Hook
Andrej Karpathy, the AI researcher and former OpenAI co-founder, just dropped a workflow that breaks every rule of prompt engineering. He calls it “long-form verbal prompting.” Instead of writing a clean, structured query, he records a 10-minute voice memo—jumping between ideas, repeating himself, leaving sentences unfinished. Then he feeds the raw transcript to a model like Claude or GPT-4, letting it ask clarifying questions before producing the final output. The result? A messy human thought stream transformed into structured, actionable analysis in minutes. For those of us who live by the ticker tape, this method is worth studying—not because it’s a new trading bot, but because it reveals how the next wave of AI tools will change the way we consume and react to crypto news. Speed beats analysis when the graph is vertical, and verbal prompts might be the fastest way to get from chaos to conviction.
Context
Karpathy built this method while co-founding OpenAI and later working at Anthropic. His core insight: the bottleneck in human-AI interaction is input friction. Typing a perfect prompt takes cognitive cycles that should go to idea generation. Speech, by contrast, flows at ~150 words per minute—triple the typing speed—and offloads the burden of structure to the AI. The AI doesn’t just receive orders; it becomes an active collaborator, probing for missing context and clamping down on ambiguity. This is a paradigm shift from “code is law” to “dialogue is law.” In crypto, where news moves in 15-minute windows and a missed nuance can cost you alpha, this method could be the missing link between raw data and trading decisions.
But let’s be clear: this is not about using AI to write headlines. It’s about using AI to think with you—to sift through the noise of a hundred Discord channels, Telegram leaks, and regulatory rumors, and distill the signal. I’ve been in this game since the Tezos FOMO sprint in 2017, where I learned that the fastest source isn’t always the smartest. Karpathy’s method directly addresses the problem: how to turn a flood of unstructured information into a concise, verified insight before the market moves.
Core
Here’s how it works in practice for crypto news analysis. Imagine a Tuesday afternoon in a bull market. You’re scanning on-chain data, scrolling through X threads, and juggling three Telegram groups. A rumor drops: “Uniswap v4 deploy on Base delayed due to smart contract bug.” Your first instinct? Panic-sell or buy the dip? In the old world, you’d open a terminal, search for the official announcement, check Etherscan, and maybe write a quick Python script to simulate slippage. That takes 10 minutes. By then, the market has already priced in the noise.
With Karpathy’s method, you instead speak your thought process into a voice recorder for 2 minutes: “Okay, Uniswap v4 delay, Base channel says it’s confirmed by core devs, but I saw a conflicting tweet from a validator. Check if the smart contract address has been updated on Base’s GitHub. Also, look at the liquidity pool—any abnormal withdrawals in the last hour? Let AI ask me for what I’m missing.” You feed the raw audio to a model like Claude 3.5 Sonnet with a system prompt: “You are a crypto analyst. Listen to my thought stream, identify gaps, and ask three clarifying questions before providing a summary and a risk score.”
The AI processes the chaos, asks: “1. Did the Base team issue a public statement? 2. Are there any governance proposals on Uniswap v4 that would override the delay? 3. What’s the current TVL in the affected pools?” You answer in voice, the AI integrates the answers, and within 3 minutes you have a structured report: “Delay confirmed by a single source, two conflicting tweets from validators. TVL unchanged. Risk score: 3/10—likely market overreaction.” You publish the news bite, your subscribers get it 7 minutes before CoinDesk picks it up. That’s alpha.
I tested this workflow personally during the recent AI agent wallet controversy in 2026. I had 15 minutes of fragmented voice notes from three sources—an auditor, a regulatory insider, and a chatbot log. I dumped the entire transcript into a model, let it ask me for clarification on wallet addresses and timestamp mismatches, and within 60 seconds it generated a timeline of suspicious transactions. The output was not perfect—it hallucinated a connection between two unrelated wallets—but the active questioning caught my error before I published. The real win is that the AI forced me to think through the gaps I had glossed over. “You said 60% of wallets funnel to mixers—what’s the data source?” My answer: “Etherscan API.” The model probed: “API has a 24-hour lag; are these recent transactions?” I had to verify. That kind of collaborative skepticism is rare in human analysis, and even rarer in automated news feeds.
But the method’s true power lies in contraction. The AI doesn’t just regurgitate your stream; it extracts the few high-signal statements—the ones that will move the price. In my DeFi Summer 2020 Uniswap v2 arbitrage deep dive, I spent 3 nights writing a Python script. Today, I would verbally describe the same logic: “I want to compare Uniswap and SushiSwap for token A, calculate optimal route using constant product formula, account for slippage, and generate a graph of profit vs. pool depth.” The AI would turn that into executable code in seconds. The bottleneck shifts from typing code to thinking clearly.
Contrarian
Here’s the blind spot everyone is missing: verbal prompts amplify confirmation bias and hallucination risks exactly when speed matters most. In crypto, a misheard number or a hallucinated contract address can trigger a cascade of bad trades. Karpathy’s method relies on the model’s ability to reconstruct “real intent” from chaotic input. But what if the model reconstructs a wrong intent because the underlying model has a built-in bias? For example, during the 2022 FTX collapse, I live-blogged a “Trust List” of solvent VCs. If I had used this method back then, a model primed to trust certain names (say, Alameda-related entities) might have regenerated a false narrative. The active questioning loop can catch some errors, but it’s not foolproof.
Another contrast: the method assumes you have a high-quality model with long context and reliable questioning capabilities. Most public models today still struggle with mid-context coherence over 10-minute transcripts. Claude excels here; GPT-4o is close. But Llama 3.1 70B? Forget it—token limits and weak intent inference will turn your voice memo into spaghetti. This creates a proprietary advantage for closed-source models, which may tighten the moat around AI-as-a-platform. In crypto, we already have a bias toward trustless, open-source solutions. This tension between the need for speed and the desire for sovereignty is exactly the kind of contradiction that makes this story compelling.
Finally, the user skill trap: over-relying on verbal prompting can atrophy your own ability to structure arguments. In bull markets, euphoria masks technical flaws. In AI analysis, convenience masks cognitive decay. The best traders still read order books, not just abstracts. The best news aggregators still verify with human intuition. Karpathy’s method is a tool, not a crutch. If you let the model do all the thinking, your edge erodes. I’ve seen this happen with junior analysts who depend entirely on ChatGPT for trade ideas. They lose the “hunter’s instinct” for spotting anomalies in on-chain data. I don't read whitepapers; I read order books. That sentence has never been more relevant.
Takeaway
The “long-form verbal prompt” method is not a silver bullet, but it is a signal. It tells us that the future of crypto news aggregation is not about writing better prompts—it’s about having better conversations with AI. The next killer app in this space will be a cross between a voice recorder and a real-time fact-checking engine that asks you “What did you miss?” before you hit publish. I’m already building it into my Crisis Watch section: every 15-minute update will start with a verbal dump, let the AI question my raw thoughts, and only then produce the final alert. It’s faster, more accurate, and forces me to reconsider my assumptions.
But the question that haunts me: will this method make us faster at the cost of deeper insight? In a bull market, speed is oxygen. But in a black swan event, the fastest analysis can be the most dangerous. The best news is the news that moves the price, but the most valuable news is the one that doesn’t crash your portfolio. Karpathy’s method is a double-edged sword—sharpen it wisely.