Eric Roby

I’m still confused by the whole “vibe coding” concept. When I’m working with Claude or Codex, I push back constantly, especially during the planning phase. We go back and forth on architecture, product decisions, and implementation before any code gets written. Ironically, I spend far less time reviewing the code itself than I do challenging the plan. So I’m curious… Do vibe coders just let the AI rip and hope the database schema, migrations, architecture, and everything else work out? Or is “vibe coding” mostly being used to build demos and static websites rather than real production applications?

I think one of the strangest parts of AI coding is being able to build something faster than you've fully decided what it should be. I'm used to implementation taking enough time that I could keep thinking through the product while the work was happening. Now I can give an agent a substantial task and let it work without being hands-on the whole time. That changes where I need to spend my attention. I need a much better understanding of the product. What we're trying to accomplish, how the feature should work, and whether the next thing I'm asking for makes sense. I think this is why the job is starting to feel more like product engineering to me. The implementation can move quickly, but I'm still responsible for figuring out what we're trying to build.

I used to default to web apps because I wanted my tools available when I was away from my computer and it was what I was familiar with But building native apps has changed how I think about that. For my own tools, I like having more control over how they work with the machine and where I can improve performance. A Mac app doesn't have to go through the Mac App Store. And on iOS, TestFlight gives you a way to test your own app on your phone, with each build available for up to 90 days before it expires. I have my internal setup connected through Tailscale. I kind of assumed wanting mobile access meant everything should be a web app.

I started building all my internal tools as separate web apps on different ports. Then I was talking to Adrian who said he now builds native apps, and I realized I hadn't even considered doing that. I'd never built native apps before. It just wasn't where my brain went when I wanted to make something for myself. With AI, that changed. I took design ideas from Claude and ChatGPT and built my own internal app around what I needed, including connecting it with applications already on my machine. It's funny how much of my original decision was just familiarity. I knew web apps, so every internal tool was going to be a web app.

I always work in a separate worktree, so if something is off I can just scrape it If I realize we're going down the wrong path, I'll stop the agent, get rid of that attempt, and start a new one with more information about what I actually want. I'd rather give a fresh agent a clearer task than keep trying to steer an implementation I already don't want.

Not why the top headlines keep saying “Karpathy built YouTube in 2 Hours” He built a UI to add his own browse list and removed thumbnails (which idk if that’s even a good idea) Everything else still uses … YouTube

I really need to automate this ... Sometimes I’m working through a plan with an agent and it keeps asking me questions with a recommended answer. And I keep choosing the recommendation. After a while, I start wishing it would just use that recommendation by default Most of the time, that’s what I want. I’m asking the agent to help me figure this out, so I’m usually happy to proceed with the approach it thinks makes sense ...

I’ve sold over 200,000 courses on Udemy, and I keep thinking about what education looks like beyond the video course. Udemy already has interactive formats like AI Role Play. But I’m especially interested in what we can build with interactive software and 3D worlds. AI can help answer a question fast. But, I don’t think getting an answer is the whole learning experience, though. I want people to enjoy exploring the idea and want to keep going. Being able to build these experiences with agents is a big part of what excites me right now. It makes me think about lessons I’d want someone to interact with, instead of only watch. I really do think more education is going to move in that direction. Fun and learning together, with software designed around the experience.

I’m definitely procrastinating by building internal tools. I’ll build something to help me get work done, then have so much fun adding features that I spend more time improving the tool instead of doing the work I built it for. And because I’m building something useful, it’s pretty easy to justify. The other thing I keep telling myself is that these subscriptions might not stay this good forever. So part of me wants to build as much as I can while I have this level of access. I don’t know if that’s a smart way to think about it or just another excuse to keep building. Either way, I really enjoy making these tools!

Once an agent and I agree on the implementation, I want it to need very little from me. For the work we’ve already agreed on, my expectation is that it keeps going. The interruptions, where the agents reaches out to me, that make sense are actual access problems. Maybe it needs computer use but my screen is locked. Maybe the database URL I gave it doesn’t work. That’s something I need to help resolve. But if the environment is available and we’ve already worked through the plan, I don’t want a stream of notifications asking me to keep the task moving. I want to come back to the completed work, or a clear explanation of what genuinely blocked it.

So today is the day to buy more Google stock?

I don’t give every model’s opinion the same weight right now. My usual approach is Opus 5.5 on high effort for the reasoning and review, then medium for implementation. I’ll usually go back to high for another code review. Even then, I’ll sometimes have another agent verify the work. If I use an OpenAI model, I ALWAYS want that second check. That’s where my confidence is with these tools at the moment.

Do you have AI connected to your email, Google Drive etc? I don’t but really thinking about trying it out

I was just asked yesterday what do I do when an AI provider goes down ... If one AI provider goes down, I’ll usually just switch to another subscription and keep building. I already have a couple connected through T3 Code, so it doesn’t slow me down too much. I definitely have models I prefer. But if the one I want isn’t available, I’d rather keep moving with another option than wait around for it to come back. That’s another reason I like having more than one provider available in the same place.

I could only imagine, that one day there will be a single programming language that agents are best at.

It’s funny how the unglamorous parts of building an app are still unglamorous with AI. Accessibility, error recovery, performance problems, etc. I’m happy to have an agent help me work through them, and I like how quickly I can get fixes. But they still aren’t the parts I’m (or anyone really) is excited to spend time on.

“We want to reserve human time and attention for things only humans can handle” - great read here

Token maxing is real for me. I have two Claude Code subscriptions, and I get a little anxious when one is nearing its weekly reset with a decent amount of usage left. I’ll start ramping up bigger features, more testing, or performance work to try to use as much of it as I can before the reset. I know that probably isn’t the best reason to choose what to work on next. But I still catch myself doing it. There’s this feeling that I’m wasting something I already paid for if I leave capacity unused. Then I end up choosing more work partly because the usage is there. I just haven’t figured out how to switch that part of my brain off.

I’m getting less interested in building a completely separate workflow for every model. I use both Claude Code and Codex. I want the same core skills and project instructions to come with me when I switch between them. Mostly because I use T3 Code now. If I’ve already worked out how I want something done, I don’t want to rebuild that whole process because I’m using a different agent today. That is why I spend a lot of time making sure my agent files and workflows are agnostic

“I didn’t know php existed until yesterday morning” These are the type of experts random companies hire as cto