Comment by whazor
5 days ago
The key is that manual coding for a normal task takes a one/two weeks, where-as if you configure all your prompts/agents correctly you could do it in a couple of hours. As you highlighted, it brings many new issues (code quality, lack of tests, tech debt) and you need to carefully create prompts and review the code to tackle those. But in the end, you can save significant time.
I disagree. I think this notion comes from the idea that creating software is about coding. Automating/improving coding => you have software at the end.
This might be how one looks at it in the beginning, when having no experience or no idea about coding. With time one will realize it's more about creating the correct mental model of the problem at hand, rather than the activity of coding itself.
Once this realized, AI can't "save" you days of work, as coding is the least time consuming part of creating software.
The actual most time-consuming parts of creating software (I think) is reading documentation for the APIs and libraries you're using. Probably the biggest productivity boost I get from my coding assistant is attributable to that.
e.g: MUI, typescript:
Tab. Done. Delete the comment.
vs. about 2 minutes wading through the perfectly excellent but very verbose online documentation to find that I need to set the "labelPlacement" attribute to "start".
Or the tedious minutia that I am perfectly capable of doing, but it's time consuming and error-prone:
Tab tab tab tab .... Done, with all bindings and fields done, based on the structure that's passed as a parameter to the method, and the tables and fieldnames that were created in source code above the current line. (love that one).
Yes, I currently lean skeptical but agentic LLMs excel at this sort of task. I had a great use just yesterday.
I have an older Mediawiki install that's been overrun by spam. It's on a server I have root access on. With Claude, I was able to rapidly get some Python scripts that work against the wiki database directly and can clean spam in various ways, by article ID, title regex, certain other patterns. Then I wanted to delete all spam users - defined here as users registered after a certain date whose only edit is to their own user page - and Claude made a script for that very quickly. It even deployed with scp when I told it where to.
Looking at the SQL that ended up in the code, there's non-obvious things such as user pages being pages where page_namespace = 2. The query involves the user, page, actor and revision tables. I checked afterwards, MediaWiki has good documentation for its database tables. Sure, I could have written the SQL myself based on that documentation, but certainly not have the query wrapped in Python and ready to run in under a minute.
what are you using for this? one thing I can't wrap my head around is how anyone's idea of fun is poking at an LLM until it generates something possibly passable and then figuring what the hell it did and why, but this sounds like something i'd actually use.
7 replies →
I think for some folks their job really is just about coding. For me that was rarely true. I've written very little code in my career. Mostly design work, analyzing broken code, making targeted fixes...
I think these days coding is 20% of my job, maybe less. But HN is a diverse audience. You have the full range of web programmers and data scientists all the way to systems engineers and people writing for bare metal. Someone cranking out one-off Python and Javascript is going to have a different opinion on AI coding vs a C/C++ systems engineer and they're going to yell at each other in comments until they realize they don't have the same job, the same goals or the same experiences.
Would you have any standard prompts you could share which ask it to make a draft with you'd want (eg unit tests etc)?
``` Run it through grok:
When I ACTUALLY wrote that code the first time, it took me about two weeks to get it right. (horrifying documentation set, with inadequate sample code).
Typically, I'll edit code like this from top to bottom in order to get it to conform to my preferred coding idioms. And I will, of course, submit the code to the same sort of review that I would give my own first-cut code. And the way initialization parameters are passed in needs work. (A follow-on prompt would probably fix that). This is not a fire and forget sort of activity. Hard to say whether that code is right or not; but even if it's not, it would have saved me at least 12 days of effort.
Why did I choose that prompt? Because I have learned through use that AIs do will well with these sorts of coding tasks. I'm still learning, and making new discoveries every day. Today's discovery: it is SO easy to implement SQLLite database in C++ using an AI when you go at it the right way!
That rely heavily on your mental model of ALSA to write a prompt like that. For example, I believe macOS audio stack is node based like pipewire. For someone who is knowledgeable about the domain, it's easy enough to get some base output to review and iterate upon. Especially if there was enough training data or you constrain the output with the context. So there's no actual time saving because you have to take in account the time you spent learning about the domain.
That is why some people don't find AI that essential, if you have the knowledge, you already know how to find a specific part in the documentation to refresh your semantics and the time saved is minuscule.
2 replies →