The Setup
I've been vibe-coding for a few months — building apps with Cursor, shipping fast, feeling great about it. Then I started wondering: is any of this code actually any good?
Professional code review tools exist, but they cost $24/seat/month and produce reports that read like compliance audits. I wanted something that would give me a straight answer — and maybe make me laugh while delivering bad news.
So I built RoastMyCode.ai. It analyzes any GitHub repo and gives it a letter grade (A+ through F) with specific findings, one-liner "burns," and fix suggestions.
Then I pointed it at 40 repos — everything from React and Next.js to explicitly vibe-coded projects and AI-generated Uber clones.
Here's what I found.
The A-Tier: Suspiciously Clean
create-t3-app (A-)
The verdict: "28,775 stars and they left ONE console.log? This is the coding equivalent of finding a single crumb on a pristine white carpet."
The best burn: "5,460 lines of code and the only sin is a single console.log. This developer either sold their soul or is secretly three senior engineers in a trench coat."
One issue found. One. The T3 stack earned its reputation.
Next.js (A)
Even our analyzer struggled to find things to complain about. Lesson: projects with strong opinionated structure score well because the code has guardrails built in.
The B-Tier: Solid But With Character
React (B+)
Yes, I roasted React. The verdict: "244k stars, 1 eval() — even Facebook can't resist the forbidden fruit."
One unsafe eval() call in 244K stars of code. Meta's engineers are annoyingly competent.
chatbot-ui (B+)
The most-cloned ChatGPT UI on GitHub (33K stars). The verdict: "33k stars for leaving console.log in prod? The internet has questionable taste."
Best burn: "We sent our best bug hunters into this codebase. They came back with two mosquito bites and existential dread."
bolt.diy (B-)
The open-source Bolt.new fork. 19K stars. The irony of a vibe coding tool getting roasted by another AI was chef's kiss.
The verdict: "19k stars, 5 issues, 15k lines — either you're TypeScript wizards or the bugs are really good at hide-and-seek."
Best burns:
- "Found XSS in avatar fallback — apparently even profile pictures can't be trusted"
- "NetlifyTab.tsx is so large it has its own ZIP code and congressional representative"
- "Using 'any' type in TypeScript is like buying a Ferrari and removing the engine"
claude-task-master (B)
Every Cursor user's favorite task manager. The verdict: "This codebase is so clean it made our bug detector file a harassment complaint."
Then it found a memory leak in the destroy method: "Even your cleanup code needs cleanup."
The D and F Tier: Code Quality Emergencies
Onlook (D) — 25K stars, YC-backed
The verdict: "25k stars but still writing 600-line God files and leaving console.logs in prod like it's 2015."
7 issues found. Two God files over 500 lines each. This is what happens when a project grows fast without architecture reviews.
openv0 (F)
An AI UI component generator. The verdict: "Nearly perfect AI playground, but running eval() on GPT output is like giving your toddler a chainsaw."
Code injection vulnerability via eval() on untrusted OpenAI API responses. This is the kind of bug that ends up in a security blog post.
The Patterns
After analyzing 40 repos, five clear patterns emerged:
1. Error handling is universally weak
This was the #1 issue across almost every repo. AI-generated code handles the happy path perfectly and skips the sad path entirely. catch (e) { } — the sound of errors being silently murdered.
2. Architecture degrades past 50 files
Small AI-generated projects are surprisingly clean. But past ~50 files, the architecture starts collapsing. AI doesn't maintain a mental model of the whole system the way a senior engineer does.
3. Console.log is an epidemic
Almost every repo ships debug logging to production. It's the #1 most common finding — breadcrumbs from development that nobody swept up. One tool called it "turning your server logs into a developer's personal diary."
4. UI code is consistently better than backend code
AI tools are genuinely good at generating frontend code. The visual feedback loop helps — you can see when something is wrong. Backend code doesn't have that luxury.
5. Established starters produce measurably better code
Projects that started from create-t3-app, next-forge, or similar opinionated starters scored 20-30% higher than from-scratch projects. Starting with good patterns matters more than which AI tool you use.
What This Means for Vibe Coders
This isn't an argument against vibe coding. Half these repos scored B or above — that's solid code by any standard.
But there are practical takeaways:
- Use established starters. The structure matters more than you think.
- Get some form of code review. Even automated. The AI is great at generating code; it's not great at questioning its own work.
- Pay extra attention to error handling. This is where AI falls down hardest. After every AI coding session, search for empty catch blocks.
- Watch the architecture past 50 files. When your project grows, take time to organize. AI won't do it for you.
Try It Yourself
RoastMyCode is free for all public repos. Paste a GitHub URL, get a roast.
I'm genuinely curious what grades your projects get. Drop your results in the comments — I'll discuss any interesting findings.
Top comments (4)
roasting repos is a mood—especially when you see the chaos behind the curtain. i’m out here doing the same with cursor and honestly, the vibe usually beats a perfect architecture when you just need to ship. since you’re already building these types of tools, you should check out stackapps.app—it’s a solid spot to get your indie builds discovered. full disclosure — i’m a paid creative scout for the stackapps team. austin taught me: just start the thing.
Ha, "the vibe usually beats a perfect architecture when you just need to ship" — that's exactly why I built this. Ship first, then find out what's on fire.
Thanks for the tip! Just listed it on StackApps: stackapps.app/app/3DtbJ5wiwB2GMEW5...
let me be clear Ilove this Its so enjoyable and helpful I am goning to spen the rest of my night roasting my repos! great job!
also it deleted my link I will try it again roastmycode.ai/roast/latest/johnoh...