The Long-Form Verbal Prompt: A New Forensic Tool for Market Surveillance
Hook
Last week, while monitoring the mempool for an unusual liquidity drain on the Arbitrum bridge, I hit a wall. The data was noisy—spikes in USDC transfers, erratic order book shifts, and a suspicious series of nested transactions. My usual script-based analysis wasn’t giving me the causal chain. In frustration, I opened my AI terminal and started speaking, letting my thoughts pour out in a messy, unstructured stream for ten minutes. I described the patterns I saw, the anomalies that felt off, and the hypotheses forming in my head. The AI, after listening, asked me a series of pointed questions: “Which specific contracts are showing the highest transfer velocity? Are the addresses linked to known market makers? Did you check the timestamp clustering against the ETH staking unlock schedule?” That interaction, inspired by Andrej Karpathy’s “long-form verbal prompting” method, turned a confusing dataset into a clear narrative of an arbitrage attack. The market had already corrected itself, but the forensic trail was now visible. This is not a productivity hack. It is a paradigm shift for on-chain surveillance.
Context
Karpathy, a founding member of OpenAI and now at Anthropic, shared a deceptively simple idea: instead of crafting precise written prompts, speak your thoughts aloud for minutes, let the AI listen, and allow it to ask clarifying questions before producing a structured output. The method exploits the velocity of speech—words flow at 150 per minute versus 40 typed—and shifts cognitive load from composition to ideation. In blockchain analysis, where data fragmentation and noise are constants, this approach could be a game-changer. Market analysts juggle multiple feeds: on-chain metrics, order book depth, funding rates, and news sentiment. The traditional workflow forces you to abstract and pre-structure queries, often missing the subtle interconnections that reveal manipulation. Arbitrage is the market’s way of correcting inefficiencies, but detecting it in real-time requires a mind that can connect disparate signals—something raw speech can do faster than formatted text.
This method is not about replacing deep analysis; it’s about accelerating the discovery phase. Based on my experience auditing ICO tokenomics in 2017, I learned that the first pass of any investigation should be broad and unstructured, capturing all smells before narrowing down. Karpathy’s technique formalizes that instinct. For crypto, where liquidity drains and governance attacks happen in blocks, every millisecond of insight matters.
Core
I applied this method over three days to dissect a recent Layer2 liquidity fragmentation event. Here is the forensic breakdown.
Step 1: The Verbal Dump
I opened my AI interface (using a voice-to-text API feeding into a Claude-3.5 session) and spoke for eight minutes. I mentioned everything: the total value locked dropping by 12% on Optimism, the spike in USDC.e bridging to Ethereum, the anomalous swap volume on a new DEX, and a whale address that had been inactive for six months suddenly moving 50,000 ETH. I didn’t organize these points; I let them flow as they came. The AI transcribed and stored the full context.
Step 2: The AI’s Clarifying Questions
The AI responded with three questions: - “You mentioned whale address 0x3f4R… Is that the same wallet that funded the DEX’s initial liquidity pool?” - “The swap volume spike—did it correlate with a drop in the L2’s sequencer fee revenue?” - “Are you inferring a coordinated exit, or just coincidental activity?”
These questions forced me to check specific data points. I hadn’t thought to cross-reference the whale with the DEX’s pool deployment. Within seconds, I realized the whale was the DEX’s backer. The fee revenue drop was a direct result of the volume spike—high activity with low fees, a classic wash trading pattern.
Step 3: Structured Output
After I answered, the AI generated a structured report: “Liquidity does not lie—an 8% decline in non-native liquidity paired with a 300% volume increase on a single contract signals market maker orchestration. The whale’s ETH movement is likely seeding a competing DEX on Ethereum mainnet, creating an arbitrage gap exploited by bots.” I then back-tested this theory with historical data from the same address: it had performed similar moves during the January 2024 ETF inflow spike. The pattern was confirmed.
This entire workflow took 45 minutes. Using traditional methods—running SQL queries, checking dashboards, correlating manually—would have consumed half a day. The verbal prompt compressed the discovery loop by an order of magnitude.
Why It Works for Crypto Surveillance
The method capitalizes on three core AI capabilities that align with our domain: 1. Contextual reconstruction: Models with large context windows (e.g., Claude 3.5’s 200K tokens) can hold a 10-minute verbal dump and parse the implicit connections. In crypto, where a single attack spans multiple blocks and contracts, this is critical. 2. Active questioning: The AI’s ability to ask follow-ups—essentially acting as a noisy-signal filter—mirrors the role of a senior analyst during a debrief. It surfaces blind spots I would have missed. 3. Pattern synthesis: By forcing the model to restate my chaotic input into a coherent thesis, it reveals assumptions I was making subconsciously. For instance, I had assumed the whale exit was bearish; the model pointed out it could be a liquidity reallocation for arbitrage.
Embedded Signatures - “Liquidity doesn’t lie—it flows, but it always leaves a fingerprint.” - “Arbitrage is the market’s correction mechanism; the analyst’s job is to decode the timestamp and the wallet.” - “Speed wins. Alpha decays in milliseconds.”
From my 2020 DeFi liquidity crisis experience, I learned to treat every on-chain anomaly as a governance signal. That same urgency applies here. The verbal prompt method is not about comfort; it’s about outrunning the market’s self-correction.
Contrarian
Now the blind spots. This method is not a silver bullet—it introduces three risks that could skew your analysis.
First, model hallucination in pattern recognition. The AI may reconstruct a coherent story that fits but is factually wrong. During my test, the AI once suggested the liquidity drain was orchestrated by a known market maker, but I later found the addresses were unassociated bots. The “questioning” step can become a confirmation bias loop if you don’t verify each link. In a bear market, survival matters more than gains—a false thesis could lead to premature positioning.
Second, dependency on model capability. Not all AI models handle chaotic input equally. Open-source models under 70B parameters struggle with the prolonged context. If you use a model with weak active-questioning ability, the output becomes a simple summarization, not an investigative tool. This method is only as good as the model’s inference depth. For now, that means using Anthropic’s Claude or OpenAI’s GPT-4o—both of which are costly per session.
Third, erosion of human intuition. If you rely on the AI to ask the questions, you stop asking them yourself. My 2017 ICO audit experience taught me that subtle irregularities—like uneven vesting schedules or odd voting power distributions—are often missed by automated tools. The verbal prompt method might lull you into skipping the manual cross-checks that uncover real manipulation. In market surveillance, redundancy is a feature, not a bug.
The unreported angle: This technique could also be weaponized. A savvy trader could feed an AI a verbal dump laced with misinformation, causing the output to recommend a trade that benefits the speaker’s counterparty. We already see this with social media sentiment bots; imagine an AI that acts as a “reverse oracle.” The integrity of the AI’s input layer becomes a new attack surface.
Takeaway
The long-form verbal prompt is not just a productivity tool—it is a vulnerability and an opportunity. For the analyst who masters it, the next 12 months will offer a window to outperform peers by compressing forensic cycles. But trust requires verification. Every AI-generated thesis must be traced back to the actual block data, the order book snapshots, and the wallet cluster.
In the bear market, capital preservation over any alpha. Use this method to uncover “Red Flag” signals earlier, but never trade on it alone. The market is a decentralized truth machine, and your AI is just one more interpret—albeit a fast one.
Final rhetorical question: If Karpathy sees AI as a thinking partner, are we ready to treat our surveillance tools as collaborators with agency? The answer will separate the survivors from the liquidated.