I came across a fascinating story this week that I wanted to discuss. It started when @adamwathan , the co-founder of @Tailwind CSS , joked on X that his decision to make every button bg-indigo-500 five years ago seems to have caused every AI-generated UI on earth to default to a shade of purple.
Stopped writing code the old way a while back. Not because I'm lazy, the bottleneck was never the typing.
Here's how I actually work now when building something like a data pipeline:
I don't start with code. I describe the input, the expected output, the constraints. Claude lays out the file structure and flow first. That alone saves me from building the wrong thing entirely.
Then I go piece by piece. One function, read it, test it, confirm the logic, move on. Bugs surface immediately instead of stacking up into that "it all worked separately but now nothing does" moment.
When I first started vibe coding, my instinct was to build broad all-in-one workspaces, but the AI agents kept drowning in context bloat and users got tired of prompt engineering their way through a generic interface. Once I pivoted to building hyper-specific micro-tools, such as single-purpose AI code helpers, dedicated document-parsing file tools, and simple automation utilities, the entire build process changed. The spec files stayed clean, Claude and Cursor stopped hallucinating, and the final products solved an exact problem right out of the box. Has anyone else shifted their vibe coding strategy away from big multi-feature apps toward single-workflow micro-tools?
Spent this week manually writing real content (facts, examples, categories) for 840+ programmatically-thin pages on a site of mine, after Google rejected AdSense twice for "low value content." Could've had an agent generate placeholder-style facts fast, but the whole point was making each page genuinely worth a visit, not just technically non-empty. Where's the line for you between "AI wrote this and it's fine" and "this needs an actual human judgment call"? Curious how others are drawing that line, especially on content-heavy or programmatic-SEO sites.
Vibe coding is amazing for getting from idea to working demo fast.
But I think there is a tricky moment after the demo starts feeling real.
The UI works. The auth flow seems fine. Maybe payments, webhooks, user data, background jobs, or database rules start getting added. At that point, the question changes from can AI build this? to can I safely maintain and trust this?
I feel a lot of projects get stuck in the middle too useful to throw away, but not reviewed enough to feel production-ready.
I recently built a browser-based tool that extracts text from screenshots using Tesseract.js, with all processing happening 100% client-side. No files get uploaded to a server, which makes it useful for sensitive content like code snippets, error logs, or internal screenshots.
I wanted to share the technical approach and a few lessons learned, and I'd love your feedback on how you'd improve it.
I've been letting Claude/Cursor do most of the typing lately and I noticed I stopped double checking what actually goes out in the API calls. then last week I caught myself pasting a real customer row into a prompt just to debug faster.
anyone else have a moment like that? where the vibe coded thing had a hole you'd never have shipped if you were writing it slower. secrets in a log, some data you didn't mean to send anywhere, whatever. My product is focusing on this and I would like to know is this a serious issue?
I'm asking because I honestly don't know if this is just my perspective as a software engineer.
I've tried several AI app builders, and while they've improved a lot, I still find myself hitting the same walls. They often feel rigid, expensive as projects grow, and I rarely feel confident I could build and maintain a production-grade application with them alone.
Early on I asked Claude Code to recommend an auth setup and just went with what it suggested since I wanted to move fast. It picked a provider that was fine for the first few hundred users, but their pricing tier jumps hard once you cross a certain MAU threshold, and we hit that wall faster than expected. Now I'm staring down a migration that touches every session and password hash in the system, during a period where I really cannot afford downtime or angry users getting logged out randomly. For anyone who's picked infrastructure based on an agent's recommendation rather than researching it themselves, how did it hold up a year later? Where do you draw the line between letting the agent decide and doing the research yourself first?
