Every project links to its Devpost submission (full write-up, images, demo video), its repo, and its live demo wherever one exists. 1st and 2nd place are verified from official sources; the rest are badge-verified winners whose exact slot was never published.
A starter template for gesture-controlled apps: MediaPipe hand, face and head tracking behind a clean Python layer, 17 ready gestures, one setup script and you are building. Two demo apps prove the skeleton: Math-O-Lantern, a pinch-to-slice math game, and Holo-Board, a tool for drawing in mid-air over your webcam and exporting one clip. Kiro's own framing of why it won: solving a specific pain point, the difficulty of vibe coding complex computer vision logic without hallucinations.


A hook is a trigger plus a runbook. It fires on new gesture files and walks the agent through seven numbered integration steps, ending with a demand to confirm what gesture was detected and list the files modified. That report-back is the accountability behind the claim that seventeen gestures were integrated in seconds.
Rules numbered like law: assets first, one test file as the final step, check gesture attribute names before using them, use Context7, respect the template structure, keep detection and game logic separate. Plus a hard rule that application code lives in its own folder and never modifies the core library, and an eight-step workflow ending in documentation.
Glossary-first: eighteen exact terms defined before any user story, with when-then-shall acceptance criteria. Every task traces to requirement numbers, and one task is a property-based test with named properties: the bank contains only valid operations, and operands stay within digit limits. The champions were doing property-based testing in 2025.
Context7 with auto-approve for documentation lookups, credited in their write-up with removing hallucinations on MediaPipe landmark indices. One well-chosen server with auto-approve beats five half-used ones.
A design-to-code platform that replaces six tools: an infinite canvas with drawing tools, wireframe-to-UI generation via AI agents, chat-based creation across four frontier models, vision references from any screenshot, a visual click-to-style editor that writes Tailwind back to source, a code explorer, and Remix the Web — a Chrome extension that captures any element from any live site and reconstructs it as an editable component. Each screen gets its own live Next.js sandbox embedded in the canvas.
A unified AI inbox for Gmail, Slack and Calendar. Every message runs a parallel LangGraph agent pipeline: summarization, intent classification, 0-to-10 priority scoring, action-item extraction, spam detection. Smart replies are drafted in your own voice using persona memory from your sent mail, with human approval before anything sends. Query it naturally — urgent emails this week, meetings tomorrow — on a live WebSocket feed.
An AI-powered migration tool for teams stuck on legacy PHP. Pitch: stuck with legacy PHP code holding your team back? Meet Frankenstein Laboratory. The direct project page resisted our extractors; the gallery card opens fine from a normal browser.
GeoCities reborn, but every page you build is pinned to a real place on an interactive 3D Mapbox globe. A drag-and-drop builder with ten-plus components, AI-generated pages and layouts, embeddable AI chatbots per page, shareable URLs, a My Sites dashboard, and a spooky mode with flying ghosts. The same democratized web publishing that made GeoCities magical — now that corner has coordinates.
A real-time voice debate coach: bidirectional audio streaming with the Gemini Live API, an adaptive AI opponent, instant post-debate feedback and progress tracking. Two modes, Coach and Fierce, a WebGL voice-reactive visualizer, and an analytics suite — confidence trends, duration distributions, activity heat maps — plus a persona and archetype system.
A Pomodoro-RPG Chrome extension: focus sessions damage Stubborn Souls — bosses with backstories about procrastination and unfinished work, each ending in a heartwarming resolution conversation when defeated. The dopamine-detox twist: game UI is hidden while you work and unlocks only during breaks. Work, then reward. Soft-discouraged versus strict-mode hard-blocked sites, ten bosses, three stats, cosmetics, idle collection, task management, streak stats.
dyc19991225 · Resurrection — a real IRC server (RFC 1459), web client and AI Ghost bot. Honesty-forward Kiro log.
mangeshraut712 · resurrect abandoned repos and dead pages via Wayback, with hash-chained receipts.
toriyama237 · Frankenstein — byte-for-byte failure replay across COBOL, FTP, Postgres and Rust, with a four-agent Kiro Crew companion.
Nobody visible competes on art or design strength, and only one extension is in the field. The Costume student lane stays open.
The pattern: every single winner paired specs with property-based tests and stated quantities — stories, criteria, properties, tasks. The 2026 judging criteria ask about property-based testing explicitly, and the field was already answering in 2025.
The Devpost winners tab is Cloudflare-walled with zero archive captures, but direct submission-page URLs and the paginated project gallery extract cleanly. Gallery Winner badges — seven in total, all on the first three pages — reconstructed the board. 1st place is verified via the repo and Kiro's official spotlight; 2nd place is verified on UnitSet's own Devpost page. Strangler's direct page resisted extraction, so it links to its gallery card, which may open fine in your browser.
Confidence labels: verified means an official source names the placement. Badge-verified means the gallery badge is present but the slot was never published. Strong match means repo evidence matches the gallery text, as with DebateMaster. The full source log lives in the notes alongside this site.