The latest episode of the VergeCast revisited the "Vibe Coding Challenge," where host David Pierce and colleagues Jake Castronakis and Hayden Field spent a month using AI coding tools to develop websites addressing personal problems. While David was "unreasonably proud" of his creations, his colleagues recounted more frustrating experiences.
Before diving into the challenge, David provided "90 Seconds on The Verge" news updates for August 10th, 2026:
* Mark Zuckerberg published a 6,500-word manifesto on AI's future, favoring government oversight without banning foreign models. He highlighted "invention, not automation" as superintelligence's greatest contribution.
* Big news in the App Store world: Aptoid, the first non-Google Android App Store (focused on games), is now available in the Play Store. This follows legal battles with Epic (Fortnite creator) and signifies a major shift in the app economy, mirroring an ongoing fight with Apple and iOS.
* Verge contributor Georgiana Torbit reported on Reflect Orbital, a company that won FCC approval to launch a "mirror satellite" designed to reflect sunlight onto Earth for light or solar power. While currently one satellite, Reflect aims for 50,000, raising concerns among astronomers about light pollution and scientific observation.
**The Vibe Coding Challenge Check-in**
Jake and Hayden both described their experiences as far from "fun and fulfilling," often involving "all caps yelling" at the AI. They found the tools autonomously tried to do too much, often going down "the wrong road."
**Jake's Project: "Farm Share Share"**
Jake aimed to solve the problem of receiving too many unwanted vegetables from his farm share (CSA), particularly "dandelion greens every week." He envisioned a website where friends could claim his excess produce.
* **Tools Used:** Switched from Claude Code to Codex (OpenAI's coding AI).
* **Experience:** Described using Codex as "horrible." He constantly had to start new chat threads because Codex would frequently encounter access issues or "give up," stating it couldn't proceed.
* **Design:** The initial logo concept went "photo realistic" and "tacky," leading to a minimalist, geometric design that Jake ultimately settled for, resembling "blobs and lines."
* **Functionality:** The site features a "friend passcode" login (which Jake noted was AI-generated copy), displaying available veggies for friends to "claim." An alleged notification system is supposed to alert Jake when items are claimed, but he hasn't tested it.
* **Major Setback:** Jake couldn't publish the website. Despite ChatGPT itself suggesting the "ChatGPT Sites" plugin for publishing, Codex consistently failed to work with it, citing permission issues. Jake concluded that the site is currently only functional for him.
**Hayden's Project: Personal Portfolio Website**
Hayden's goal was to update her "embarrassing" personal website from 2022.
* **Tools Used:** Started with Claude Free (found it "horribly ugly" and quickly ran out of tokens), then switched to Gemini, which became her primary tool.
* **Experience:** Found the process frustrating, leading to moments of "yelling in all caps" at Gemini.
* **Design:** Requested a "sleek" site with "nineties, like Easter egg elements." Gemini delivered a "Nineties Mode" with a teal and pink color scheme and a chunky, star-emitting cursor.
* **Challenges:** Gemini struggled with simple design requests, like a "cloud-shaped" button, producing "crazy cutouts insane" results. Ironically, it nailed a more complex 90s-style button. Hayden also faced significant difficulty with Gemini scraping her published articles, forcing her to manually compile links from "29 pages of pieces." Additionally, Gemini's "guidelines" prevented it from embedding YouTube videos, even from The Verge, despite embed codes being readily available.
* **Publishing:** Successfully published her site using GitHub Pages and her existing domain, moving away from Wix.
**David's Projects: Secure Photo Sharing & Newsletter Archive**
David, using Claude Code, built two projects.
* **Project 1: Family Photo Sharing Site**
* **Problem:** Inefficiently sharing photos of his young children with family, leading to cumbersome Google Photos albums or camera roll scrolling.
* **Solution:** A secure, straightforward grid of photos and videos.
* **Security:** This was paramount. Claude Code implemented admin/user passwords, revocable "magic links" for individual family members (like Google Docs sharing), access logs, and Cloudflare integration.
* **Key Learning:** Claude Code *automatically stripped EXIF data* from uploaded photos as a security measure without being asked. This initially prevented sorting photos by "date taken." David had to re-engineer the system to first extract the EXIF date, then strip the rest of the metadata.
* **Functionality:** Supports drag-and-drop uploads, an Apple Shortcut for quick sharing from an iPhone camera roll, sorting by date added or date taken, video uploads (with imperfect thumbnails), and a "download all" button for easy backups.
* **Assessment:** David found this project perfectly solved his problem and was functional enough that he considered selling it.
* **Project 2 (Bonus): "Installer Archive"**
* **Problem:** No searchable, sortable archive of the numerous links featured in his weekly "Installer" newsletter.
* **Solution:** A website archiving 5,997 links from 136 newsletter issues.
* **Functionality:** Imperfectly sorted and categorized list with tags, a search function (e.g., for "Obsidian" or "movies"), and links to external databases like TMDB. All categorization and linking were automatic.
* **Challenges:** The project consumed "400,000 billion tokens" over a week, requiring extensive iteration and token usage due to the AI's need to understand context and match links to external data.
* **Assessment:** Despite its imperfections, David found it "incredibly useful" for his personal needs.
**Overall Takeaways**
All three hosts concluded that AI coding tools are exceptionally good for building *personal* tools that solve individual problems. However, they highlighted a "vast" and "unbelievably daunting" gap between such a functional personal prototype and a polished, secure, user-friendly, and scalable commercial product. This gap requires significant additional skills in clean code, security, user experience, and scaling that AI currently cannot autonomously provide. They also noted the AI's tendency to implement illogical guardrails or refuse basic tasks, often leading to user frustration and a shift towards more direct, demanding interaction with the chatbots.