Bartek Kunowski’s Post

Three months ago I would have flagged vibe coding as a red flag in any codebase. Then I shipped a working integration tool in 4 hours. My usual estimate was 3 weeks. I still cannot explain half the code. But the tool runs fine and my clients do not care. That forced me to ask a harder question: is software engineering about understanding how it works, or making sure it keeps working? Curious how other founders and builders are thinking about this.

A team of engineers gets paid to delete your code. $10,000 a week. The project: a 100,000-line AI-assisted codebase. The goal: 35,000 lines, same functionality. They use agents to find and collapse the redundancy - GitClear tracked the same pattern across 623 million code changes, block duplication up 81% since 2023. But the first thing Slopfix does before touching a single line is write down what the app is supposed to do. Every screen, every endpoint. A regression checklist. Written after the ship. The teams paying for this didn't skip it on purpose. The coding agent was fast, the checklist felt like friction, and by the time the codebase was three times larger than it needed to be, nobody wanted to go near it. The cleanup crew writes your test spec retroactively. The market price for that, apparently, is $10k a week. Link below

I told my team to stop coding and answer four questions first. Not because they weren't shipping. They were shipping more than ever. Claude handles the code. Tests pass. PRs get merged. Everyone feels good. Then I check what actually moved our numbers. And it's a fraction of what went out the door. This kept happening. So I started doing something that felt annoying at first. Before any feature gets built, I ask for a one-pager. Just four questions: What are we building? Why does it matter? What metric moves if this works? What user behavior should change? I get it. Writing a one-pager feels like slowing down when you could be shipping. But here's what I noticed. When someone knows what behavior they're trying to change, they stop building the feature everyone asked for. They build the version that might actually work. Still figuring out if this scales. But so far, the features that start with those four questions look different than the ones that don't. Comment "one-pager" and I'll send the template & sample your way

Coding agents have changing a codebase really fast. The harder part now is keeping up with the speed and volume of changes that accumulate, knowing what to pay attention to, and maintaining the right depth of understanding as generation happens. This includes understanding what changed, why it changed, and which details actually matter. Today we’re launching three new features in Nuanced: 1. chat directly with a selected part of your spec 2. see spec changes as they happen, and  3. use Spec Tour to walk through a dense spec without reading the whole thing at once. As code gets faster to generate, developers need a faster and more intuitive way to understand the system and keep agents on track. We're working toward that goal at www.nuanced.dev!

I vibe coded a tool this weekend. But vibe coding is creating false expectations. Vibe coding is brilliant. I think it’s fun. But when you can add a feature in minutes, it has you thinking this is easy. So why is [blank] app so rubbish? Adding a feature to an app I just vibe coded for me, on my personal computer, that no one is using is one thing. Adding a feature to an app used by actual customers isn’t a prompt, a cuppa, job done. 😱 It’s research, design, the build (maybe involving multiple teams), testing, release governance and a need to be prioritised ahead of a host of other ideas competing for the same engineering resource. I’m not saying it’s impossible to go faster. I’m saying right now for many organisations it’s 100% not a prompt, a cuppa, job done. And nor would we want it to be. ⬇️ Has vibe coding changed expectations of your team? Enjoy the vibe! 😎

This Claude Code workflow cleaned up my changes fast. I started treating Claude like a senior dev pair. I used to let Claude Code make a few edits, then review everything at the end. The code looked fine at first, but small mistakes stacked up quietly. What changed was making the loop brutally tight. One change. Typecheck. One focused test. Then continue. Here are the Claude Code tricks I wish I used earlier: ☑ Put repo rules in CLAUDE.md before touching code ☑ Ask Claude Code for one change at a time ☑ Run typecheck after every edit, not every task ☑ Run one focused test before expanding the scope ☑ Use isolated test environments Claude can execute safely ☑ Let Claude write tests for your framework first ☑ Save progress, then continue only after green feedback The output got cleaner the moment the feedback loop got shorter. I made a free workflow guide. Comment "CLEAN" and I'll DM it. Is your team checking Claude Code changes after every edit, or only at the end? 1 or 2. #ClaudeCode #ClaudeAI #SoftwareEngineering #EngineeringManagement

Everyone's debating whether vibe coding is real engineering. We stopped debating and shipped it as a product feature. Users describe what they want in plain language: a filtered view, a custom report, a specific slice of data. The platform generates it on the fly. No SQL exposed. No schema leaked. No business rules bypassed. Just a GraphQL query, scoped and sandboxed, doing exactly what the user asked within exactly the boundaries we defined. The vibe coding debate misses the point. The interesting question isn't "is it real code?". It's: can you give users the power to build for themselves, without giving up control? Turns out, yes.

When I first saw all the debates about vibe coding, I honestly wondered... "Is it really that bad?" Maybe I thought that because every vibe-coded project I'd seen was created by software engineers. Sure, the code wasn't always pretty, but they still understood architecture, security, data models, and deployment. Then I recently reviewed a project built by someone from a business background. Great domain knowledge. Zero software engineering experience. (Let's just say they couldn't tell JSON from Jason.) That's when it clicked. The app worked. Until you looked under the hood. No backend. No security boundaries. No proper authorization. No data model. No deployment strategy. No audit trail. Personal data exposed. Every feature added another layer of technical debt. The scary part? It looked finished. Vibe coding is an incredible way to validate an idea or build an MVP. But shipping software that stores people's data is a different game entirely. Software engineering isn't just writing code. It's everything that stops the code from becoming tomorrow's disaster. AI has dramatically lowered the barrier to building software. It hasn't lowered the bar for building software well.

The cost of writing code is approaching zero. The cost of building is different than the cost of writing code.

The cost to build is not zero. Try building something and then get back to me. Did it just work? Did it survive exposure to your first customer? How about your second? How about your 100th? Does it scale beyond that? When you go to extend it, is it extensible or do you have to refactor everything? How good is the automated testing coverage? If you let a coding agent define and write all the tests, they probably aren't very good. The cost to build *a* thing is getting cheaper. If the result is we build *many* more things, then the cost to build and maintain *things* is going to skyrocket. Proceed with caution.