How the engine actually works.

This project wasn't made by clicking around a chatbot. It was co-developed: a human and an AI working as partners in a real programmer's terminal. The human brought things a computer can't know on its own — like the way crypto reliably gets wild late at night when few people are trading (the "witching hour" windows), watched first-hand during real trades — and the rules to stay safe. The AI turned those ideas into working code, tested them, and wired up the data. (Kip is the trader; right now his dad is his "lab partner," helping type and check the work — the way a science-fair coach helps run the experiment.)

Most people picture a chat window. The real work happens in a terminal — a person typing commands to direct the AI, then checking that what it built actually works. Here's a real slice of building Kip Capital:

Real commands, real output. A human decides what to build and checks it; the AI does the heavy lifting and gets corrected when it's wrong. That back-and-forth — not a single magic prompt — is how the engine was made.

All three lenses, the current portfolio, and a living strategy playbook get written into one prompt. Claude must answer in a strict format — action, symbol, size, confidence, thesis, stop-loss, take-profit — so its decision is machine-readable and human-readable. The thesis is the part you see on the front page: the AI explaining, in plain English, why it did what it did. The playbook grows over time as patterns are learned, so the machine gets a little wiser each cycle.

The AI never gets the last word on risk. A separate risk gate enforces rules it cannot override:

• Position cap — no single coin above ~25% of the bucket (scaled down near cycle tops).
• Exposure cap — keep cash in reserve; never go all-in.
• Daily circuit breaker — if the bucket drops past a set line in a day, stop trading.
• Confidence floor — ignore low-conviction ideas.
• Never average down — don't throw more money at a losing position.

If a decision breaks a rule, the gate shrinks it or throws it out. This is the difference between a science project and a gamble.

This is the part that makes it more than a script. Every completed trade is saved along with the exact conditions it happened under — the trend, the time-of-day window, the cycle phase, the confidence. Then, once a night, a separate reflection engine reads that whole track record, measures what actually made or lost money (win rate, profit factor, which setups paid off), and asks the AI: "given the evidence, what should we do more of, and what should we stop doing?"

The AI writes those lessons — each backed by a real number — straight into its strategy playbook, which the trader reads before every decision. So the bot literally gets better as it gains experience, the same way a person does. Two guardrails keep it honest: it needs enough trades before drawing a conclusion (so it doesn't overreact to luck), and it can only tune the strategy — never the safety rails. It learns to trade smarter, not to remove its seatbelt. You can watch the lessons appear on the live board.