> but rather prompting very small tasks after you have made the initial design, something along the lines of "Create a database migration that adds column [x], not null", "add a field to the this form that will accept string value, validate it on the backend against this format, and save it to column [x]". So basically, writing code in plain English.
I find such tasks easier to just do in the programming language directly than describe them in English.
> I believe that writing something yourself helps you learn it better, but at the same time, if you did all the thinking and now just prompting for a fancy auto-complete, we could argue that the hardest part (the design) was already done, and you skip the writing for the sake of writing. On the other hand, I am afraid that if I will stop writing code in environments where I can still do it, I will lose the ability to think about code and understand code.
Merely completing the "design" is by far not "doing all the thinking," and this is really the core of the problem. When people code with LLMs, the LLMs are making nontrivial decisions and thus are replacing part of the programmer's intelligence. That's why calling LLMs "fancy autocomplete" is so misleading, because autocompletion isn't supposed to make decisions, just to facilitate typing out the code you already have in your head.
So yeah, I agree with the last sentence about losing the ability to understand code, and that's a sufficient reason to avoid AI assistance whenever possible. Since I prioritize education and learning above almost everything else, it's not a close call. Where the line gets blurry for me is when people say "I would never have learned the tech stack needed to make this tool myself, but I vibe-coded it and now it's mine." It's really difficult for me to argue with that, to tell someone simply to live without something that's potentially improving their day-to-day life. To be clear, there are many good reasons not to do that thing, but it's also not my place to judge people for doing it, since I don't live my own life perfectly either.
It's like any other tool. Use it when you need it.
We don't insist that HVAC installers use hand-powered drills to "feel closer to the work," but we also don't insist that they use battery-powered drills and impact drivers on every single task.
We don't insist that a chef use a Vitamix blender to chop a single onion when a couple seconds with a sharp knife will do, but we certainly wouldn't only use a sharp knife to create a salsa.
We don't insist that a carpenter use only hand-saws to build a house, but we also don't insist that they use a chainsaw to make a chair.
We don't insist that an accountant write out taxes by hand, but we also don't insist that they use a calculator to determine how old you are from your birthdate.
You use the power tools when appropriate, and you use more delicate, surgical tools when appropriate. It's not any more difficult than that.
The problem with this analogy is an LLM isn't very much like a "tool" at all.
- In cases of practical relevance, you pay a recurring fee to keep using it - not very tool-like.
- Rather than doing a predictable action in response to user inputs, you give it imprecise instructions and it does something unpredictable but vaguely correct in response, which then must be thoroughly inspected.
- In cases of practical relevance, its actual operation (not just design) is entirely handled by a third party, not by the user.
These things describe a subcontractor, not a tool. We don't insist that a chef at a restaurant use this tool or that tool, but if the restaurant was actually just outsourcing all its cooking to some other company...yeah, that would absolutely factor into my decision to go there or not.
Imagine a carpenter needed to be proficient in hand-saws to use their chainsaw well, and using their chainsaw required knowing details about the wood they were working with, in only the way using a hand-saw gave.
...and it's fine to not use the loud, vibrating power saw and instead take some extra time to use a handsaw, and thus get more practiced at it.
It's fine to hire movers, but sometimes they end up damaging stuff. My solution is have a lot less stuff, and move it myself.
I increasingly have more respect for the Amish community one of my relatives lives near, at least in terms of their considered, intentional limits set on resource use. I don't agree with all their choices, and some seem arbitrary, but I respect their autonomy to choose.
I choose to move around the world primarily under my own power or augmented by a battery (supplies 2/3 of the Wattage I need to make my cargo bicycle move); harm reduction in the face of fossil-fuel addiction. I choose not to use generative AI because I'd rather human do more work for themselves instead of selling short future humans with profligate resource consumption.
It's fine to strive for things, but the most important thing in my life is my relationships with other humans. If I'm a person who can't walk, and am in physical therapy towards being able to walk, it's fine to keep working at it, and it's fine to never be able to walk; what matters is how I get on with my mates, and spending massive amounts of electricity on data centers is arguably harmful to the health of interpersonal relationships.
I really like your analogy as-is, but LLMs aren't reliable power tools. Asking for external help requires good communication and can be overkill for small tasks.
Hiring a team of subcontractors carries risks unless you babysit the whole project.
I still think I haven’t gotten much more productive than a year ago:
I wrote the barebones, scaffolding, and where I cared about the details. Agents stamped out the rest from those patterns.
Everyone speaks about this stuff in anecdotes. Or heroic attempts to have agent only projects.
But very little actual data on real gains in actual quality as experienced by users? Harder to measure IMO. And may even be subjective to each developer.
I am a solo developer. I write most of my code by hand. I use LLMs to help me do research and to help find stuff in documentation. I also use it to do come reviews since I don't have anyone else to review the code. I do get a lot of value out of the code review. It often finds little things I missed or will tell me if my code sucks.
Really for what I am doing writing the code is never the bottleneck. I spend little of my time actually writing code. Where I spend most of my time on the design and requirements.
> you do it on your own terms, rather than on the terms of the employer
You've setup a condition where essentially every answer is true. I know someone who spins wool and dyes it and makes socks. Does it still make sense? Only once you add in the fact she enjoys.
If you're real question is what can be done better by hand in July 2026?
afaik optimization problems are still getting superior results from humans. Things like if you do this algo you can loop the array once instead of 3x. If you align this struct like this it's faster on a certain arch + os. Or if we do these 2 steps in microservice A we can only touch the memory and network once instead of twice etc.
a 2x speed up is still meaningful at scale. (I'd bet there are 100s of software companies out there that would pay 7 to 8 figures if you could 1/2 their AWS bill)
> If you're real question is what can be done better by hand in July 2026?
The answer to that question is "everything". Humans are still noticeably better than LLMs at writing code, and just as fast once you factor in the time it takes to review and correct the LLM.
Even though the translation from English algorithm to code is the "boring" part that an LLM can do, the process of doing that works as a kind of REM sleep where I'm coming up with ways to make the program better
And for a lot of simple tasks, there's plenty of stuff where it's quicker to type a command than an English description of the same thing.
I do all of my code by hand. I don't even use an LSP. It makes sense as long as it makes sense to you.
You don't need to change the way you do things because of some perceived thing you're missing out on. That is what the AI companies want you to feel and think because it benefits them.
This applies to more than AI, but programming language choices and more. Heck, it even applies to social issues and politics :)
I am already being forced to use Claude for all coding in my job. All personal projects/code I still write by hand. This has always been the fun part of software development for me.
> prompting very small tasks after you have made the initial design, something along the lines of "Create a database migration that adds column [x], not null", "add a field to the this form that will accept string value, validate it on the backend against this format, and save it to column [x]". So basically, writing code in plain English.
Are you presenting this as the option that allows you to have better understanding (smaller tasks, closer to the changes), or the option with which you don't have understanding (because you're still getting assistance, not solving every problem on your own and not typing the code out yourself)? I genuinely can't tell from your post, people can and do take both of these positions.
> I will lose the ability to think about code and understand code
If you lose it (whatever "it" is), it's because you don't use it. If you don't use it but you're still getting your job done, is that really a loss?
We are in a hype bubble and it can be hard to go against the grain. The bubble will calm down, as it has every other time: crypto, low code, outsourcing, SPAs, Java, AI that last time around (search for Lighthill report).
I agree with your concern about skill stagnation or erosion. Years ago we used to make fun of "architectural astronauts": developers who stopped implementing and instead used English to design systems for others to implement.
I'm guessing you have had the experience writing code where you throw away the first, second and maybe third implementations? While you "work out" what the real problem is? Our human brains need that niggling at the problem before we can make a good design. Don't short circuit it.
In the context of coding on your own terms, I think it makes sense to still code to help learn and reenforce understanding. There are a lot of times I read something or see code, I think I get it, but it’s not until implementing it, and making changes to that implementation, that I actually understand it.
There is also a question of enjoyment. While you enjoy it, and while you’re learning, write it yourself. If you’ve already done something many times, understand it well, and it’s become tedious in a bad way, maybe that’s where some strategic use of AI comes in. Of course, I tend to like a bit of tedium from time to time.
I have a related question: In a contest where all the code is generated by LLMs and not hand-coded anymore, does it still make sense to use high-level languages and other abstractions that were created for humans?
I believe so, with Rust being exhibit 1. Rust is known for finding rather a lot of bugs while compiling instead of during debugging. The quote you'll often see is: "If it compiles, it usually runs."
The Rust compiler enforces many code style conditions that you most probably want in your finished program... Type checking, only one thread writing to memory at a time to avoid race conditions, etc.
It is possible to imagine an AI version of Rust that takes this even further in the future.
Maybe not the ones created for humans, but high level languages, probably. More abstract languages ("high level" basically means more abstraction away from the underlying machine, with "low level" being closer to the machine) allow for compression relative to unabstracted languages.
LLMs are known to do worse the more code they have to go through. Therefore, using a high level language where it doesn't have to work with as much code is going to be far more effective.
lot of times I feel the code that is generated has better readability and better documentation then if I had coded. so if something is doing better job then I will use that, but reviewing the code that is generated I feel is still my job, sometimes code generated is not correct, so I need keep checking.
Yes, there is still a need to learn to write code by hand.
Maybe not to the detailed nitty-gritty level in a computer language that was needed pre-LLM era but nevertheless still a must.
In this context, it is good to revisit the practices from early days of programming where "System Analysts" and "System Designers" provided detailed specifications to "Programmers" who translated it into computer code. The key point to understand here is that the specifications were detailed i.e they were code just in a non-machine executable form.
So detailed Specifications followed by Verification is mandatory. Understanding the code implementation in-between has lesser need but still important.
Its rare for me now. Most by-hand stuff I do is setting up initial repositories and folders for the base structure I want, copying logos, assets, images, content, and then I tell AI to finish the rest for example.
Configuring secrets for CICD, .env files, hosting configurations/backups.
I see the argument here constantly "But ur skills!". OK. Well back to work actually shipping software using this new tool that is never going away. My client doesn't really care about my skills, just my speed and they have money for me.
Outside of work (where they force me to use the clanker), why on God's green earth wouldn't I write code by hand? That is the fun part; I'm not going to voluntarily give up the fun part of my hobby project. If you are a person who is wired to care more about getting to the end than the process (as some here have said they are), then fine, use an LLM if you want. But there's no "makes sense to" involved here - if you like writing code then write code, if you don't then don't.
> but rather prompting very small tasks after you have made the initial design, something along the lines of "Create a database migration that adds column [x], not null", "add a field to the this form that will accept string value, validate it on the backend against this format, and save it to column [x]". So basically, writing code in plain English.
I find such tasks easier to just do in the programming language directly than describe them in English.
> I believe that writing something yourself helps you learn it better, but at the same time, if you did all the thinking and now just prompting for a fancy auto-complete, we could argue that the hardest part (the design) was already done, and you skip the writing for the sake of writing. On the other hand, I am afraid that if I will stop writing code in environments where I can still do it, I will lose the ability to think about code and understand code.
Merely completing the "design" is by far not "doing all the thinking," and this is really the core of the problem. When people code with LLMs, the LLMs are making nontrivial decisions and thus are replacing part of the programmer's intelligence. That's why calling LLMs "fancy autocomplete" is so misleading, because autocompletion isn't supposed to make decisions, just to facilitate typing out the code you already have in your head.
So yeah, I agree with the last sentence about losing the ability to understand code, and that's a sufficient reason to avoid AI assistance whenever possible. Since I prioritize education and learning above almost everything else, it's not a close call. Where the line gets blurry for me is when people say "I would never have learned the tech stack needed to make this tool myself, but I vibe-coded it and now it's mine." It's really difficult for me to argue with that, to tell someone simply to live without something that's potentially improving their day-to-day life. To be clear, there are many good reasons not to do that thing, but it's also not my place to judge people for doing it, since I don't live my own life perfectly either.
It's like any other tool. Use it when you need it.
We don't insist that HVAC installers use hand-powered drills to "feel closer to the work," but we also don't insist that they use battery-powered drills and impact drivers on every single task.
We don't insist that a chef use a Vitamix blender to chop a single onion when a couple seconds with a sharp knife will do, but we certainly wouldn't only use a sharp knife to create a salsa.
We don't insist that a carpenter use only hand-saws to build a house, but we also don't insist that they use a chainsaw to make a chair.
We don't insist that an accountant write out taxes by hand, but we also don't insist that they use a calculator to determine how old you are from your birthdate.
You use the power tools when appropriate, and you use more delicate, surgical tools when appropriate. It's not any more difficult than that.
The problem with this analogy is an LLM isn't very much like a "tool" at all.
- In cases of practical relevance, you pay a recurring fee to keep using it - not very tool-like.
- Rather than doing a predictable action in response to user inputs, you give it imprecise instructions and it does something unpredictable but vaguely correct in response, which then must be thoroughly inspected.
- In cases of practical relevance, its actual operation (not just design) is entirely handled by a third party, not by the user.
These things describe a subcontractor, not a tool. We don't insist that a chef at a restaurant use this tool or that tool, but if the restaurant was actually just outsourcing all its cooking to some other company...yeah, that would absolutely factor into my decision to go there or not.
Imagine a carpenter needed to be proficient in hand-saws to use their chainsaw well, and using their chainsaw required knowing details about the wood they were working with, in only the way using a hand-saw gave.
Makes it a littttttle more complicated.
...and it's fine to not use the loud, vibrating power saw and instead take some extra time to use a handsaw, and thus get more practiced at it.
It's fine to hire movers, but sometimes they end up damaging stuff. My solution is have a lot less stuff, and move it myself.
I increasingly have more respect for the Amish community one of my relatives lives near, at least in terms of their considered, intentional limits set on resource use. I don't agree with all their choices, and some seem arbitrary, but I respect their autonomy to choose.
I choose to move around the world primarily under my own power or augmented by a battery (supplies 2/3 of the Wattage I need to make my cargo bicycle move); harm reduction in the face of fossil-fuel addiction. I choose not to use generative AI because I'd rather human do more work for themselves instead of selling short future humans with profligate resource consumption.
It's fine to strive for things, but the most important thing in my life is my relationships with other humans. If I'm a person who can't walk, and am in physical therapy towards being able to walk, it's fine to keep working at it, and it's fine to never be able to walk; what matters is how I get on with my mates, and spending massive amounts of electricity on data centers is arguably harmful to the health of interpersonal relationships.
I really like your analogy as-is, but LLMs aren't reliable power tools. Asking for external help requires good communication and can be overkill for small tasks.
Hiring a team of subcontractors carries risks unless you babysit the whole project.
I mean, it's the same idea, no? You hire contractors to build your house, but not to put together an IKEA bookshelf.
Yes. Someone's gotta create new grist for the LLMs so they don't Habsburg themselves.
I still think I haven’t gotten much more productive than a year ago:
I wrote the barebones, scaffolding, and where I cared about the details. Agents stamped out the rest from those patterns.
Everyone speaks about this stuff in anecdotes. Or heroic attempts to have agent only projects.
But very little actual data on real gains in actual quality as experienced by users? Harder to measure IMO. And may even be subjective to each developer.
I am a solo developer. I write most of my code by hand. I use LLMs to help me do research and to help find stuff in documentation. I also use it to do come reviews since I don't have anyone else to review the code. I do get a lot of value out of the code review. It often finds little things I missed or will tell me if my code sucks.
Really for what I am doing writing the code is never the bottleneck. I spend little of my time actually writing code. Where I spend most of my time on the design and requirements.
> you do it on your own terms, rather than on the terms of the employer
You've setup a condition where essentially every answer is true. I know someone who spins wool and dyes it and makes socks. Does it still make sense? Only once you add in the fact she enjoys.
If you're real question is what can be done better by hand in July 2026?
afaik optimization problems are still getting superior results from humans. Things like if you do this algo you can loop the array once instead of 3x. If you align this struct like this it's faster on a certain arch + os. Or if we do these 2 steps in microservice A we can only touch the memory and network once instead of twice etc.
a 2x speed up is still meaningful at scale. (I'd bet there are 100s of software companies out there that would pay 7 to 8 figures if you could 1/2 their AWS bill)
> If you're real question is what can be done better by hand in July 2026?
The answer to that question is "everything". Humans are still noticeably better than LLMs at writing code, and just as fast once you factor in the time it takes to review and correct the LLM.
> just as fast once you factor in the time it takes to review and correct the LLM
Do you have a source for that claim?
When your learning it does and in work no
Yeah, still write by hand.
Even though the translation from English algorithm to code is the "boring" part that an LLM can do, the process of doing that works as a kind of REM sleep where I'm coming up with ways to make the program better
And for a lot of simple tasks, there's plenty of stuff where it's quicker to type a command than an English description of the same thing.
I do all of my code by hand. I don't even use an LSP. It makes sense as long as it makes sense to you.
You don't need to change the way you do things because of some perceived thing you're missing out on. That is what the AI companies want you to feel and think because it benefits them.
This applies to more than AI, but programming language choices and more. Heck, it even applies to social issues and politics :)
Make your own choices and walk your own path.
I am already being forced to use Claude for all coding in my job. All personal projects/code I still write by hand. This has always been the fun part of software development for me.
This, the FOMO is real, but giving in to it based on others perceptions is naive
> prompting very small tasks after you have made the initial design, something along the lines of "Create a database migration that adds column [x], not null", "add a field to the this form that will accept string value, validate it on the backend against this format, and save it to column [x]". So basically, writing code in plain English.
Are you presenting this as the option that allows you to have better understanding (smaller tasks, closer to the changes), or the option with which you don't have understanding (because you're still getting assistance, not solving every problem on your own and not typing the code out yourself)? I genuinely can't tell from your post, people can and do take both of these positions.
> I will lose the ability to think about code and understand code
If you lose it (whatever "it" is), it's because you don't use it. If you don't use it but you're still getting your job done, is that really a loss?
It makes sense to write code by hand.
We are in a hype bubble and it can be hard to go against the grain. The bubble will calm down, as it has every other time: crypto, low code, outsourcing, SPAs, Java, AI that last time around (search for Lighthill report).
I agree with your concern about skill stagnation or erosion. Years ago we used to make fun of "architectural astronauts": developers who stopped implementing and instead used English to design systems for others to implement.
I'm guessing you have had the experience writing code where you throw away the first, second and maybe third implementations? While you "work out" what the real problem is? Our human brains need that niggling at the problem before we can make a good design. Don't short circuit it.
In the context of coding on your own terms, I think it makes sense to still code to help learn and reenforce understanding. There are a lot of times I read something or see code, I think I get it, but it’s not until implementing it, and making changes to that implementation, that I actually understand it.
There is also a question of enjoyment. While you enjoy it, and while you’re learning, write it yourself. If you’ve already done something many times, understand it well, and it’s become tedious in a bad way, maybe that’s where some strategic use of AI comes in. Of course, I tend to like a bit of tedium from time to time.
If it’s a preference question then certainly yes, but how to push back on x productivity boost argument when a company enforce this as policy?
Yes a soft policy but it’s a smart move to encourage the rat race.
I think you could make a high level brainstorm session / plan on paper of things to prompt and that this was valuable even before AI
and occasionally manually writing out the code could be useful as a mental exercise to keep coding skills sharp?
Turns out I think I always liked the end product more than getting there. I would kind of hate having to only write code by hand for a paycheck again.
I still write code by hand because sometimes it's faster and cheaper.
I have a related question: In a contest where all the code is generated by LLMs and not hand-coded anymore, does it still make sense to use high-level languages and other abstractions that were created for humans?
I believe so, with Rust being exhibit 1. Rust is known for finding rather a lot of bugs while compiling instead of during debugging. The quote you'll often see is: "If it compiles, it usually runs."
The Rust compiler enforces many code style conditions that you most probably want in your finished program... Type checking, only one thread writing to memory at a time to avoid race conditions, etc.
It is possible to imagine an AI version of Rust that takes this even further in the future.
Maybe not the ones created for humans, but high level languages, probably. More abstract languages ("high level" basically means more abstraction away from the underlying machine, with "low level" being closer to the machine) allow for compression relative to unabstracted languages.
LLMs are known to do worse the more code they have to go through. Therefore, using a high level language where it doesn't have to work with as much code is going to be far more effective.
lot of times I feel the code that is generated has better readability and better documentation then if I had coded. so if something is doing better job then I will use that, but reviewing the code that is generated I feel is still my job, sometimes code generated is not correct, so I need keep checking.
doesnt matter if you use AI as long as you check the code afterword and know what you are doing
Would you ever want to be as helpless at a skill you care about as the typical person is about math, since calculators were invented?
I'd love to vibe code with an army of agents, I just don't know if I could afford to vibe debug the resulting mess lol
Does it make sense to still walk? I mean, we have cars. You want to go somewhere that's 1km away from home. Does it make sense to walk?
Someone wants to give you a phone number. Does it make sense to still try to memorize it or must you hunt down paper and pen to write it down?
... you get the drift? Do what makes sense to you.
Yes, there is still a need to learn to write code by hand.
Maybe not to the detailed nitty-gritty level in a computer language that was needed pre-LLM era but nevertheless still a must.
In this context, it is good to revisit the practices from early days of programming where "System Analysts" and "System Designers" provided detailed specifications to "Programmers" who translated it into computer code. The key point to understand here is that the specifications were detailed i.e they were code just in a non-machine executable form.
So detailed Specifications followed by Verification is mandatory. Understanding the code implementation in-between has lesser need but still important.
See also my previous comment chains from here - https://news.ycombinator.com/item?id=48800484
Its rare for me now. Most by-hand stuff I do is setting up initial repositories and folders for the base structure I want, copying logos, assets, images, content, and then I tell AI to finish the rest for example.
Configuring secrets for CICD, .env files, hosting configurations/backups.
Writing instructions files, specs, copy/pasting context, requirements, examples.
I see the argument here constantly "But ur skills!". OK. Well back to work actually shipping software using this new tool that is never going away. My client doesn't really care about my skills, just my speed and they have money for me.
Outside of work (where they force me to use the clanker), why on God's green earth wouldn't I write code by hand? That is the fun part; I'm not going to voluntarily give up the fun part of my hobby project. If you are a person who is wired to care more about getting to the end than the process (as some here have said they are), then fine, use an LLM if you want. But there's no "makes sense to" involved here - if you like writing code then write code, if you don't then don't.