The “correct” way to use AI for coding (and anything really) is to ask for explanations / tutorials when you can’t find one online, then learn from that.
Never let it do something for you. That’s how you lose. If you’re not actively learning, you’re actively rotting, and that goes for life in general too.
What you said is fine for people learning, but there’s nothing wrong with having AI do something for you when you’ve done it a hundred times and don’t want to do it again. Some of us are actually out here working, not just learning.
I have started using LLM tools recently after taking a new job where a lot of people do it. I’ve discovered that it’s actually fairly helpful not only for explanations, but in two other respects
Sifting through immense amounts of documentation. I have to deal with some datasheets that are hundreds of pages, where there will be info scattered throughout. It’s very helpful sifting through those.
Doing boiler plate “plumbing work” in my code. I’m mostly drawing a line where I don’t want it doing the “core” work in that which I’m an expert, since I agree that if I stop doing that, I’ll atrophy. However it can help accelerate my process if I pass off some of the minutiae that I don’t feel the need to do.
However all that said, I am honestly pretty impressed how well it works. I’ve mostly been using Claude, and damn, it’s honestly pretty competent. I had it make me a helper Python GUI program for me to test some stuff (I’m not a UI/high level engineer like that, I’m an FPGA Engineer), and it did a decent job. It definitely needed a good amount of massaging and guidance. However I can definitely see the appeal, and I think it’s a slippery slope, and I need to make sure I remain disciplined in not letting it do everything
One trap is to trust it as a means to accomodate unreasonable schedule pressure.
Sure - this thing looks like it works, hell it probably does work, do you really want to launch a probably works product? If your management does - consider shopping around for a raise/promotion under different management. It’s never easy to move, but if you’re moving on your own terms you can often make the effort worth your while.
Another note: I find the LLMs to be wickedly detail oriented code reviewers - like, they’ll point out the tiniest little discrepancies and edge cases, and what they (Claude, at least) report is usually “real.” Now, that doesn’t mean they find everything that’s wrong on the first pass, but once you’ve addressed everything in the first pass, you can make a second pass, and a third, etc. each time with different focus: documentation complete? implementation functions as intended? technical debt? test coverage? security issues? issues with maintainability? documentation in sync with implementation? specific aspect of implementation functions as intended? etc. - if you address all the findings after each review cycle (and addressing a finding can be clarifying a requirement to relax about certain unimportant aspects…) eventually the findings slow down / only find ridiculously unimportant things.
A thing I found quite amusing about the AI agents I’ve toyed with is that they have a step where they do a code review of their changelist, usually switching to a different “persona” when they write it so that they’re not seeing it as “their own” code. It’s funny reading at the critiques and compliments it gives the “other agent” it’s checking the changes for.
I haven’t seen this feature yet, but it might be a good future enhancement to ensure that the harness literally uses a different model for the code review from the one that wrote the code in the first place. If Claude wrote the code have GPT do the review, and vice versa, for example. Wouldn’t be surprised if the feature exists and I just haven’t spotted it yet though, things change fast.
I use Cursor for work (Claude Code at home), and Cursor gives the option to select your model. I’ve dabbled a bit with GPT for the review of Claude code - haven’t found anything dramatically better doing that than just Claude prompted to “wear the reviewer hat now.”
Yeah, I wouldn’t use a framework that didn’t let you select the basic model. I’m just thinking about having it automatically switch to a different one during the review “phase”. It’s not as popular a coding agent these days but I like using Google’s Antigravity and it’s capable of being told to go through the sequence of steps “plan - > write documentation -> implement the plan -> run unit tests -> do a code review” automatically without needing to be prompted at each step. That’s where it would be nice to have it automatically switch for the review.
“Wear the reviewer hat now” does seem to work quite well with the same model, but if more models from different lineages are available it just seems like the right thing to do to switch to another one.
I don’t think that’s a good idea, if you can’t find an explanation online that means that there’s not much info available in which case the best thing would be to ask on a forum, that way other people that look for that info will find it.
Usually, the LLM’s response will be incomplete or partially incorrect, but it’s often good enough to get un-stuck.
Usually it will have some keywords you can look up, some bits that bring up further questions for you to answer (and for which the LLM should also not be your first choice).
There’s also the aspect of giving the LLM another prompt, and another, and another - have it build up a local documentation set based on its internet research, continue to research and refine the local documentation set keeping the things you trust, filtering out the sketchy stuff…
Ultimately, the LLM is a tool you are using. If you use it like a 4th grader copying a paragraph out of the Encyclopedia - you’re probably not going to get great results; especially because today’s reference materials aren’t highly edited / vetted material like encyclopedias were, today’s reference materials are internet forums full of self-confident idiots blathering on about whatever they think they know something about (like me, here…) So, back to the tool thing: if you use it well, you can make nice things. If you use it lazily, don’t be surprised when your boss decides he doesn’t need you at your salary to push that button.
The “correct” way to use AI for coding (and anything really) is to ask for explanations / tutorials when you can’t find one online, then learn from that.
except the “explanation” frequently will be 100% “hallucinated” bullshit
For what it’s worth, I’ve been working on (yet another) ActivityPub based micro blogging application and LLMs have been enormously helpful and so far as I can tell, correct. Often it cites the AP specs and its extensions, as well as specific implementations from existing major AP apps. It can show me expected outputs, what responses from my app should look like in response to different requests from other servers, and quickly give context for features like Mastodon’s shared inbox. I’m not having it simply generate code, but I think I’m still moving way faster than I otherwise could. I don’t recall it ever giving me incorrect information.
It’s the first time I’ve used an LLM as a tool this way, and I’m pretty impressed with it. I’m using the assistant made available through Kagi.
Tip: check those citations yourself before publishing with your name on the product. Yeah, they’re usually correct - do you only usually not want to be perceived as a lazy idiot?
People say the best way to see this is asking AI about subject you’re expert of.
This is not always possible, I had people who said “but I’m not expert at anything”. Another way is to ask them about yourselves. For example if you have reddit account that is has some age, Gemini has deal with reddit and feeds them everything that’s posted. First response might even look good, but continue talking (as it is getting more ridiculous), don’t try correct, you can see how it is making shit up.
Since they are feeding it with everything lemmy might also work.
I’ve seen very mixed results depending on which model I’m using. The newer ones, since about November of 2025, have been getting significantly better - but some of the “free class” tools are still using older ones today.
Free Gemini gave me extremely ridiculously bad advice about how to get through a traffic jam today. Free Gemini also drew the crudest sketch imaginable for a prompt, same prompt fed to ChatGPT yielded a really nice quality cartoon panel of basically exactly everything in the prompt, with some nice/appropriate embellishments.
I’ve become rather disillusioned with Gemini’s use of search tools lately. It’s odd given that it’s a Google model, you’d think Google would be at the top of the search engine game. But honestly, Deepseek’s been my go-to lately when I want an answer that’s likely to be synthesized from a lot of web searches. I’ve had it search over a hundred different pages for a generic “how does this work?” Sort of query. It didn’t read them all, but it’s casting a wide net and it’s letting me actually see the details. Gemini seems more willing to just tell me what it “thinks” the answer to a question is based off of its training data, which is not a particularly reliable thing for an LLM to do.
Gemini seems more willing to just tell me what it “thinks” the answer to a question is based off of its training data, which is not a particularly reliable thing for an LLM to do.
Yeah. I pay for Claude, my company pays even more for Cursor, so comparing them to free Gemini probably isn’t fair.
Gemini is very useful for offhand queries while Claude is chewing on a bigger problem, but if it’s something that needs complex analysis and/or extensive research… the tools that let you build up a folder full of files related to the task are vastly superior to chatbots. Gemini does have a Claude Code command line tool that does that kind of development in a folder, I didn’t install it until last week. Gave it a coding problem to work on (lookup realtime weather radar data from NOAA, present recent data on a map on a webpage)… it sort of succeeded, but with poor user experience. Again, I’m in “Free mode” which can do quite a bit on a day’s allowance of tokens, but… I don’t feel like their paid modes would be particularly higher quality. If they are, they’re doing themselves a tremendous disservice by demoing such substandard performance in free mode.
That’s why I always ask it to cite sources. Basically googld ATP since google is turning to shit and all other search engines still aren’t quite as good
It could very easily use a completely different or hallucinated source.
But a lot of LLM products are now providing source links right in the response. I’ve found them useful, and hopefully they aren’t produced just by feeding the text back in and asking for a link.
The “correct” way to use AI for coding (and anything really) is to ask for explanations / tutorials when you can’t find one online, then learn from that.
Never let it do something for you. That’s how you lose. If you’re not actively learning, you’re actively rotting, and that goes for life in general too.
What you said is fine for people learning, but there’s nothing wrong with having AI do something for you when you’ve done it a hundred times and don’t want to do it again. Some of us are actually out here working, not just learning.
I have started using LLM tools recently after taking a new job where a lot of people do it. I’ve discovered that it’s actually fairly helpful not only for explanations, but in two other respects
However all that said, I am honestly pretty impressed how well it works. I’ve mostly been using Claude, and damn, it’s honestly pretty competent. I had it make me a helper Python GUI program for me to test some stuff (I’m not a UI/high level engineer like that, I’m an FPGA Engineer), and it did a decent job. It definitely needed a good amount of massaging and guidance. However I can definitely see the appeal, and I think it’s a slippery slope, and I need to make sure I remain disciplined in not letting it do everything
One trap is to trust it as a means to accomodate unreasonable schedule pressure.
Sure - this thing looks like it works, hell it probably does work, do you really want to launch a probably works product? If your management does - consider shopping around for a raise/promotion under different management. It’s never easy to move, but if you’re moving on your own terms you can often make the effort worth your while.
Another note: I find the LLMs to be wickedly detail oriented code reviewers - like, they’ll point out the tiniest little discrepancies and edge cases, and what they (Claude, at least) report is usually “real.” Now, that doesn’t mean they find everything that’s wrong on the first pass, but once you’ve addressed everything in the first pass, you can make a second pass, and a third, etc. each time with different focus: documentation complete? implementation functions as intended? technical debt? test coverage? security issues? issues with maintainability? documentation in sync with implementation? specific aspect of implementation functions as intended? etc. - if you address all the findings after each review cycle (and addressing a finding can be clarifying a requirement to relax about certain unimportant aspects…) eventually the findings slow down / only find ridiculously unimportant things.
A thing I found quite amusing about the AI agents I’ve toyed with is that they have a step where they do a code review of their changelist, usually switching to a different “persona” when they write it so that they’re not seeing it as “their own” code. It’s funny reading at the critiques and compliments it gives the “other agent” it’s checking the changes for.
I haven’t seen this feature yet, but it might be a good future enhancement to ensure that the harness literally uses a different model for the code review from the one that wrote the code in the first place. If Claude wrote the code have GPT do the review, and vice versa, for example. Wouldn’t be surprised if the feature exists and I just haven’t spotted it yet though, things change fast.
I use Cursor for work (Claude Code at home), and Cursor gives the option to select your model. I’ve dabbled a bit with GPT for the review of Claude code - haven’t found anything dramatically better doing that than just Claude prompted to “wear the reviewer hat now.”
Yeah, I wouldn’t use a framework that didn’t let you select the basic model. I’m just thinking about having it automatically switch to a different one during the review “phase”. It’s not as popular a coding agent these days but I like using Google’s Antigravity and it’s capable of being told to go through the sequence of steps “plan - > write documentation -> implement the plan -> run unit tests -> do a code review” automatically without needing to be prompted at each step. That’s where it would be nice to have it automatically switch for the review.
“Wear the reviewer hat now” does seem to work quite well with the same model, but if more models from different lineages are available it just seems like the right thing to do to switch to another one.
Ask it what the Helvetica scenario is.
So Using it as my emotional dumbing machine is wrong ?
I don’t think that’s a good idea, if you can’t find an explanation online that means that there’s not much info available in which case the best thing would be to ask on a forum, that way other people that look for that info will find it.
Usually, the LLM’s response will be incomplete or partially incorrect, but it’s often good enough to get un-stuck.
Usually it will have some keywords you can look up, some bits that bring up further questions for you to answer (and for which the LLM should also not be your first choice).
There’s also the aspect of giving the LLM another prompt, and another, and another - have it build up a local documentation set based on its internet research, continue to research and refine the local documentation set keeping the things you trust, filtering out the sketchy stuff…
Ultimately, the LLM is a tool you are using. If you use it like a 4th grader copying a paragraph out of the Encyclopedia - you’re probably not going to get great results; especially because today’s reference materials aren’t highly edited / vetted material like encyclopedias were, today’s reference materials are internet forums full of self-confident idiots blathering on about whatever they think they know something about (like me, here…) So, back to the tool thing: if you use it well, you can make nice things. If you use it lazily, don’t be surprised when your boss decides he doesn’t need you at your salary to push that button.
except the “explanation” frequently will be 100% “hallucinated” bullshit
And that’s why you have to check everything.
For what it’s worth, I’ve been working on (yet another) ActivityPub based micro blogging application and LLMs have been enormously helpful and so far as I can tell, correct. Often it cites the AP specs and its extensions, as well as specific implementations from existing major AP apps. It can show me expected outputs, what responses from my app should look like in response to different requests from other servers, and quickly give context for features like Mastodon’s shared inbox. I’m not having it simply generate code, but I think I’m still moving way faster than I otherwise could. I don’t recall it ever giving me incorrect information.
It’s the first time I’ve used an LLM as a tool this way, and I’m pretty impressed with it. I’m using the assistant made available through Kagi.
Tip: check those citations yourself before publishing with your name on the product. Yeah, they’re usually correct - do you only usually not want to be perceived as a lazy idiot?
People say the best way to see this is asking AI about subject you’re expert of.
This is not always possible, I had people who said “but I’m not expert at anything”. Another way is to ask them about yourselves. For example if you have reddit account that is has some age, Gemini has deal with reddit and feeds them everything that’s posted. First response might even look good, but continue talking (as it is getting more ridiculous), don’t try correct, you can see how it is making shit up.
Since they are feeding it with everything lemmy might also work.
I’ve seen very mixed results depending on which model I’m using. The newer ones, since about November of 2025, have been getting significantly better - but some of the “free class” tools are still using older ones today.
Free Gemini gave me extremely ridiculously bad advice about how to get through a traffic jam today. Free Gemini also drew the crudest sketch imaginable for a prompt, same prompt fed to ChatGPT yielded a really nice quality cartoon panel of basically exactly everything in the prompt, with some nice/appropriate embellishments.
I’ve become rather disillusioned with Gemini’s use of search tools lately. It’s odd given that it’s a Google model, you’d think Google would be at the top of the search engine game. But honestly, Deepseek’s been my go-to lately when I want an answer that’s likely to be synthesized from a lot of web searches. I’ve had it search over a hundred different pages for a generic “how does this work?” Sort of query. It didn’t read them all, but it’s casting a wide net and it’s letting me actually see the details. Gemini seems more willing to just tell me what it “thinks” the answer to a question is based off of its training data, which is not a particularly reliable thing for an LLM to do.
Yeah. I pay for Claude, my company pays even more for Cursor, so comparing them to free Gemini probably isn’t fair.
Gemini is very useful for offhand queries while Claude is chewing on a bigger problem, but if it’s something that needs complex analysis and/or extensive research… the tools that let you build up a folder full of files related to the task are vastly superior to chatbots. Gemini does have a Claude Code command line tool that does that kind of development in a folder, I didn’t install it until last week. Gave it a coding problem to work on (lookup realtime weather radar data from NOAA, present recent data on a map on a webpage)… it sort of succeeded, but with poor user experience. Again, I’m in “Free mode” which can do quite a bit on a day’s allowance of tokens, but… I don’t feel like their paid modes would be particularly higher quality. If they are, they’re doing themselves a tremendous disservice by demoing such substandard performance in free mode.
That’s why I always ask it to cite sources. Basically googld ATP since google is turning to shit and all other search engines still aren’t quite as good
Then why not ask just for the sources and read them yourself?
I do that to make sure the sources are worth my time and not just hallucinations
It could very easily use a completely different or hallucinated source.
But a lot of LLM products are now providing source links right in the response. I’ve found them useful, and hopefully they aren’t produced just by feeding the text back in and asking for a link.
That’s exactly how those links are produced.