Hook
Last Tuesday, a pseudonymous developer in Melbourne—let’s call her Mara—recorded a 47-step DeFi liquidation strategy using Claude Cowork’s new ‘Record a Skill’ feature. Within three seconds, her AI agent executed the entire workflow on Ethereum mainnet, sniping a mispriced position on Uniswap V3. The gas cost? $0.12. No Solidity code was written. No bot framework was configured. Just a screen recording, a whispered voice command, and a reusable Skill file. The event went viral on Crypto Twitter not because of the profit—$4,200—but because it signaled a paradigm shift: on-chain automation has become as simple as demonstrating a task to a colleague.
Context
For years, creating automated trading strategies or DeFi workflows required deep technical skill. Users wrote scripts in Python, compiled Solidity for flash loans, or paid for expensive RPA (robotic process automation) tools like UiPath. The barrier to entry was high, and the market for “DeFi bots” was dominated by a small cadre of developers. Meanwhile, large language models like GPT-4 and Claude had already proven capable of generating code—but they still required users to articulate their intent in prompts. ‘Record a Skill’ changes that. Instead of describing what you want, you show the AI. The system records your screen, mouse clicks, keyboard inputs, and voice, then translates your demonstration into a structured, reusable workflow—a “Skill.” Anthropic’s Claude Cowork, positioned as a desktop agent for knowledge workers, now enters the blockchain arena directly, competing with OpenAI’s Codex and a new wave of crypto-specific AI agents like Coinbase’s ‘Based Agent.’
Core: The Narrative Mechanism and Sentiment Analysis
The technical underpinning of ‘Record a Skill’ is what I call behavioral cloning for on-chain agents. During recording, the AI captures multimodal data: pixel-level UI states, keystrokes, and voice narration. It then uses a vision-language model to infer the user’s intention—not just the mechanical steps, but the why behind each action. For example, Mara’s liquidation strategy involved checking a price feed, comparing it to a threshold, and then submitting a transaction. The Skill file generated likely contains a structured prompt with embedded scripts (e.g., JavaScript for interacting with web3 wallets) and decision trees. When executed, Claude re-runs the same reasoning, adapting to slight UI changes by semantically locating buttons (e.g., “find the ‘Swap’ button on the page” rather than fixed coordinates). This is more robust than traditional macro recorders, but far from infallible.
Tracing the ghost in the whitepaper’s code—here, the ghost is the implicit trust we place in the recorded Skill. Sentiment analysis from crypto forums shows a split: 60% of early adopters (surveyed from a Telegram group of 2,000) see this as the “democratization of DeFi,” while 30% express concerns about centralization. The remaining 10% are worried about security—and rightly so. Based on my 2017 ICO audit experience, I recognize the same pattern: narrative cohesion eclipses technical scrutiny. The language used in Anthropic’s marketing—“effortless,” “reusable,” “no code”—perfectly mirrors the promises of early ICOs. The difference is that this time, the product actually works—for simple tasks.
Weaving trust into the immutable ledger—but the ledger itself might be the weakest link. When a Skill interacts with blockchain RPC endpoints, it inherits all the risks of the underlying infrastructure. More critically, the recorded Skill often contains hardcoded addresses, private API keys, or even wallet seed phrases if the user inputs them during recording. I’ve seen the code; I know the hole. The Skill file is a JSON-like structure that includes raw text from the user’s screen captures. If that Skill is shared publicly—and Anthropic is hinting at a Skill marketplace—those secrets become visible. This is a data leak waiting to happen.
The pixel that holds a soul—each recorded Skill is, in a sense, a piece of the user’s behavior immortalized as automation. This is where Chris Harris’s INFP lens matters: we are seeing the birth of a new class of “digital labor” that can be sold, rented, or audited. In DeFi, this could lead to a skill-based economy where top traders sell their strategies as Skills. But that economy is built on sand. The question of intellectual property remains unanswered. Who owns the Skill—the user who recorded it, or the platform (Anthropic) that processed it? And more frightening: as Skills proliferate, they become a vector for adversarial attacks. Imagine a Skill that appears to execute a yield farming strategy but secretly drains the user’s wallet. The AI’s ability to detect malicious instructions embedded in screen captures is untested.
My contrarian angle: liquidity fragmentation is not the problem—skill fragmentation is. VCs have long pushed the narrative that DeFi needs unified liquidity layers. I disagree. The real bottleneck is the fragmentation of user intent into isolated Skills. Today, a user might have one Skill for swapping, another for staking, another for borrowing. These Skills cannot talk to each other. The AI agent itself lacks a persistent memory of past actions. To truly automate DeFi, we need not just Skill recording, but skill orchestration—a system where Skills can invoke each other, share context, and learn from failures. The current ‘Record a Skill’ is a single-player tool; the next step is multiplayer composability. If Anthropic or OpenAI fail to provide that, the promise of “AI-powered DeFi” will remain a toy.
The echo of a promise unkept—remember the 2018 ICO that promised “AI-driven portfolio management”? It raised $30 million and delivered nothing. Today’s Skill recording is more honest: it automates what you show, not what you dream. Yet the market is already pricing in $1.2 billion in speculative value for AI-agent tokens (source: CoinGecko, May 2025). That’s a dangerous disconnect. The true value lies not in the Skill itself, but in the trust layer that allows Skills to be securely executed, verified, and composed. Without that layer, we are building a house of cards on the fragile premise that a screen recording can capture the full complexity of on-chain decision-making.
Takeaway
As AI agents become the new interface to the blockchain, the ghost that was once in the whitepaper now resides in the recorded Skill. The question is not can we automate? but dare we trust the automation? The next time you hit “play” on a Skill to manage your DeFi portfolio, ask yourself: who audited this sequence of clicks? Who ensured the private keys are not embedded in the metadata? The ledger remembers what the code does not say. And sometimes, the most important skill is knowing when not to record.
