Hook
Red candles don't lie, but the machines learning to shape them? That's a whole new ballgame. This morning, two of the most well-funded AI labs on the planet—Anthropic and OpenAI—simultaneously dropped a feature that reeks of copy-paste engineering: "Record a skill." Both Claude Cowork and OpenAI Codex now let you record your screen, clicks, keystrokes, and voice to create a reusable automation workflow. No code. No scripts. Just a demonstration of a repetitive task, which the AI then turns into a ghost that can run that task forever.
For a market surveillance analyst who spends 7x24 staring at order book manipulation and wash trading patterns, this is the digital casino finding a new way to rig the slot machines. The immediate impact? The barrier to deploying automated DeFi bots just collapsed from a week of Solidity debugging to a five-minute screen recording. Exit liquidity is someone else's problem now—but whose? Let's peel the layers off this engineered flash crash before the liquidity drains.
Context
To understand why this matters, you need to remember that the current DeFi automation landscape is a graveyard of half-baked scripts. Retail traders use Telegram bots that scrape Discord signals. Fund managers run custom Python bots that execute perpetual swaps. But the critical bottleneck has always been the "cold start" problem—creating a reliable automation pipeline from scratch requires both domain expertise and a tolerance for debugging at 3 AM when a sushi swap fails because the Görli testnet faucet was empty.
The "Record a skill" feature promises to kill that bottleneck. On the surface, it's a product play: Anthropic positions Claude Cowork as the office assistant that can handle your Excel and browser tasks, while OpenAI Codex targets the developer who wants to script their entire workflow with a single prompt. But both are now converging on the same core: a multimodal recording->understanding->replaying loop. They watch your screen, they hear your voice, they see your mouse, and they hallucinate a step-by-step plan to do it again.
Wash trading: The digital casino of AI agents is about to get its first automated dealers. And the house always takes a cut.
Core
Let's cut through the marketing and look at the actual technical meat. The "Record a skill" feature is not a breakthrough in model architecture. It's an engineering-level combinatorial hack that glues together existing components: screen recording APIs, UI element detection, speech-to-text, a large language model for intent parsing, and a code execution sandbox. The output is a structured "Skill"—likely a mix of natural language instructions, Python/Powershell scripts, and UI element selectors—that the model reads like a cookbook to generate real-time actions.
From my experience analyzing on-chain data, I've seen similar patterns in the evolution of MEV bots. The early bots were hardcoded with arbitrage logic. Then came the GPT era where bots used LLMs to read mempool and generate transactions. This feature takes it a step further: it's behavioral cloning for the desktop GUI, where the AI learns a policy from your demonstration and replays it in a similar environment. The key question is how robust that policy is when the environment changes—when the "Swap" button moves two pixels to the left, or when the liquidity pool has a different curve.
The article—a classic PR fluff piece—states that users can "directly execute similar tasks next time." But it never mentions the failure rate when the GUI changes. In my audits of DeFi protocols, I've seen exactly this pattern: a script that works perfectly on testnet but falls apart on mainnet because the order of dropdown options differs by one. The hidden challenge is semantic robustness: can the AI understand that the "Confirm Swap" button on a Uniswap clone is functionally identical to the one on PancakeSwap, even though the styling is different? Based on current multimodal capabilities, I'd estimate a 60-70% success rate for identical tasks, dropping to 30% for similar tasks with UI variation. That's not production-grade.
Another layer: the internal representation of a Skill. The original article mentions that previously creating a Skill required manually writing a SKILL.md file. This suggests that Skills are not freeform neural networks but structured prompts that the model parses. The recording process is essentially generating a high-quality prompt with attached media (screenshots, audio transcript). When the Skill runs, the model re-ingests the current screen and the Skill prompt, then generates actions. This is computationally expensive—each execution step requires a full forward pass of a large multimodal model. The inference cost is likely 10-100x the recording cost. For a crypto trader running a bot that checks price every minute, that adds up quickly in API fees.
Now, connect this to the current bear market. Survival matters more than gains. Over the past 7 days, several DEXs have lost 40% of their LPs because automated market making bots couldn't react fast enough to volatility. The promise of "Record a skill" is that anyone can now create a bot that does one thing—say, rebalance a stablecoin pool every hour—without writing code. But the risk is that these bots are brittle. A single UI change or a mislabeled button could cause the bot to execute a trade on the wrong pair, turning a DCA strategy into a rug. I've seen it happen with amateur scripts. Now it will happen at scale.
And what about the true monsters: AI agents that record other agents? Imagine a scenario where a sophisticated player records a Skill that performs a complex arbitrage across three exchanges. Then they sell that Skill to retail users. The Skill works for a week, then the creator introduces a subtle bug—maybe adds a tiny delay that makes the arbitrage unprofitable for everyone else but profitable for themselves. That's the essence of wash trading: the digital casino of automated Skills will have its own market makers who front-run the recorded bots.
Contrarian
Everyone will focus on the productivity gains—how "Record a skill" will democratize automation. But the unreported angle is the systemic risk to data privacy and financial security. Let's be cynical.
The feature records your screen, keystrokes, and voice. If you're a crypto trader, that means it captures your API key entry, your seed phrase typing, your wallet address, your private Discord channels. All that data is uploaded to Anthropic's or OpenAI's servers to be processed into a Skill. The article mentions nothing about data local processing or encryption at edge. For a retail user, that's a massive attack surface. For a professional fund, that's a compliance nightmare. Imagine a Skill that automatically claims airdrops—the recorded Skill contains your wallet interaction pattern, which could be reverse-engineered to derive your trading strategy.
The contrarian take: this feature is not a tool for automation; it's a tool for mass surveillance and model extraction. The AI labs use your recordings to train their next models. The Skills themselves become a vector for data exfiltration. In a bear market, when trust is thin, the last thing we need is another vector that can drain wallets. Exit liquidity is someone else? Not if you just recorded your private key transfer method.
Furthermore, the skill-execution reliability issue is a ticking time bomb. In a competitive market like crypto, milliseconds matter. A Skill that has even a 10% chance of failing mid-execution will cause more losses than savings. The original article's tone is overwhelmingly positive, but as a surveillance analyst, I see the negative feedback loop: users create bots, bots fail at critical moments, users lose funds, trust in the platform evaporates. This is the typical hype cycle of every DeFi innovation.
Another blind spot: regulatory arbitrage. The SEC hasn't defined AI agents as financial actors yet, but a "Record a skill" that executes swaps automatically is essentially an unregistered bot. The moment a Skill executes a buy order based on recorded conditions, it could be classified as an algorithmic trader. The labs are ignoring these implications. The bear market reduces regulatory pressure, but the seeds of future lawsuits are being planted.
Takeaway
So what do you do? Don't record your trading workflows on any AI desktop agent right now. Treat them as unvetted third-party scripts. If you need automation, stick to audited, open-source bots that you can review. The race to sell "Record a skill" as a productivity hack is a distraction from the fundamental truth: in a system where the house controls the recording and the replay, the only winning move is not to play. But if you must play, watch for the first leak of a Skill that empties wallets. That will be the signal to short the hype. Red candles don't lie—the wash trading of AI skills has already begun.