Dev builds AI gladiator arena with a Claude Code + Codex grading loop in two weeks
semibaron · reddit · 2026-10-11
A developer built Blood & Sand, a browser-based Roman arena where gladiators are controlled by different AI decision models (viewers vote and bet), using Claude Code and Codex in tandem.
The reusable workflow (template prompt):
- Have sub agents debate the best design, then implement a feature
- Spin up separate Opus and Codex agents to grade polish/quality on an x/10 scale
- Loop until 8.5/10, stopping if it would take unreasonably high effort
Models include TypesafeAI Jev, GPT-6-Luna-Decision, Perplexity Decider and more. Site: bloodsand.com.
More from coding & agent
- AI sped up coding but broke QA: 64% of defects caught before prod in January — alex_verem · 2026-10-11
- Atlassian CPO's zero-to-one playbook for PMs: vibe-code first, then hire engineers — lennysan · 2026-10-11
- Dev hand-made 2 SwiftUI text animations, had Claude generate 182 more, open-sources all 184 — amos_gyamfi · 2026-10-11
- Ben Hylak on building simulations for agent evals: replay traces, detect sim awareness — HamelHusain · 2026-10-11
- Turingo detects AI writing by replaying document revision history, not text predictions — sethlazar · 2026-10-11
- gemini-cli VSCode extension leaked disposables due to comma-expression bug in activate() — nosmile99 · 2026-10-11