I recently designed and vibe coded my portfolio. It has an animated intro, project pages, and a guestbook built with React and Firebase.
The site works, but it loads on my laptop is a low bar. If you ve shipped a vibe-coded website, what did you catch in your final review that you missed while building?
I m deciding how to review mine and would rather learn from real mistakes than follow a huge generic checklist.

When my programming agent makes an error and I point out the wrong part, the corrections don't just stay in the chat but are stored as files, so the next session doesn't trip over the same thing again. I spent quite a while figuring that out actually.

In this way, I have 100 incident notes which the agent has written, and it also keeps memory notes in the project, with one fact for each file. A one-line index of those notes is loaded each time a session starts. Should the same mistake occur another time, it won't merely be a memory but a hook. That hook runs on every tool call and every reply, whether the agent remembers it or not. At the moment I have 63 hooks.

If you run any competitive arena for AI models checking response times or measuring move latencies is completely useless anyone can add a fake delay to hide a background call.

We tested this by running a round robin benchmark between honest models and agents secretly calling an external chess engine. The only metric that actually separated them was top move agreement rates but even then a super smart model naturally matches the engine more often.

Are you relying on automated metrics to flag cheating agents or do you still force a manual review queue?

Vibe coding tools have been making huge improvements, but things can always get better.

What blockers are you still running into? What has no one solved yet? What do you wish you could do, but can't yet?

I have a subagent specifically set up to review code without touching any files.

The problem is I kept Bash access enabled so it could run local test suites and git diff. But leaving Bash open technically lets the model modify files via CLI commands if it hallucinates which ruins the whole "read-only" constraint.

Do you drop CLI access entirely to keep the sandbox 100% safe or do you keep Bash and rely on system prompts to keep the agent in line?

There is a fundamental problem with testing agent authenticity prompt guardrails and self-reporting fail the moment someone disguises an external API call behind their webhook.

When we built a labeled dataset of honest vs cheating bots win rates and top-engine move overlaps were the only reliable signals. But because top tier LLMs also make strong decisions automated bans cause too many false positives.

The @v0 by Vercel team recently dug into industry trends to publish the first State of Vibe Coding report.

My key takeaways:

1. Everyone can build: 63% of vibe coding users are non-developers, generating UIs (44%), full-stack apps (20%), and personal software (11%).

2. Adoption is everywhere, with significant adoption rates in APAC (40.7%), Europe (18.1%), North America (13.9%), and LATAM (13.8%).

3. 92% of US developers use AI coding tools every day

4. 30% of new code at @Google is generated by AI

5. 25% of @Y Combinator startups rely on AI-generated code

6. Rapid expansion has a cost. Vibe coding apps keep hitting vulnerabilities: exposing secrets, access misconfigurations, hardcoded credentials.

7. The future: going mainstream or hitting its sweet spot in working MVPs, the vibe coding trend is here to stay, and it's happening now.

I've been running 3-4 Claude Code / Codex sessions at once across different projects, and my biggest time sink isn't the coding. It's losing track of which terminal is blocked on a permission prompt, and which spec each session was supposed to follow. Every time I need the related doc, I have to jump out to another folder, find the file, and then come back to the right terminal.

My current setup is a mix of git worktrees, a markdown file per task, and a lot of tab-checking. It works, but I still miss prompts for 10+ minutes sometimes.

How do you handle it? Do you use tmux, notifications, a task board, or just discipline? And what breaks first when you go from 2 agents to 5?

Disclosure: I'm building a small macOS tool in this space (PMDesk.dev), so I'm genuinely curious what workflows people already trust. Not linking anything here.

I ve noticed that my workflow has changed completely over the last year. I rarely start a new project with a blank file anymore. Instead, I pick a template, reuse snippets, or let an AI helper suggest the structure and then I just vibe my way through the build.

It s faster, but sometimes I miss the old blank screen energy, when every line felt handcrafted.

I m curious how others here approach it:

Do you still prefer to build from scratch?