Anthropic's hackathon recap shows where Claude Code is actually landing
Anthropic published the winners of its Built with Opus 4.7 Claude Code hackathon. The projects — medical training sims, electronics repair, CS pedagogy, puppet theater, and Chilean home-repair triage — are a useful read on what individual builders are doing with agentic coding tools right now.
Anthropic has posted the winners of its Built with Opus 4.7 Claude Code hackathon, alongside a note that a follow-up event, Claude Build Day, has already happened in San Francisco using Opus 4.8. The post is part marketing, part field report. The field report is the interesting half.
Five projects, all built by solo developers, all shipped in roughly a week:
- Medkit (Bedirhan Keskin, Istanbul) — a simulated-patient trainer for medical residents, with an agentic grader scoring encounters against published clinical guidelines. Three Istanbul medical faculties and a pharma company are reportedly lined up for pilots.
- Wrench Board (Alexis Chapellier, France) — an electronics-repair assistant that ingests schematics and boardviews, builds a unified electrical graph, and walks a technician to the exact pad to probe.
- Maieutic (Paula Vásquez-Henríquez, Concepción, Chile) — a teaching IDE that deliberately slows students down: spec before code, no autocomplete, Socratic chat, and an "Intent-Diff Review" that compares spec to final code. Researchers at the University of Houston have reportedly reached out about co-authoring a paper.
- Virtual Puppet Theater (Rene Hangstrup Møller) — a browser puppet show driven by webcam tracking, Three.js, and Opus 4.7 for spatial reasoning and improvised dialogue. Open source, MIT.
- MaestrIA (Benjamin Torralbo, Chile) — a home-repair triage app aimed at Chilean tradespeople, with photo-based diagnosis, severity scoring, budget estimates, and a WhatsApp handoff to nearby maestros.
What the workflows tell us
The recurring pattern across the writeups is worth more attention than any single demo.
Most of the winners describe a spec-first, multi-agent loop rather than vibes-based prompting. Keskin ran four parallel Claude Code sessions, one per subsystem, working largely by voice. Chapellier prototyped in Claude Design, then ran "five or six agents in parallel" in Claude Code's multi-agent mode, one per domain, using a skills framework called Superpowers to structure brainstorm-then-plan. Vásquez-Henríquez spent two of her seven days writing specs before touching code. Møller used Claude across concept, planning, and code while keeping architecture and review for himself.
This is closer to how a small engineering team actually operates than to the autocomplete-and-pray story that dominates a lot of the agentic-coding discourse. The tooling — parallel sessions, design-to-code handoff, skills frameworks, voice input — is doing real work, but so is the discipline of writing things down before generating them.
What we'd flag
This is Anthropic's own hackathon, written up on Anthropic's blog, so the selection is curated and the framing is favorable. The projects are prototypes; "set to start running pilots" and "researchers reached out" are not deployments. None of the cost, latency, or failure-rate detail you'd want to evaluate these as products is in the post.
Still, the throughline is consistent with what we're hearing elsewhere: the people closest to a domain — a physician, a repair technician, a CS lecturer, a carpenter's son — are increasingly able to ship working software about that domain in days. Whether that translates into durable products or stays at the demo stage is the question the next recap should answer.
