Cast by jrf
tldr: i went from max bullish to max bearish on vibecoding while building @atlas real software is built by real developers and that's not really debatable exceptions to the rule: internal tools, data / ai analysis apps, single flow apps w no persistent memory thoughts?
18 replies
Time to harvest your activity! 🌾💰 The new Rewards Hub is distributing tokens to all active users. 🔹 High Activity = Higher Tier 🔹 Old Account = Multiplier See what you’ve earned so far: 👉 farcaster.xyz/miniapps/3sEURhbb3FFe/farcaster-rewards-hub 🚀 ref: 60284
Time to harvest your activity! 🌾💰 The new Rewards Hub is distributing tokens to all active users. 🔹 High Activity = Higher Tier 🔹 Old Account = Multiplier See what you’ve earned so far: 👉 farcaster.xyz/miniapps/3sEURhbb3FFe/farcaster-rewards-hub 🚀 ref: 78808
I felt the same way when rebuilding /moxito I am taking a different approach where I’m focusing on prompts first and getting the foundations set for: - The app vision - The architecture - Focus on quality and CI checks and so on - Being precise on what I really want and iterate I learned that letting AI decide everything was a big mistake on my end. For a smaller app, it works well, but there are some essentials that need to come from the dev(s)
big agree ai coding tools want to write code they want you to keep needing them and paying for credits they don't care how much complexity or bugs they add bc their motto is *don't worry i can fix it!*
They fix it but can add more bugs along the way. I let them review as much as they want but only stop and fix if it’s actually critical. Huh that makes absolute sense. They want you to keep using them unless explicitly told. Ideally I’d like a system that I own and that I built, that is not traced, tracked and does absolutely everything that I require for each project. Maybe that’s asking too much and could take months to build..?
no time at all i wrote my first lines of code in nov 2022 copying verbatim from chatgpt to vs code
Ah yeah then this makes a lot more sense and I agree. Even with the major advancements we’ve seen in AI coding tools, a deep, fundamental understanding of how to develop software is still necessary to roll production apps
You probably meant something more like consumer products than real software, so I won't push back too hard. That said, it's a limited definition of "real software." That I constantly hear of real devs vibe coding tells me it's a skillset worth climbing the learning curve for.
i ofc think ppl can build apps now i have some basic coding familiarity and i've built mini apps used by tens of thousands of ppl but that's different than building real software proposed definition of *real software* : software that can be built on scalably and extended without rewriting the whole thing p.s. i'm very reluctant to listen to vibecoding advice from real devs who know how to evaluate architecture and code snippets (i.e. code it themselves if needed), i don't think real devs even understand the term *vibecoding*
man you wouldn't believe this, but all software is real software 99% of the mega-corps an unicorns of today started with non-scalable shitty codebases that they outgrew and rewrote 100% when it made business sense PMF > Tech debt
My definition of real software is much broader, but I also think software is and should be seen as art. Your point is taken, I just think you discount use cases with a term like that! Why not say consumer-grade or something? And that's a keen insight re: your ps. It took me a long time for vibe coding to fully click with me as a framework. Still weird though lol
The closest I can to actual vibecoding was last week trying to get Rust to deploy a smart contract to Base using Alloy. It was like asking your assistant that knows as much as you in a couple areas and less in most to give suggestions. I did get there, but only by fighting the code and fighting AI until I found, bit by bit, things that worked that contradicted the documentation, all project provided examples, and every suggestion by AI. But AI provided additional things to try that helped me find the real answer, more from having someone to talk to about it than being helpful.
i think for you it's a tool for most ppl it's a mess maker and similarly i will still use vibecoding software to work on frontend, which i can reliably do without breaking things (if i check out a branch and make a pull request) related: farcaster.xyz/jrf/0xdedd20a0
This is accurate, I think, both what you said about it being a tool for me and a mess maker for most people, and the statement about real devs not really understanding the term vibecoding. And I fall in a weird category, I taught myself BASIC from a book at 4, and took classes off and on through the end of hugh school, then got my COSC degree in more old school type programming (in PASCAL, C, C++, LISP, assembly, etc), but my career has been in infrastructure, sysadmin, networking, database, cloud based devops. Most of my programming has been either scripting or small projects in work, and more backend type things with raw HTML or CLI as front end. So I tend to approach everything as a tool for a task.
claude code / opus made me *double down* on vibecoding in july of this year but i believe vibe coders hit an inevitable brick wall and claude code's response: "i can fix this, don't worry!" i still pay for claude max and will use it for prototyping related: farcaster.xyz/jrf/0xdedd20a0
I do think you have to be able to go deep enough with what it is building to orient it towards the right things and away from the wrong things. And I think debugging / security review agents are going to help a lot too. Harnesses up next
AI-generated code allows me to multiply the code I "produce" concurrently, but always has to be gated on its way to production by a human engineer who understands the nuances of the product and how it integrates with the broader ecosystem, because the need to understand that scales with the complexity of the solution. Maybe it's just a matter of us humans needing to brain dump more into documentation so that agents can be trained on that information to lessen the need to have a human be present to make sure those nuances are accounted for in the generated code, though.
I will say that I am excited for the potential of vibe coding to allow people with no programming skills to automate smaller, low- or no-risk processes and make their lives easier or more fun.
Agreed if you’re purely just vibecoding and giving prompts without really paying attention to the code itself If you feel like you’ve hit a plateau though I’d encourage you to look into agentic coding using Claude Code or something similar You can definitely build maintainable apps with agentic coding, just takes some practice
i use claude code exclusively it's not as good as we all want it to be
Your mileage may vary I guess, a lot of different way to use it I think it’s pretty exceptional if you have the right subagents and skills setup
Thought: You're absolutely right! Jk, yeah, experiments, short timeline it help. But as soon as you want to go forward and scale to implement more things, if you keep vibe coding then you'll require exponential time on the go. You can't sail without knowing where you're going and how you’re going. As soon as you know that, then you'll require exponential can delegate part to AI to help you going. But you need to know and oversee. Without that, it's just randomly sailing and following the wind. You may land where you want, but god damn, that will be long. Delegate the right thing.
perfect metaphor and it's ironic also bc the thing most ppl struggle w in vibecoding is that they are sailing w a destination in mind but all the foundational stuff they ignored leaves them lost at sea need to go back to port and find a crew who know how to actually sail and what the different parts of the boat do
Exactly, the destination is important But if you are going to sail by yourself with autopilot, then you still need to know a bit how the weather and your boat is working! And the metaphor can go forward, with small stop on the journey. Step by step forward! (Yes, I love boats.)
I use it as a tutor, when I first dived in it was just pushing buttons and then realised my git was a monstrosity and I hadn’t actually learnt anything. Now I approach things more like a tutor where I ask what and why a lot more
that is definitely the best approach i wish i had done it more over the last three years but i have been mostly concerned with UX which is still what i hope to do, and will use claude code to help with rapid UX prototyping
I agree but I am very bullish in what it’s done for lowering the bar to entry for people who would never have tried to begin with
100% that's a good perspective, too and better coding frameworks and SDKs will emerge based on where coding LLMs are today
From my experience over the last year, you gotta at least understand how the things you’re using work (and work together) or you’re gonna get rekt and have a bad time after a certain point.