Pencils down: We share our vibe-coded websites

Episode Summary

Executive Summary: The episode revisits Vergecast's month-long vibe-coding challenge, where David Pierce and two colleagues built personal websites using AI tools. The conversation compares Codex, Claude Code, and Gemini, highlighting both the usefulness and frustrating limitations of vibe coding for small, personal projects. The hosts conclude that these tools are great for prototypes and self-use, but far from making anyone a production-ready software engineer.

Main Topics: Revisiting the vibe-coding challenge (Priority: 5/5): David checks in on the experiment where each host used AI coding tools to build a website solving a real-life problem, framing it as a test of what regular people can create with vibe coding. Tool comparison: Codex, Claude Code, Gemini (Priority: 5/5): The hosts compare their experiences across different AI coding platforms, focusing on reliability, permissions, design quality, and how much handholding each required. Hayden's portfolio site and public publishing (Priority: 4/5): Hayden built and published a revamped personal portfolio site with 90s-inspired styling, embedded media, and manually gathered links, despite tool frustrations. Jake's farm-share exchange site (Priority: 4/5): Jake built a private-ish site for sharing excess CSA produce with friends, but ran into hosting and permissions issues that prevented full public deployment. David's family photo-sharing app (Priority: 5/5): David created a secure family photo timeline with revocable links, EXIF-based sorting, and mobile shortcuts, solving a practical personal workflow problem. Installer Archive as a personal knowledge system (Priority: 4/5): David also showed a second, more ambitious project: a searchable archive of every link featured in his newsletter Installer, demonstrating the power of AI for personal information management. Limits of vibe coding for real products (Priority: 5/5): All three agree that AI tools are strong for personal prototypes but the leap to secure, polished, scalable products for others remains large and requires human judgment and software engineering skill.

Key Arguments: AI coding tools are most effective for building small, personal utilities that solve specific individual problems. Different models/platforms have different strengths and weird failure modes: design generation, permissions handling, and access to hosted services can all break down. A lot of the work in vibe coding is not code generation but persistent prompting, troubleshooting, and repeated clarification. Publishing a website for public use adds a large gap in security, UX, reliability, and scalability that AI tools do not close automatically. The tools can still be highly productive: the hosts successfully built functional systems including a portfolio site, a produce-sharing board, a family photo app, and a searchable archive. Good software for the public requires more than a prototype; it involves maintaining structure, consistency, and product thinking beyond what vibe coding currently handles well.

Data Points: Challenge length: About 1 month - The hosts spent roughly a month building their vibe-coded projects before reconvening. Jake's Codex threads: 7 separate threads - Jake said he ended up with seven Codex chat threads over several days because the tool repeatedly lost access or stalled. Hayden's Gemini conversation: About 9 hours - Hayden described a very long single Gemini session to complete her portfolio site. David's first prompt length: ~1500 words - David said his first prompt for his family photo app was about 1,500 words long to constrain the model and avoid mistakes. Installer Archive size: 5,997 links - David's archive project indexed nearly 6,000 links from his newsletter Installer. Installer issues ingested: 136 issues - David said he fed the system 136 issues of Installer. Obsidian mentions in Installer: 18 times - The archive found Obsidian mentioned 18 different times across the processed newsletters. Reflect Orbital proposed satellites: 50,000 - In the opening news segment, Reflect Orbital's long-term vision was described as deploying 50,000 mirror satellites. FCC approval: 1 mirror satellite approved - Reflect Orbital recently won FCC approval for a single test satellite meant to reflect sunlight. Zuckerberg manifesto length: 6,500 words - The news segment mentioned Mark Zuckerberg publishing a 6,500-word AI manifesto. Vanta customer count: 16,000+ companies - A sponsor read cited Vanta as being used by over 16,000 companies. Blinds.com review count: 50,000+ five-star reviews - Another sponsor read claimed Blinds.com has more than 50,000 five-star reviews.

Pivotal Quotes: "What you're supposed to do is actually have a folder full of instructions, and you're supposed to have a different chat for every single feature you want to build." — David Pierce: David reflects on the prevailing advice for using AI coding tools, contrasting it with his own more conversational approach. "This is like a problem that even Google Photos breaks down with." — David Pierce: David praises his family-photo app by comparing it to a consumer product that struggles with the same sorting problem. "The gap between this thing that I made for myself and like good, published, user-facing business product is so vast that I can't even comprehend how to get from here to there." — David Pierce: The hosts conclude that prototype-building with AI is very different from shipping robust software for large audiences.

Implications: Vibe coding is proving useful for personal tools and fast prototyping, but not yet a replacement for software engineering. For listeners, it suggests AI can help build real things quickly—if the goal is narrow, private, and forgiving.