAI quotas. Big users burn lots of tokens - with Anthropic it quickly gets both expensive and restrictive. True that Claude Code won the popularity contest a majority agree on. Doesn't mean it's the only one, or even the 1st one.
I think the main benefit that CC currently enjoys is that there is a lot of experimentation resulting in tons of plugins and skills - most of which would be applicable to Codex too but not always easily installed. I am not convinced the harness itself is on a fundamentally different level or the LLM, for that matter. Compared like for like, the OpenAI models (codex or regular gpt) are definitely trading blows with Anthropic (see https://swe-rebench.com/) and Gemini is not all that far off, either.
However, the explosion of skills makes it tricky (impossible?) to actually converge on best practices right now.
True that. Yeah I agree it's the mind share that matters. Even people in AI and writing s/w for living - I was surprised to find them not that interested in 'agent writing code independently' back in Oct. For many Fill-In-the-Middle in Cursor sufficed.
I tend to think that 1/2 of the smarts is in Claude Code, and the other 1/2 is in the Opus model. For C/C++ jobs I actually found Codex 5.2-high do better for me. When paused on CC, I used alternative Zai API. Their glm-4.7 seemed between Sonnet and Opus. While the quota was multiples more generous. (switching forth/back is 3 vars/lines add/delete in ~/.claude/setting.json or shell env)
Since I've seen OpenRouter offering PAYG models under CC too - openrouter .ai/docs/guides/guides/claude-code-integration, but have not tried it. PAYG is more expensive than fixed price plans, but sometimes there are freebies to try. Or to try-before-you-buy in the exact agent (CC) one will use the model in.
IDK how much in keeping with Anthropic CC T&C is this underlying model/api back and forth. I presume they will crack down if it becomes widely used.
Lately I get the impression that opencode maybe as good as CC. I try it from time to time and seems to work, but have not needed to use it fr on a real project. Opencode is api/model agnostic, looks to me as long as the model is trained with 'interleaved thinking and tools use' (latest such being minimax-m2.1), it should 'just work' comparable to CC.
Opencode bonus point - one can plugin local llm too. Without any special gymnastics with configs and proxies, local is fully supported. Big local model is too still too slow for me (on aging m2 mbp ultra-quantised 55gb minimax-m2.1 is 10x slower than their api). But good to have local llm an option in the background for dependency risk reduction.
I've been using GLM a fair bit - seems to work best on CC. It sometimes get stuck, and isn't very good at debugging but usually one of the frontier LLMs can quickly fix it. Using multiple ones is generally helpful anyhow, they all eventually get stuck.
That would be a strong point toward opencode (or maybe codex-cli) as switching is easier than the .env hoops with CC. Local LLM is not really a consideration for me, the hardware needed for a decent one is a fundamentally unsound investment and the speed remains bad. But multi provider for open weights is a definite plus.
FWIW, if you want model selection on a plan, there is always chutes or nano-gpt....
FWIW, seems like OpenAI now allows using subscriptions with third party agents https://xcancel.com/thdxr/status/2009803906461905202
That definitely moves me toward resubscribing to Plus and towards opencode or maybe back to roo which supposedly will get it real soon now (TM).
Wish Google did the same as I have Gemini anyhow. Between those two, I feel little need for Claude...
Saved and will review thanks. I used chatgpt to help me create a matlab-like interface to a Raspberry Pi so I can drive a 3-axis mill using L-Sysytem scripts (fractal grammars) to execute CNC commands. I am pulling that back up and will take a look at Claude to help write the remaining code.
“When you hit auto-compact, Claude Code does its best to condense the prior conversation and keep going, but you will lose important context. Daniel San disabled auto-compaction for this reason, instead choosing to restart sessions if and when limits get hit.”
it’s a frustrating thing either way. context drift is real, and i appreciate anthropics approach over openai’s. chatgpt just pretends like it remember and it makes hallucinations worse.