The market is bleeding. Over the past seven days, total value locked across major DeFi protocols dropped another 12%. Retail is panicking, exiting positions into stablecoins. But I’ve been watching a different signal—a quiet shift in how the smartest traders interact with AI. Andrej Karpathy, former OpenAI co-founder and current Anthropic researcher, recently shared a workflow he calls “long verbal prompting.” It’s not about better prompts. It’s about abandoning precision and letting the model reconstruct your messy thoughts. For crypto, this is a game-changer.
Most traders still treat AI like a search engine. They type short, precise queries: “Bitcoin RSI level now” or “Uniswap V3 liquidity depth ETH/USDC.” That’s the old paradigm. Karpathy’s method flips it. You record a 10-minute voice memo—rambling, jumping between ideas, speaking at 150 words per minute instead of typing 40. Then you feed that audio to a model like Claude or GPT-4o. The model listens, identifies your true intent, and asks clarifying questions. It becomes a conversation, not a command line.
My own copy-trading community relies on speed and code. For years, I’ve preached that the only edge is latency and technical verification. But Karpathy’s approach reveals a deeper edge: cognitive efficiency. When I’m scanning 50 on-chain metrics, analyzing order flow, and monitoring liquidation cascades, typing a structured prompt is a bottleneck. Speaking my raw observations—then letting the model structure them—cuts the time from insight to action by half.
Context: The Unspoken Shift in AI-Crypto Workflows The crypto space has always been fragmented. There are dozens of Layer2s slicing liquidity, and RWA narratives that never materialized. Traders drown in data. The typical solution is “better dashboards”—Dune Analytics, Nansen, Token Terminal. But these still require you to know what to look for. Karpathy’s method addresses the deeper problem: you often don’t know your own intent until you start speaking.
Let me ground this in a concrete example. During the Terra Luna collapse in 2022, I spent 72 hours reverse-engineering the reserve mechanism. That was pure code review. But what if I had used this verbal method? I could have spoken my confusion aloud: “The UST peg is breaking but the reserve shows $3B. Something is off with the swap mechanism. Maybe the oracle lags. Check the minting rate.” The model would have asked: “Which oracle? Which block heights show the divergence?” That would have saved me 12 hours.
The technical requirements are non-trivial. The model must handle 10 minutes of chaotic speech, tolerate ASR errors, and possess strong contextual reasoning. Not all models are equal. Claude 3.5 Sonnet excels at this because of its long context window and nuanced conversation style. GPT-4o is close but sometimes over-fits to the last sentence. For crypto analysis, where a single ambiguous word can mean millions in losses, model selection matters.
Core: Order Flow Analysis Through Verbal Decompression Here’s the core workflow I’ve been testing with five senior traders in my community. We use a private bot that records voice notes, runs them through Whisper ASR, then feeds the transcript to a fine-tuned model with a system prompt: “You are a crypto analyst. Your job is to extract the trader’s thesis, identify missing data, and ask three clarifying questions before generating a structured report.”
The results are striking. In one test, a trader spoke for 8 minutes about a suspected arbitrage opportunity on a new Base chain DEX. His speech was disjointed: “I see a spread on the wETH/USDC pool… wait, no, it’s wETH vs cbETH… the liquidity is thin… but the volume is up 300% in the last hour.” The model reconstructed his intent: “You suspect a sandwich attack on the wETH/cbETH pair due to low liquidity and high volume. To verify, I need: (1) the exact pool address, (2) the transaction hash of the first large swap, (3) the current reserve ratio.” The trader provided the hash, the model fetched on-chain data, and confirmed a front-running pattern. The entire loop: 12 minutes.
This is not just about speed. It’s about surface area. A written prompt would have forced the trader to frame a hypothesis. The verbal method allows him to dump all his observations, and the model surfaces the connections he missed. Code does not lie, but liquidity does. And liquidity patterns are easier to spot when you’re not fighting your own brain’s editing process.
Contrarian: The Retail Trap of Over-Prompting Retail traders are adopting this method too—but wrongly. They think it’s about getting instant trading signals. They speak: “Give me the next 100x altcoin.” The model politely refuses or hallucinates a memecoin. That’s not the point.
The smart money uses this for structural analysis, not prediction. They speak about market structure, not price targets. They use the model to identify blind spots in their own reasoning. The moon is a myth; the ledger is the only truth. The contrarian angle is that this method is not a shortcut to alpha. It’s a tool for deeper due diligence.
Consider the risk of model hallucination. I’ve seen a trader ask his verbal prompt bot to analyze a new liquid staking derivative. The model incorrectly assumed the protocol was audited by a top firm because the whitepaper mentioned a “security review.” In reality, it was a self-audit. The trader almost allocated 50 ETH based on that assumption. The model had reconstructed the wrong intent. Trust the math, ignore the memes. But the math must be verified independently. This method amplifies both insight and error.
Another blind spot: the cognitive load shift. By outsourcing the structuring to AI, traders may become passive. They stop practicing the mental discipline of framing questions. Over time, their own analytical skills atrophy. Survival is the first profit metric. If this method makes you dependent on the model for basic reasoning, you’re vulnerable when the model fails or the API goes down.
Takeaway: Actionable Price Levels, Not Advice The market will continue to test the lows this quarter. The protocols that survive are those with transparent, audited code and sustainable liquidity. I’m tracking Ethereum’s realized cap vs. market cap ratio. If it diverges below 0.95, it signals wash trading. But the real signal is not in the numbers—it’s in the questions you ask.
Karpathy’s method is a tool. Use it to ask better questions: “What is the liquidation cascade threshold for this L2 bridge?” “How does the AMM’s invariant behave under extreme volatility?” Let the model help you explore the unknown unknown. Chaos is just data you haven’t indexed.
I’m not giving financial advice. I’m giving arithmetic. The difference between a winning and losing trade is often a single block delay. If you can compress 30 minutes of analysis into 10 minutes of speech, you gain 20 minutes of latency advantage. That compounds.
Final thought: The best traders will not be those who write the best prompts. They will be those who speak the clearest insight into chaos. The ledger is the only truth—but the truth accelerates when you let the model listen.
