Eric S. Raymond on X: "My life has changed this week. There's finally an AI coding tool that's good enough to keep up with me. It's called "Aider". It's a front end that knows how to talk to any model that speaks the standard JSON API, and knows how to work git, and has been cleverly and effectively prompted to behave like someone you are pair-programming with. In the last four days I have used Aider to vibe-code a non-trivial web application - a manager for a public database of capabilities about a certain class of hardware. I'm being a bit coy about this because I don't want to announce it before I have it fully ready for production, and it's not there yet - I need to do initial population of the product database and run a bunch of tests before I'm willing to let the public beat on it. But... Aider did a pretty good job magicking up an entire running web application from my original description. It has since written two moderately complex test and utility scripts to go with the server, and I'm routinely telling it to generate significant chunks of code and getting something that is not only testably correct but easy to read. Part of this is probably down to me choosing good tools that are popular enough for it to have seen a lot of code in them. Python and the Flask web framework in this case, with SQLlte's Python bindings for the database backend. Part of it is down to me writing good prompts, I think. Decades of programming experience equips me well to specify what I want crisply. But even with these tailwinds I'm finding Aider/Sonnet's performance quite impressive. It's open-source, has obviously had a lot of thought and good design put into it, and is even well documented. It still has the problem to some extent that I've noticed with LLMs in general. It doesn't have much design sense. It knows how to emulate good code by pattern matching, but it doesn't know, for example, that you really ought to encapsulate all your SQL access stuff into a class with methods rather than having exposed query generation all mixed up with your business logic. But it's my job to supply the design sense. I'm still learning more about how to prompt it in ways that convey that. In the meantime, it is more than good enough for most boilerplate code. And it's good at spotting errors - one of the things that routinely does is ask you if you wanted to lint the code, and if you tell it yes about one in three times it's going to find a minor error that probably would have taken you a lot longer to spot. It can even - gasp! - write decent documentation. When I add a user-visible feature I tell it to update the manual, and it's done a passable job every time. It can't write as fluently and concisely as I can, but most humans can't manage that either, and what it does write is good enough to get the job done. So, my workflow is permanently changed now. Going back to coding without AI would be plain stupid - it makes me much faster and delivers constant correctness checks. I haven't found the limits of what it can do yet. Two big things I'm going to ask it to try, once I have the database populated enough, is writing end-to-end tests and CSS-styling the website. Which right now has a brutally functional 1990s look that is not going to fly for production in this century. Of course I also have on my to-do list composing a prompt that tells an LLM to do the product research needed to add a device to the database. We'll see how that goes. This website is also a social-engineering hack. But I'll write about that part when I have it on a public site where y'all can see it. So, if you write code and you can run Python programs, grab a copy of Aider and check it out. It's going to rock your world."

Conversation

My life has changed this week. There's finally an AI coding tool that's good enough to keep up with me. It's called "Aider". It's a front end that knows how to talk to any model that speaks the standard JSON API, and knows how to work git, and has been cleverly and effectively prompted to behave like someone you are pair-programming with. In the last four days I have used Aider to vibe-code a non-trivial web application - a manager for a public database of capabilities about a certain class of hardware. I'm being a bit coy about this because I don't want to announce it before I have it fully ready for production, and it's not there yet - I need to do initial population of the product database and run a bunch of tests before I'm willing to let the public beat on it. But... Aider did a pretty good job magicking up an entire running web application from my original description. It has since written two moderately complex test and utility scripts to go with the server, and I'm routinely telling it to generate significant chunks of code and getting something that is not only testably correct but easy to read. Part of this is probably down to me choosing good tools that are popular enough for it to have seen a lot of code in them. Python and the Flask web framework in this case, with SQLlte's Python bindings for the database backend. Part of it is down to me writing good prompts, I think. Decades of programming experience equips me well to specify what I want crisply. But even with these tailwinds I'm finding Aider/Sonnet's performance quite impressive. It's open-source, has obviously had a lot of thought and good design put into it, and is even well documented. It still has the problem to some extent that I've noticed with LLMs in general. It doesn't have much design sense. It knows how to emulate good code by pattern matching, but it doesn't know, for example, that you really ought to encapsulate all your SQL access stuff into a class with methods rather than having exposed query generation all mixed up with your business logic. But it's my job to supply the design sense. I'm still learning more about how to prompt it in ways that convey that. In the meantime, it is more than good enough for most boilerplate code. And it's good at spotting errors - one of the things that routinely does is ask you if you wanted to lint the code, and if you tell it yes about one in three times it's going to find a minor error that probably would have taken you a lot longer to spot. It can even - gasp! - write decent documentation. When I add a user-visible feature I tell it to update the manual, and it's done a passable job every time. It can't write as fluently and concisely as I can, but most humans can't manage that either, and what it does write is good enough to get the job done. So, my workflow is permanently changed now. Going back to coding without AI would be plain stupid - it makes me much faster and delivers constant correctness checks. I haven't found the limits of what it can do yet. Two big things I'm going to ask it to try, once I have the database populated enough, is writing end-to-end tests and CSS-styling the website. Which right now has a brutally functional 1990s look that is not going to fly for production in this century. Of course I also have on my to-do list composing a prompt that tells an LLM to do the product research needed to add a device to the database. We'll see how that goes. This website is also a social-engineering hack. But I'll write about that part when I have it on a public site where y'all can see it. So, if you write code and you can run Python programs, grab a copy of Aider and check it out. It's going to rock your world.