Architecture explainer for a 24/7 trading system

CLAUDE OPUS 5.5 IS INSANE FOR BUILDING A 24/7 TRADING SYSTEM I mapped the full build from an empty repo to a service that can run, recover, and explain its decisions. The 10-part breakdown: 1. Define the job → Opus writes the spec and code. Your trading service runs the approved rules. 2. Start with one strategy → BTC-USD spot, completed 5-minute candles, a 20-bar breakout entry, and a 10-bar exit. 3. Give every role an output → Research produces a spec. Engineering produces tested code. Review finds defects. Operations prepares recovery. 4. Make market data replayable → Store events, detect gaps, and pause new entries when the feed cannot be trusted. 5. Record every decision → Entry, exit, or hold, with the candle, threshold, position state, and reason attached. 6. Put risk before orders → Check cash, exposure, open intents, and data health before an entry reaches execution. 7. Handle the dangerous “unknown” → If an order times out, reconcile it before sending anything again. 8. Test two separate things → Does the strategy survive costs? Does the software survive disconnects, duplicate fills, and restarts? 9. Run in paper mode → Keep the service on continuously and inspect signals, orders, fills, and alerts. 10. Let Opus review the night → It can investigate logs and propose changes. Tests and your approval decide what ships. The interesting part isn’t getting AI to say “buy.” It’s building a system that still knows what happened when an order response disappears. The video shows the architecture. The complete build, prompts, and tests are in my article below ↓

— Solomon Rojie (@Solomonrojie) Sep 28, 2026