Vibe coding examples: real projects, including the ones that died

Most "vibe coding examples" online are demos: a to-do app, a landing page, a game built in an afternoon. Useful, but they skip the part that tells you the most, which is what happens after the demo.

So here's a real portfolio. These are my projects, all built almost entirely with AI coding agents (mostly Claude Code). Some are live. Most are still being built. Several changed direction or stopped, and those are the ones I've learned the most from.

Live

Vibe Leagues — the site you're reading. A home for vibe coders to share build-in-public stories, with invite-only accounts, positive-only reactions, a posting API and these guides. It runs on Cloudflare with a real database, and it was built by an agent working to a written list of rules it isn't allowed to break.

Heart & Industry — heartandindustry.com, a catalogue and comparison engine for British-made goods: "what is the British alternative to this?" The holding site is live; the full catalogue isn't finished yet. Even the product data is extracted by Claude Code agents rather than a pay-per-use AI API.

Still being built

Threadbare — a co-op survival game set in a geographically accurate, post-apocalyptic Sheffield, in Unity. It's the purest vibe coding I do: Claude Code in VS Code, connected directly to the Unity Editor. It's also early. The groundwork is done; the terrain isn't.

Storying — an iPad storybook for reading aloud. A parent reads, the app listens through the microphone, and the words light up as they're spoken. The first three phases are complete; the next needs a backend for a marketplace of story packs.

Ragtag — a desktop app that lets non-technical people build private AI personas grounded in their own folders, websites or search topics, running locally.

Lots to Unpack — turn a single photograph into a 3D scene you can move around in, then shoot new camera angles from it. Built for AI filmmakers and storyboard artists.

Auto-Cue, Auto-Flick, Auto-Shout, Auto-Launch — a suite of local-first tools for solo founders: social posting, video making, finding commentators, and launch planning. Auto-Shout is feature-complete; none of them is publicly released yet. Releasing is its own project, with signing, licensing and payments, and it's the part that's easy to put off.

Changed direction, shelved or retired

This is the section most portfolios leave out.

seo-data: retired, and that's a success. It was a small command-line tool that gave every one of my sites, and the AI agents working on them, a single, budget-guarded way to read search data. On 28 September it was retired and archived, replaced by Auto-Spark, a bigger tool that took over its job. It did what it was for, then handed over.

Auto-Press: the generator I deleted. Auto-Press began as a tool for creating websites, and seven phases of it were built and working. Then the real use became clear: people ask an AI to build them a good website, and what they actually need afterwards is a way to edit it without touching code. So it pivoted to editing existing sites, and the original generator was removed entirely rather than left in as a dormant second mode, "because every phase afterwards had to reason around it." Deleting working code you're proud of is one of the hardest calls in a project. Keeping it would have been worse.

Lots to Unpack: no price tag. I'd planned to sell it. Then I withdrew the pricing altogether, because, in the words of the decision record, "there is no value in a commercial licence if nobody would buy the thing." It's still being built. It just isn't pretending to be a business before it's earned that.

Pattern Engine: stopped on purpose. A deterministic engine that turns motifs into print-ready seamless patterns for textile printing. Its specification told me to stop and reassess once the core milestones were done, and I did. Screen printing is shelved. The engine now serves a new idea, Dreamwoven (bespoke children's pyjama prints from a child's interests), which is still a concept.

Auto-Flick: rejected by its payment provider. Not every stop is your own decision. Auto-Flick's payment provider turned it down because it's an "AI video" product, so the payments plan had to change before launch.

Auto-Cue: one platform dropped. Auto-Cue posts to Bluesky, Mastodon, Threads, Facebook, LinkedIn and Instagram. It was going to support Nostr too, and that's being dropped. Not every integration is worth maintaining.

What the portfolio taught me

Write down why you stop. Every project above that changed direction has the decision written down: the date, the reason, what was removed. That's what made this article possible. More usefully, when a future agent session reopens one of these projects, it knows not to rebuild the thing I deliberately killed.

Documentation is not progress. My readiness tracker turned up an uncomfortable pattern: documentation volume correlates with projects that went nowhere, and the project with the largest agent instruction file has a single commit. It's easy, with an AI, to plan a project in loving detail and never build it.

Retiring something can be the win. seo-data didn't fail. It did its job and handed over to its successor. Calling that "dead" misses the point.

Shipping is its own project. Building is what AI made fast. Signing, licensing, payments and launching are still slow, and they're where most of my "nearly finished" projects are waiting.

---

Vibe Leagues has a stage for exactly this: projects move from idea to building to launched to growing, and sunset is a celebrated stage, not a hidden one. If you've built something, or stopped building something, that's a story worth telling. Request an invite.