Deep Dive
Why execution automation matters more than strategy selection
Miles opens by stating that most traders lose not because their directional ideas are wrong, but because execution breaks down — missed entries, poor exits, profit-taking at the wrong moment. He pivoted to trading bots in 2023 after noticing he was round-tripping too much capital. Earlier AI models made bots impractical, but recent improvements in Claude and GPT have made the approach viable again. Critically, exchanges like Bybit have only recently shipped native Model Context Protocol (MCPs) and AI toolkits; without these, connecting an LLM to an exchange was technically impossible. This convergence of better models plus exchange infrastructure means what Miles is showing was literally impossible six months ago. He's clear this isn't a set-and-forget money printer, but a customizable execution layer where you plug in any strategy — whether algorithmic, discretionary, or hybrid.
Setting up the technical stack: Claude, MCP, exchange, and memory
The setup requires a Claude subscription (or ChatGPT with Codex), Node.js installed locally, and an exchange account supporting MCP natively — Miles uses Bybit. The architecture flows: Claude acts as the brain sending instructions to the MCP, which relays commands to the exchange for order execution, and the exchange reports back data through the MCP to Claude. Before any trading, you build a memory system in a folder on your computer containing a context file describing your financial situation and goals, an instructions document defining how Claude should behave, and a memory.md file that updates based on your learnings. Miles recommends voice prompting with Claude first to co-create this profile, converting the conversation into markdown. The key advantage over Bybit's built-in grid and DCA bots is full customizability — you can ask Claude to execute a specific Pine Script backtest, buy Bitcoin when price drops 10% and volatility hits 1.5 on a custom index, or run any rule you define. Once you create exchange API credentials with trading permission enabled and withdrawals off for security, Claude can verify the connection and make test orders to confirm the pipeline works end-to-end.
Building feedback loops: ledgers, learnings, and reflection before every trade
The execution layer is only half the system; the other half is structured learning. Miles implements two critical files: a ledger that logs what fired, when, why, and the outcome, and a learnings file that the bot writes in plain English after every closed trade summarizing what was learned. He goes further by encouraging traders to add personal reflections before entering a trade — stating your thesis, the catalyst, and your invalidation price — then reflecting after the position closes. This forced reflection prevents the impulse trading that plagued him earlier. However, he warns against over-adjusting: most great traders have 40% win rates, so a single loss shouldn't spawn a new rule. Instead, observe patterns over time, and only introduce a rule once you're confident in the pattern. Over-engineering kills profitable strategies by adding too many conditions. The power emerges over years of compounding data — every trade logged, every outcome recorded, every decision documented. AI excels at processing this data retroactively, allowing you to recall exactly why you entered a trade months ago and assess whether the pattern held. This removes the human brain's limitation of forgetting specifics and creates a systematic, auditable approach to trading.
Remote execution and integration with Telegram for hands-free trading
Miles extends the system to work from his phone via Telegram. He sets up a Telegram bot using BotFather, connects it to his Claude session (running 24/7 on a separate device like a Mac Mini or $10 VPS), and uses WhisperFlow to send voice messages that transcribe into text and reach Claude. This allows him to query account balances, portfolio changes, and place trades while traveling without touching a computer. The constraint is that Claude only accesses files on devices currently running; cloud-based solutions like Claude Dispatch face the same limitation. He prefers Telegram since he's already messaging on it for business. This setup extends to stock portfolios too — he's built a similar system for Interactive Brokers using TWS (the trading workstation application), allowing him to monitor and trade equities using the same execution framework. The pattern generalizes: any brokerage or exchange with an API or MCP connection can be automated, whether crypto, stocks, or commodities. By automating execution and making it accessible from his phone, Miles spends far less time at a computer while staying continuously updated — a major quality-of-life win given his travel schedule and conference attendance.