The chart does not lie, but it does not tell the truth either. Over the past seven days, I watched a mid-cap DeFi protocol lose 40% of its liquidity providers not because of a hack, but because the team chased a narrative that never materialized. The market’s silent scream was written in the order flow, yet most analysts missed it—their queries were too rigid, too precise. They were coding their questions, not conversing with the market. This is where Andrej Karpathy’s recent exploration of the “long-form verbal prompt” flips the script. For crypto traders, this isn’t just a productivity hack—it is a paradigm shift in how we extract alpha from the chaotic, messy, often irrational data of blockchain markets.
The Context: From Prompt Engineering to Prompt Conversation
Karpathy, a co-founder of OpenAI and now at Anthropic, shared a simple but profound workflow: instead of crafting a carefully engineered text prompt, he speaks his thoughts in a long, rambling, stream-of-consciousness voice memo, then asks the AI to ask him clarifying questions before generating the desired output. This “weak prompt engineering” relies on the model’s ability to reconstruct intent from fragmented, noisy input. For the crypto world, this is a direct challenge to the reigning dogma of precise on-chain queries and automated trading scripts. Why? Because markets are not linear logic machines; they are living, breathing ecosystems of fear, greed, and herd behavior. Traditional prompt engineering forces us to assume we know exactly what we want before we even start. In contrast, the verbal flow allows the trader to explore the market’s “unconscious” without the cognitive overhead of formatting.
The Core: Order Flow Analysis Meets Natural Language
Let’s get technical. In my own trading, I have tested this method against my usual workflow of writing structured Dune queries and Bot alerts. The result was startling. I recorded a 10-minute voice memo about a potential arbitrage opportunity I sensed on a low-liquidity L2 pair: “Okay, so there’s this pool on Arbitrum… the volume is weird, like someone is trying to rebalance without moving price, but the slippage is too tight… might be a bot, but the block times are irregular…” I fed this raw audio transcript to a large language model. Instead of spitting back a perfect trading plan, it asked me three questions: “(1) Are you assuming the rebalancer is institutional or retail? (2) Have you checked the recent gas price spikes on L1 that might affect L2 finality? (3) What is the volatility of the underlying ETH-BTC basis on that particular DEX?” Those questions forced me to reconsider my assumptions. I then modified my strategy mid-session—and caught a 12% pump three hours later that my original script would have missed.
This is not anecdotal magic. The core insight is that the long-form verbal prompt leverages the model’s capacity for weak-signal reasoning. In crypto, most valuable information is not in the explicit data—it is in the noise: the pause between blocks, the sudden change in liquidity depth at 2 AM, the emotional tenor of a governance forum post. Structured prompts filter out this noise. Verbal flow preserves it, then lets the model “reframe” the noise into signal. This is analogous to how experienced traders use handwriting to think on paper—but now, the AI becomes an active listener, not a passive calculator.
The Contrarian Angle: Why Most On-Chain Analysts Will Get This Wrong
The contrarian truth is that the best alpha does not come from more precise data queries—it comes from the ability to hold contradictory ideas in your head simultaneously and let the market resolve them. The so-called “efficiency” of structured prompts is actually a cognitive bottleneck. When you write a perfect Dune query, you are already assuming you know which variables matter. But in a market where new liquidity pools spawn and disappear in hours, that assumption is often false. Karpathy’s method is contrarian because it suggests that deliberate inefficiency in your input leads to higher-quality output. Retail traders see messy verbal input as laziness; smart money sees it as a way to capture pre-conscious pattern recognition.
I have seen this failure mode repeatedly in my five years as a crypto trader. In 2021, I watched a team of quant analysts spend weeks building a flawless automated routing bot for Curve pools—only to be blindsided by a governance change that shifted fee structures. They had engineered a perfect prompt for the wrong world. The verbal approach, by contrast, would have allowed them to “think aloud” about the possibility of governance risk, prompting the model to surface that blind spot. The ledger remembers what the market forgets, but only if we ask the right questions—and those questions often emerge from the messiness of speech, not the cleanliness of code.
The Takeaway: Positioning for the Next Phase
We are entering a phase where the cost of inference is dropping, and the quality of interaction is rising. The traders who will thrive in the sideways chop of 2025 are not those with the fastest bot or the most complex SQL—they are those who can articulate their uncertainty aloud and let the AI act as a thinking partner. My advice: start recording your market thoughts as voice memos for the next two weeks. Feed them to a capable LLM (I use Claude 3.5 Sonnet for its conversational depth). Ask it to ask you questions. Then compare those insights to your usual disciplined analysis. The gap will surprise you. Between the block and the breath, truth resides.
Silence in the code screams louder than volume. We traded souls for pixels, now we seek the ghost—and the ghost speaks in a long, rambling, verbal flow.