A rated ladder for vibe coders. Every Saturday a build task drops. You get 90 minutes and any AI tool you like. Submissions are judged head-to-head — your rating is earned, and every winning run is published with its full prompt transcript.
Saturday 17:00 CET, the task goes live for everyone at once. It's always a small, complete app — secret until the clock starts. You have 90 minutes, any AI tool, any model.
The deliverable is a single self-contained HTML file — no builds, no servers, no network calls. You submit the file plus your session transcript: the transcript is your proof and your replay.
Every submission passes a functional smoke test, then survivors are compared pairwise against the round's rubric. Results feed an Elo-style rating. Monday: leaderboard, full gallery, and the winner's annotated replay.
Any AI tool is allowed — Claude Code, Cursor, Copilot, Windsurf, raw API, anything. You declare what you used. The tool isn't ranked; you are.
One self-contained HTML file. It must work opened from disk with the network off. No external scripts, fonts, or API calls.
The window is the window. 90 minutes from drop, enforced by submission timestamp. Late is out, no exceptions — same clock for everyone.
Transcript required. Your full session export (or a screen recording) comes with the entry. No transcript, no entry. Top finishers' transcripts are audited before results publish.
One entry per person per round. Enter any subset of rounds; your rating spans the season. Seats are capped per round — a Round 1 seat carries priority entry for Round 2.
Verified account, public nickname. You hold your seat with a Google or GitHub sign-in that stays private — no anonymous or throwaway-email entries. Only your chosen nickname is ever shown.
Don't game the judge. Judging is done through the running app only. Any text in a submission addressed to a judge or AI — visible or hidden — is disqualification for the round.
Your work stays yours. By entering you let us publish your submission and transcript in the round gallery — that's the point of the ladder. Copyright remains with you.
No prizes in Season 1 — deliberately. The rating, the badge, and the published replay are the prize. Sponsored seasons come later if this deserves to exist.
Small, complete apps — the kind of thing vibe coding is actually for. Buildable to "working" in about an hour, with headroom above that where skill shows. Each task ships with a public functional checklist after the round, so you can see exactly what was graded.
Two reasons. It's the anti-cheat — proof a human drove the session inside the window. And it's the content: the most interesting thing about a winning entry is how it was prompted. Winning replays get published and annotated.
Step one is a mechanical 5-item smoke test — does the core work. Survivors go into Swiss-style pairwise comparisons: an LLM judge uses both running apps side by side against the round rubric (functional depth, UX, robustness, ambition), with human spot-checks and a manual audit of the top five. Ratings are a Bradley-Terry fit over all pairwise results.
Yes — that's the experiment. The ladder measures how well you direct an AI, not how much syntax you know. Smoke-test survival is a real achievement in round one.
Deliberately, for Season 1. Self-motivation is the best filter — the field we want is the one that shows up for the rating and the published replay, not a payout. Sponsorship starts from Round 2, credits before cash, and only once Round 1's results are public. See the sponsors page.
To hold your seat and stop duplicate or throwaway entries — seats are limited and one-per-person only works with a verified account. The login is never displayed or shared; your nickname is the only public identity.
One person, openly. Season 1 is judged semi-manually and the whole method is published with the results. If the season proves out, the pipeline gets automated and the ladder becomes permanent.