We might be in the wrong community here to discuss a positive attitude towards AI coding… But anyway… Do you like it? I think I’m more and more coming to the conclusion that I don’t really fancy it. It’s somewhat fulfilling to code something. But my experience with AI is I’ll spend 90 minutes arguing with it and making it have countless shots at the one problem, and then I end up reading all the code, refactoring it and rewriting snippets and it’s super tedious and I’m annoyed because I like computers for doing exactly what I tell the to do, and now I have to argue with the darn thing about the specifications or how memory allocation or maths works a certain way or if we can pull in random libraries for a simple task… So I’m a bit split on this. At first it was very exciting and fascinating. But I think for coding that kind of got old quickly. At least for me and the stuff I do. These days I’ll use it for quick tech-demos, templates, placeholders, to google the documentation, translate Chinese and the like but I’ve cut down on the actual coding mostly because it takes the fun out of it and turns it upside-down into reviewing and correcting code.
To the extent I have grown more comfortable, it’s accepting that the AI is usually wrong and giving up on trying unless it’s obvious and short. I won’t “argue” with it, I just discard and do it myself. I’ll also click “review my code” and give it a chance to highlight mistakes. Again it is frequently wrong. But once it did catch an inconsistency that I know would have been frustrating when it eventually reared its head.
The thing that I’m thinking of turning off is code completion with tab. Problem is that the lag means I didn’t know if the tab key is going to do a normal thing or if by the time I hit it an AI suggestion pops up and I have to undo the unexpected modification. Also sometimes the suggestions linger and make the actual code hard to read long after I already decided to ignore the suggestion.
Yesterday was a fair amount of tab completing through excessively boilerplate crap thanks to AI, but most days it’s next to useless as I am in low boilerplate scenarios. Some frameworks and languages make you type a novel to do something very common, and AI helps with those. I tend to avoid those but I didn’t have a choice yesterday. Even then the AI made some very bad suggestions, so I have to be in the lookout at all times.
Not OP but I had great success letting it repeat stuff we already have, for example we have a certain pattern on how we place translations. So I just hardcode everything and in the end tell it, using a pre-written task I can just call up, to take all the hardcoded labels and place it im our system in the same way it has already been done. It then reads the code of a few existing components and replicates that. Or I let it extract some code into smaller components. Or move some component around, it can do that batter than the IDEs integrated move action.
Completely novel stuff is possible but I am uncertain if I am actually not slower using it to achieve that. I mostly do it step by step, really small steps that is.
I have to measure my performance at some point, it is certainly possible that I am actually slower than before. But overall I never liked typing out the solution that is in my head, so using it as writer is nice.
Sonnet 4.5 is what I use. Some colleagues like GTP-5 but it struggles real hard to do the most basic things right in my experience. Claude is just miles ahead.
Yeah, I sometimes find the same loop with “this thing just don’t understand what I’m asking for” - I’ve had luck with breaking it down into smaller steps, and being specific about the requirements helps. I use Claude Sonnet 4.5 which is pretty decent, the OpenAI models really don’t compare and are pretty bad at best at coding.
Thanks. Yeah I didn’t try Claude. They want my phone number to sign up and I’m not providing that to people. But you’re not the only person suggesting Claude Sonnet, I’ve read that several times now. I wonder if they’re really that much better. I’ll try some more throwaway phone numbers to get in, but seems they’ve blocked most of them.
I’ve tried breaking down things as well. That’s usually a good strategy in programming. Though I guess at some point they’re small enough so I could have already typed it in myself instead of talking about doing it. And I find it often also struggles to find the right balance with the level of detail of a function and whether it’s clever to do a very specific singular thing or do it a bit more general so the function can be reused in other parts of the code. So it’ll be extra work later to revise it, once everything is supposed to come together and integrate/combine.
I don’t think it’s the model, it’s more the process I don’t like. There is some appeal to programming for me in understanding things and creating something with my mind, not dealing with people but solid logic problems which are properly right or wrong and it’ll all come together from a small defined set of very basic things and then I combine them in some craft to achieve arbitrary things. It’ll come together in my mind and then I type everything down and figure out the rest and that’s kind of satisfying to me. AI programming feels more like sitting in endless meetings to discuss details and revise things, then fix stuff and that’s all the tings I don’t like, and in the meantime we’ve brushed over the one thing I did like. I’d rather it did the tedious things for me. Write specifications, documentation, sift through data and convert it, do correspondence and handle finances. Organize the mess on my computer and bring coffee… Idk. I guess I could break up tasks and delegate if it had the “intelligence” for it.
I think I tried most free commercial ones which didn’t need my phone number… Gemini (AIstudio), ChatGPT, Grok. I’ve experimented a bit with local ones like DeepSeek but I don’t really have the hardware for that, so it’s smaller variants and takes ages to ingest the code and write something.
(And btw I think the individual experience also depends on the task and programming language involved. Seems to me AI’s performance varies a lot depending on if it’s currently writing Python, JavaScript or C++ code for an embedded system… And sometimes coding problems are fairly contained and sometimes you need to have some overview of a larger codebase to tackle one single task. So I guess we might get different results depending on the specific project?!)
Got it, yeah I hated all of those too. Claude Max is a game changer if you’re looking for a better experience, since that experience is heavily tied to model choice.
I’ve found that I can spend two hours writing or 30 minutes editing, and so long as I understand every line I commit I’m keeping up my end of the bargain - but often that’s the bit where everyone gets lazy. I am also used to managing a bunch of junior engineers though, so this motion does feel very natural to me.
Thanks, yeah I can understand that. I’ll try to get ahold of an Anthropic account and see what it’s about. It’s not high priority for me, but eventually I’d like to know the truth. I have some random hobby projects with stuff I could throw at it.
We might be in the wrong community here to discuss a positive attitude towards AI coding… But anyway… Do you like it? I think I’m more and more coming to the conclusion that I don’t really fancy it. It’s somewhat fulfilling to code something. But my experience with AI is I’ll spend 90 minutes arguing with it and making it have countless shots at the one problem, and then I end up reading all the code, refactoring it and rewriting snippets and it’s super tedious and I’m annoyed because I like computers for doing exactly what I tell the to do, and now I have to argue with the darn thing about the specifications or how memory allocation or maths works a certain way or if we can pull in random libraries for a simple task… So I’m a bit split on this. At first it was very exciting and fascinating. But I think for coding that kind of got old quickly. At least for me and the stuff I do. These days I’ll use it for quick tech-demos, templates, placeholders, to google the documentation, translate Chinese and the like but I’ve cut down on the actual coding mostly because it takes the fun out of it and turns it upside-down into reviewing and correcting code.
To the extent I have grown more comfortable, it’s accepting that the AI is usually wrong and giving up on trying unless it’s obvious and short. I won’t “argue” with it, I just discard and do it myself. I’ll also click “review my code” and give it a chance to highlight mistakes. Again it is frequently wrong. But once it did catch an inconsistency that I know would have been frustrating when it eventually reared its head.
The thing that I’m thinking of turning off is code completion with tab. Problem is that the lag means I didn’t know if the tab key is going to do a normal thing or if by the time I hit it an AI suggestion pops up and I have to undo the unexpected modification. Also sometimes the suggestions linger and make the actual code hard to read long after I already decided to ignore the suggestion.
Yesterday was a fair amount of tab completing through excessively boilerplate crap thanks to AI, but most days it’s next to useless as I am in low boilerplate scenarios. Some frameworks and languages make you type a novel to do something very common, and AI helps with those. I tend to avoid those but I didn’t have a choice yesterday. Even then the AI made some very bad suggestions, so I have to be in the lookout at all times.
Not OP but I had great success letting it repeat stuff we already have, for example we have a certain pattern on how we place translations. So I just hardcode everything and in the end tell it, using a pre-written task I can just call up, to take all the hardcoded labels and place it im our system in the same way it has already been done. It then reads the code of a few existing components and replicates that. Or I let it extract some code into smaller components. Or move some component around, it can do that batter than the IDEs integrated move action. Completely novel stuff is possible but I am uncertain if I am actually not slower using it to achieve that. I mostly do it step by step, really small steps that is.
I have to measure my performance at some point, it is certainly possible that I am actually slower than before. But overall I never liked typing out the solution that is in my head, so using it as writer is nice.
Sonnet 4.5 is what I use. Some colleagues like GTP-5 but it struggles real hard to do the most basic things right in my experience. Claude is just miles ahead.
Yeah, I sometimes find the same loop with “this thing just don’t understand what I’m asking for” - I’ve had luck with breaking it down into smaller steps, and being specific about the requirements helps. I use Claude Sonnet 4.5 which is pretty decent, the OpenAI models really don’t compare and are pretty bad at best at coding.
Thanks. Yeah I didn’t try Claude. They want my phone number to sign up and I’m not providing that to people. But you’re not the only person suggesting Claude Sonnet, I’ve read that several times now. I wonder if they’re really that much better. I’ll try some more throwaway phone numbers to get in, but seems they’ve blocked most of them.
I’ve tried breaking down things as well. That’s usually a good strategy in programming. Though I guess at some point they’re small enough so I could have already typed it in myself instead of talking about doing it. And I find it often also struggles to find the right balance with the level of detail of a function and whether it’s clever to do a very specific singular thing or do it a bit more general so the function can be reused in other parts of the code. So it’ll be extra work later to revise it, once everything is supposed to come together and integrate/combine.
What model are you using that caused you to “not fancy it?”
I don’t think it’s the model, it’s more the process I don’t like. There is some appeal to programming for me in understanding things and creating something with my mind, not dealing with people but solid logic problems which are properly right or wrong and it’ll all come together from a small defined set of very basic things and then I combine them in some craft to achieve arbitrary things. It’ll come together in my mind and then I type everything down and figure out the rest and that’s kind of satisfying to me. AI programming feels more like sitting in endless meetings to discuss details and revise things, then fix stuff and that’s all the tings I don’t like, and in the meantime we’ve brushed over the one thing I did like. I’d rather it did the tedious things for me. Write specifications, documentation, sift through data and convert it, do correspondence and handle finances. Organize the mess on my computer and bring coffee… Idk. I guess I could break up tasks and delegate if it had the “intelligence” for it.
I think I tried most free commercial ones which didn’t need my phone number… Gemini (AIstudio), ChatGPT, Grok. I’ve experimented a bit with local ones like DeepSeek but I don’t really have the hardware for that, so it’s smaller variants and takes ages to ingest the code and write something.
(And btw I think the individual experience also depends on the task and programming language involved. Seems to me AI’s performance varies a lot depending on if it’s currently writing Python, JavaScript or C++ code for an embedded system… And sometimes coding problems are fairly contained and sometimes you need to have some overview of a larger codebase to tackle one single task. So I guess we might get different results depending on the specific project?!)
Got it, yeah I hated all of those too. Claude Max is a game changer if you’re looking for a better experience, since that experience is heavily tied to model choice.
I’ve found that I can spend two hours writing or 30 minutes editing, and so long as I understand every line I commit I’m keeping up my end of the bargain - but often that’s the bit where everyone gets lazy. I am also used to managing a bunch of junior engineers though, so this motion does feel very natural to me.
Best of luck
Thanks, yeah I can understand that. I’ll try to get ahold of an Anthropic account and see what it’s about. It’s not high priority for me, but eventually I’d like to know the truth. I have some random hobby projects with stuff I could throw at it.