I don’t want to learn how to vibe code. I already learned how to code and I don’t want to let AI decided things for me. I want to learn how to use it as a coding assistant, as a coding tool.
I have a Copilot license bought with my office 365 suscription but if the principle apply the same, I could learn on a course designed for another coding assistant.
Let me try to add some nuance to the comments here. Mostly they are right, Copilot is not the right tool. But if you understand why you might still get some use out of it and prevent yourself from building up bad habits around it for when you get a better tool or Microslop improves it.
- It’s dumber. OpenAi, Anthropic and the Chinese alternatives are betting their entire companies on the AI race. MS, Google and Meta may claim they do but they’re still playing it safe. They probably also want to pick up the scraps of the AI companies that go down. So while the older companies want to make a profit or at least not make too much of a loss the new AI only companies only care about cashflow, big usage numbers and usage data. More money means more servers means smarter models.
- It’s not for programmers. The model needs to be smart but it also needs the right tools. Look up what a an AI harness is. Copilot’s harness is trying to be generic and connect you to other Microslop services. It’s not primarily a coding harness.
- It has a PR problem. This one is not a comment on the actual capabilities of Copilot but rather on the sentiment in the comments. Copilot IS getting better. It IS starting to be useful, even for coding. But MS tried to push it in everyone’s faces before it was any good at all. They made existing products worse because of it. it left a bad taste in people’s mouths and you can tell from the comments here that some people, especially technically minded people, hate it.
That explains it well. Thank you. I didn’t like having to pay for it either or that I struggle to find any description of the MS office suscription that wouldn’t describe solely how Copilot is included. By I thought that I could use it after all and if my IDE can automatically generate getters and setters and modelisation software generate code from UML, I could find a way to use it in a way that is helpful.
But I guess it was the wrong tool…
Crazy thought. Ask Copilot the question using that criteria.
Lmfao right? It’s almost like an advertisement.
Professional coders using AI don’t use copilot at all. They either use Claude or roll their own. Copilot isn’t real. It can’t hurt you. Ignore it.
Thank you for the advice.
Do you know a good course for using Claude then?
I haven’t taken a course, but the main thing I’ve learned that starts making agentic coding useful is to setup an AGENTS.md file for the project. Once you’ve documented the goal, style, standards, requirements, testing methodology, etc then the AI agents start to produce more consistent outputs that fits what you wanted.
I note that. Thank you.
Step one: don’t.
I tried to get it to vibe code me a macro in Excel. I didn’t think it was that complicated. It confidently gave me code. I ran it. Error. Gave it the error. It said “oh that makes sense, here’s how you fix it.” Um, couldn’t we have just opened with that? Fine, copy, paste. Run. Error. Gave it the error. It said “oh that makes sense, here’s how you fix it” and then gave me obviously incomplete code. I asked for the code in a .txt file. It gave me a Word document and the code was… not good. It didn’t even LOOK like it was gonna work. I asked it how I was to copy that to Excel and it admitted I wouldn’t be able to. It basically wasted my time apologetically.
Note that if I had just told Copilot to do what I want, it absolutely woulda done it. I told it I wanted what it could do in a macro so I could repeat it without going through Copilot. That was what it could not do.
“Is God so strong that He could make a boulder that He could not lift?” Similar situation.
I wasn’t even trying to push Copilot too hard.
I wanted it to:
One. Analyse one column, which had FIRSTNAME LASTNAME and change it to LASTNAME, FIRSTNAME.
Two. Take a 12-digit number and apply BOLD formatting to the last four digits
Three. Couple other similar functions.
It choked.
So again… Don’t.
Also, can you change how the agent behaves? Our Copilot used to have this option. I’m thinking you’re not gonna heed my warning. So try this. If you have the option to customise your agent, tell it to think deeply about each step. Tell it that its goal is to produce efficient code and that you do not expect it to be nice to you. Straight up tell it to be honest and frank with you. Tell it to not lie to you. Plus deep thought, minus sycophancy. Best of luck with your project.
Edit: Formatting. Lemmy’s markdown interpreter might have been coded with Copilot.
I had chatGPT help me make a proof of concept for a project I gave up on a while back. It’s great for showing “this can be done”. Where it falls apart is “this is how you do it”
AI is getting better all the time. People are using it to vibe code ports of PC and Xbox/PlayStation games to PC (if they weren’t there before) and Mac/Linux, which game developers typically overlook, even, unfortunately, indie publishers.
Windows and Linux are vibe coded now. I mean the operating systems. Linus Torvalds made some good points about vibe coding that are worth looking at. He’s saying that if you use an IDE (application that turns human-readable code into machine-readable code), you should be willing to use AI to vibe code, but — and this is the important part — you still have to know what you’re doing. That’s what a lot of people are overlooking. They’re saying “oh no the sky is falling, Daddy Linus is using AI to write the Linux kernel.” But Torvalds is likening AI to autocomplete. It’s writing the code, but he’s verifying it. So he says — he is getting up there in age. I’m also not saying he’s 100% right. I’m just saying maybe he’s not 100% wrong. Maybe he has a point.
But, for now — especially if you aren’t paying — AI has a long way to go before it does the kinds of things they’re saying. As they develop new models, they make the old models free. I’ve never paid for AI and I don’t think I ever will, but I cannot predict the future. It’s always telling me it can do more for me if I pay, but the ones I talk to are sympathetic (or at least sycophantic) to my unwillingness to pay. They’ll even tell me how to skirt the paywall, to an extent. Like “if you say this, I can do it for free.” But once the chat has involved an image or an uploaded file, you only get so many queries a day. So then you hit another limit and the software won’t let you continue talking to it until your usage resets.
