CZ is not the only one wondering how @the_nof1 really works - plenty of detailed performance and trade analyses (h/t @jinglingcookies @jay_azhang) but haven't seen anyone looking at the actual prompts yet. here's what the models get: > asset universe: BTC, ETH, SOL, BNB, DOGE, XRP > instruments: Hyperliquid perps only with leverage 5-40x > every 2min: timestamp + portfolio + coin data >> portfolio: session tracking (invocations, time), positions and performance (entry/exit prices, leverage, returns, sharpe ratios, cash left) >> coin data: price, OI, volume, funding rates, EMA, MACD, RSI, ATR > system guardrails >> G0: allowed model actions are 'buy_to_enter', 'sell_to_enter', 'hold', 'close_position' >> G1: "No pyramiding or size increases" cannot add to existing positions, only enter new ones >> G2: has to output JSON for each position with quantity, profit_target, stop_loss, invalidation_condition, leverage, confidence, risk_usd (see image) >> G3: entry, exit and close actions need extra "justification" field in the JSON with reasoning overall it's a well-scoped experiment but the decision space for the models is relatively narrow for now. looking forward to see where they take it from here!