Using Claude to Review Your Trades the Right Way

Most people who try to use AI to review their trades make the same mistake: they paste a spreadsheet into a chat window and ask the model what it thinks. The output is usually the kind of pleasant, hedged generality that sounds useful until you actually try to act on it. Brendan, who runs the AI Pathways channel on YouTube, has built something more disciplined than that — and the engineering logic behind it is worth understanding whether or not you ever build the thing yourself.

The instinct to just dump your trade history into Claude is understandable. AI chatbots are conversational, responsive, and genuinely impressive at pattern recognition. But Brendan's central argument cuts straight to the problem: "You can't just feed your trades to Claude and then have it review, because one, it doesn't have minute-by-minute bars for the past however many days, weeks, or months. And then also, it just can't compute all that information in one query."

That's not a limitation unique to Claude. It's a structural problem with asking any language model to do work that belongs to a database and a calculator. A year of one-minute price bars — each recording what a stock opened at, reached at its high and low, and closed at, along with how many shares traded — runs past half a million rows of data. That's not going to fit in a chat window. And even if it did, having the model perform arithmetic across that many data points is asking it to do something it's genuinely not good at. Language models are trained to predict text. Deterministic math — calculations that produce the same answer every single time, no interpretation involved — is better handled by code.

So Brendan built a five-layer system where the hard math never touches Claude at all.

The structure is cleaner than it sounds. A trader exports their history from a broker like NinjaTrader and uploads it to the system. Separately, real price bar data gets pulled from a source like Yahoo Finance, which offers around 60 days of one-minute data for free. The system then runs what Brendan calls the "deterministic engine" — code that checks every single trade against what the market was actually doing at the time: whether the stop-loss would have been hit before the target, what the exact exit price would have been accounting for slippage, how the trade's timing aligned with the rules the trader set in advance. None of this requires AI. It's arithmetic, run reliably and cheaply.

The behavioral layer is where it starts to get interesting from a human perspective. The system tracks things like whether a trader increased their position size right after a loss — revenge sizing, in trader parlance — or whether they kept cutting their winning trades short while letting losers run. It calculates a "discipline cost": what your performance would have looked like if you'd followed your stated strategy exactly, compared to what actually happened. That comparison is the kind of mirror most traders prefer not to look in.

Claude enters only at the end of this pipeline. The engine's outputs — win rates, drawdown figures (peak-to-trough losses), risk-adjusted return measures, trade classifications, behavioral flags — get compressed into a text summary of a few kilobytes and sent to Claude via API. Claude's job is to read the findings and translate them into coaching. As Brendan puts it: "The code does the math. AI reads the results."

The demo walks through a sample of 328 trades. The system classifies each one by strategy family — momentum, trend continuation, mean reversion, pullback — then maps them against what the trader said they were trying to do. The coaching output produced in the demo flagged a mismatch between the trader's stated approach and where the actual profits were coming from. It suggested specific changes: don't force a rigid exit model, revisit position sizing, look at which hours of the day are generating returns versus bleeding them away. That kind of layered interpretation — reading the numbers and then reasoning about what to do differently — is genuinely what Claude is good at. It's pattern recognition and synthesis, not arithmetic.

The build-it-yourself pitch is where broader trends in AI tooling become relevant. Claude Code — Anthropic's agentic coding tool — is the construction vehicle here. Brendan walks through five layers of prompts that assemble the full system, and his core claim is that you don't need to be a programmer to use them. That claim aligns with a broader shift observers have been noting: according to one recent analysis, Claude Code is actively turning non-programmers into builders. Whether any particular viewer can replicate Brendan's system by following his prompt sequence is a harder question — complexity tends to accumulate in the gaps between layers — but the scaffolding is at least more accessible than it would have been two years ago.

The system includes twelve preset strategy families — breakout, mean reversion, trend following, scalping, and others — or the option to describe your approach in plain English and let the AI translate it into machine-readable rules. Type "I trade momentum long only, half a percent stop, aim for two-to-one from 9:30 to noon," and the system fills in the session parameters and rule schema automatically. That input gets stored so you don't have to describe your strategy again the next time you upload a new batch of trades.

What I find genuinely useful about this approach isn't any single feature — it's the underlying discipline of keeping AI in its lane. The failure mode Brendan is designing around is real and common. Traders who dump their history into ChatGPT or Claude and get back general commentary aren't getting analysis; they're getting plausible-sounding text. The system Brendan built forces specificity: the math runs on real price data, produces real numbers, and only then asks the language model to help make sense of them. That's a principled architecture, not just a clever trick.

The limitations are worth naming plainly. This is a post-hoc review system, not a predictive one. It tells you what you did wrong after you've done it. That's valuable — most traders don't have a systematic way to audit their own behavior — but it doesn't address the harder problem of whether a strategy that worked in the sample period will hold up going forward. The system can tell you whether you followed your rules; it can't tell you whether your rules are any good in a regime it hasn't seen yet. Brendan acknowledges this, at least partially, when he notes the coaching output will tell you whether your strategy has a "real edge" — but edge is regime-dependent, and no amount of historical analysis fully resolves that.

There's also the behavioral question that the system surfaces but can't answer: knowing you revenge-size after losses and actually stopping are two different problems. Data can diagnose a habit. Changing it is something else entirely.

That's the part of this story that tends to get buried under the architecture diagrams. The tool is useful. The discipline it requires to use it honestly — to look at your own discipline cost and sit with what it shows — is the part that actually determines whether any of this moves the needle.

---

Bob Reynolds is Senior Technology Correspondent at BuzzRAG.