Hook
On a quiet Tuesday, just weeks after the FTX collapse, I watched a non-technical friend record a sequence of clicks, keystrokes, and voice commands on Claude Cowork. Within minutes, the AI had generated a reusable “Skill” that could, on command, swap ETH for USDC on a DEX, bridge it to Arbitrum, and deposit into a yield farm. No code, no terminal, no YAML file. Just a demonstration. Then I remembered: OpenAI Codex had launched the exact same feature on the same day. Two of the world’s most advanced AI labs had converged on the same product — and the crypto industry was not paying attention.
Context
For years, crypto automation has belonged to the technical elite. DeFi strategies are executed via bots written in JavaScript or Rust, linked to Chainlink keepers or Gelato networks. Smart contract developers spend weeks writing and testing scripts for simple recurring tasks. The marginal user — the lawyer managing a DAO treasury, the artist stacking royalties — is locked out of automation entirely. The result is a fragmented landscape where only 12% of DeFi users actively deploy any form of automated strategy, according to Dune Analytics data from Q1 2025. The rest rely on manual trades, exposed to slippage, timing errors, and emotional decision-making.
Now, Anthropic and OpenAI have introduced a mechanism that directly addresses this bottleneck. By recording screen activity, mouse clicks, keyboard inputs, and voice narration, their models convert human demonstration into a structured “Skill” — a reusable, parameterizable workflow that can be executed autonomously. This is not a new model architecture. It is an engineering-level combinatorial innovation that grafts behavioral cloning onto a multimodal foundation. But for crypto, it could be the bridge that finally brings non-technical users into the world of on-chain automation.
Core Insight
The technology behind Skill recording is deceptively simple and profoundly powerful. At its heart, it treats a user’s demonstration as a sequence of observations (screen state, audio transcript) and actions (mouse clicks, keyboard entries). The AI learns a policy that maps these observations to the next action, then stores the entire mapping as a structured prompt — a combination of natural language instructions, script snippets, UI element selectors, and resource paths. When the user later invokes the Skill, Claude re-initializes the environment, captures the current screen state, and generates the next action step by step, mirroring the original demonstration but adapting to minor UI changes through semantic understanding.
For crypto, the implications are double-edged. On one hand, this capability could democratize DeFi automation. Imagine a user recording a Skill for a “harvest and compound” loop on a yield aggregator: connect wallet, review pending rewards, click “harvest,” approve transaction, then re-deposit. The Skill would capture not just the GUI interactions but also the voice explanations of why each step matters — turning the recording into a teachable artifact. Over time, skill libraries could emerge: “Uniswap V3 rebalancing,” “Lido staking rewards sweep,” “Gnosis Safe multisig approval flow.” Non-technical treasury managers could build workflows by showing, not writing.
But here is where the crypto context amplifies the risk. The recording process captures everything on the user’s screen — wallet addresses, private key pop-ups (if poorly designed), portfolio balances, transaction hashes, and even passwords typed into web forms. This data is uploaded to Anthropic or OpenAI servers to generate the Skill. The resulting Skill may contain hardcoded addresses, API keys, or even audio recordings of sensitive discussions. If Skills are shared — and both companies are hinting at future Skill marketplaces — those secrets travel with the workflow. The privacy risk is not theoretical; it is structural.
I have seen this movie before. In 2020, during DeFi Summer, I audited a protocol that had embedded API keys directly into their smart contract metadata. The keys were visible on Etherscan for three weeks before I reported it. “Code is law, but humans are the protocol” — and humans make mistakes. The same principle applies here: a Skill that automates a complex transaction sequence may inadvertently expose the user’s entire financial history to a third-party server. The security model must be redesigned from the ground up.
Yet there is a path forward. The most promising development is the potential for local execution. If the inference can be run on-device (using a smaller model like Claude Haiku or a quantized variant), screen data never leaves the machine. The Skill itself could be encrypted and signed by the user’s wallet, ensuring that only they can decrypt and execute it. This would preserve the automation benefits while containing the privacy cost. Based on my experience building ChainBridge workshops in 2017, I learned that the most trusted tools are those that hand control back to the user. Education is the antidote to exploitation.
Contrarian Angle: The Real Bottleneck is Not Tech — It’s Trust
The industry narrative is that AI Skill recording will “unlock mass adoption” for DeFi automation. I disagree. The real bottleneck is not technical capability; it is the lack of a trusted verification layer. Even if the AI flawlessly replays a recorded sequence, the user still must authorize each transaction via their wallet. The Skill cannot sign on its behalf — and for good reason. But new users, desperate for convenience, may be tempted to store private keys within the Skill’s prompt or rely on the AI to “handle the signing.” This is a catastrophe waiting to happen.
Moreover, the probabilistic nature of LLMs conflicts with the deterministic execution environment of a blockchain. A Skill that succeeds 99% of the time for a GUI application might fail catastrophically on-chain: a misread button label could result in the wrong token swap, or a hallucinated step could cause a bridge to send funds to an unintended address. The industry is not ready for these failure modes. We built trust in the chaos, not despite it — meaning we must embrace the chaos of probabilistic automation with rigorous guardrails, not pretend it doesn’t exist.
The contrarian opportunity lies in building a verification layer that sits between the AI-generated Skill and the actual blockchain transaction. Think of it as a “Skill sandbox” that simulates the intended workflow on a testnet, compares the expected outcomes with a user-defined threshold, and only then submits the real transaction. This is where blockchain education meets product design. The platform I founded in 2023, ChainLink Edu, now offers a module that teaches users to audit their own Skills — because if we don’t teach them, the market vendors will exploit them.
Takeaway: The Future Belongs to Those Who Teach Together
Anthropic and OpenAI have handed the crypto industry a powerful tool, but it is a tool without a manual. The race to acquire users by ease-of-use will be won by the team that also invests heaviest in user education and safety protocols. As I wrote in my 2024 whitepaper “Beyond the Bullion,” the transition to mainstream adoption requires not just better interfaces, but better frameworks for understanding responsibility.
“Hold through the noise, build through the silence.” The release of Skill recording is noisy hype. The quiet work — building local encryption, sandbox verification, and teachable skill libraries — is where value will compound. Education is the antidote to exploitation. Those who teach together will inherit the future of on-chain automation, because they understand that trust is earned in drops and lost in buckets.
The question, then, is not whether AI can record your workflow. It already can. The question is: will you have the wisdom to verify before you execute? “Code is law, but humans are the protocol.” That is a lesson no AI can teach us — only experience can.