Vibe coding only works for “I need a one time use python script that processes this data and I can’t be assed to spend 2 hours remebering how to use Pandas right now.”
Not for million dollar apps.
Of all the AI use cases, coding is honestly the one that makes sense since like 90% of coding is just copying some other code and massaging it into place. Even as a “Vibe Coder” you kind of still need to know the idea of what code you need and what you want it to do.
Even as a “Vibe Coder” you kind of still need to know the idea of what code you need and what you want it to do.
I’ve fiddled with it a bit. IM(limited)E, you tend to get the best results if what you tell it you want broadly, and then in a significant level of detail, and especially if you tell it to ask you any questions it has about the design. So something like “I want you to build an app in $LANGUAGE that does $TASK on $PLATFORM. As I see it, I think the interface should look like $DETAILED_DESCRIPTION_OF_INTERFACE and here’s what should happen when those elements are interacted with, Please ask as many questions as you think are necessary for clarity.”
Also, unless you’re setting it to require personal approval for every terminal command, I’d only run one of those in a VM of some kind, where any potential damage from any potential fuckups are limited.
I mostly use the web version of Claude. But generally my approach has been to start simple and build bigger.
“I want a website that does X.”
“Can you add a dark mode toggle.”
“I don’tlike those colors, can you make it more blue.”
“Can you add a log in function with a simple database backend that doesn’t use a full SQL install. Users should be able to add a profile image and username.”
“The first user should be an admin and can add or remove other admins.”
“Admins can change the theme colors on a settings page.”
That sort of thing. Start basic, add features, usually one by one, test as you go, see what breaks, if anything.
Vibe coding only works for “I need a one time use python script that processes this data and I can’t be assed to spend 2 hours remebering how to use Pandas right now.”
Not for million dollar apps.
Of all the AI use cases, coding is honestly the one that makes sense since like 90% of coding is just copying some other code and massaging it into place. Even as a “Vibe Coder” you kind of still need to know the idea of what code you need and what you want it to do.
I’ve fiddled with it a bit. IM(limited)E, you tend to get the best results if what you tell it you want broadly, and then in a significant level of detail, and especially if you tell it to ask you any questions it has about the design. So something like “I want you to build an app in $LANGUAGE that does $TASK on $PLATFORM. As I see it, I think the interface should look like $DETAILED_DESCRIPTION_OF_INTERFACE and here’s what should happen when those elements are interacted with, Please ask as many questions as you think are necessary for clarity.”
Also, unless you’re setting it to require personal approval for every terminal command, I’d only run one of those in a VM of some kind, where any potential damage from any potential fuckups are limited.
I mostly use the web version of Claude. But generally my approach has been to start simple and build bigger.
“I want a website that does X.”
“Can you add a dark mode toggle.”
“I don’tlike those colors, can you make it more blue.”
“Can you add a log in function with a simple database backend that doesn’t use a full SQL install. Users should be able to add a profile image and username.”
“The first user should be an admin and can add or remove other admins.”
“Admins can change the theme colors on a settings page.”
That sort of thing. Start basic, add features, usually one by one, test as you go, see what breaks, if anything.