VibeLadder Round 2: Fourteen Minutes, Three Prompts, and a Tied Season

Summary

VibeLadder’s second round asked two entrants to build Connect Four against a computer opponent in one HTML file. The secret task required a seven-by-six board, at least two genuinely different difficulty levels, win and draw detection, a new-game control, and a harder mode that actively tried to win. Both entries passed all five functional checks, avoided network violations and injection attempts, and were then judged by an LLM playing both applications through their interfaces without seeing the source. The entrant using OpenCode with GPT-5.6 Sol won a close head-to-head against the entrant using Claude and ChatGPT. The judge found both hard-mode engines capable of creating threats and both easy modes observably weaker, but gave the overall edge to the second entry because it fit a 1280-by-800 viewport, kept result messages and the winning four visible, and tracked per-difficulty records and streaks. The first entry won functional depth by adding a who-goes-first selector, keyboard controls, score reset, and an unlockable Perfect tier, but required scrolling and exposed a debug badge whose hidden attribute was overridden by CSS. The season is now level at one win each, so the Bradley-Terry ratings place both entrants at the 1100 baseline; the earlier 1192 rating had rested on only one comparison. The winning session used three prompts and about fourteen minutes: planning, refining storage-backed convenience features, then requesting an initial implementation. The entrant submitted at 15:26, 27 minutes after the task opened, and did not resubmit. The runner-up spent nine planning prompts and a 95-minute build that included a bitboard solver, a large opening book, and extensive tests, but submitted a mid-build snapshot after the window had closed. The report’s lesson is to time-box AI-assisted builds, submit a working version, and iterate within the contest window. Round 3 was scheduled for 26 September at 17:00 CET.