Fair warning. This week’s issue is more technical than what I usually send.

But if you spend the time with it, this is actually very doable. And I think this is what you should be aiming to get to with AI for finance.

I built a full options flow screener using Claude Code.

The idea is: insiders with informational advantage may place "unusual" options activity which can be picked up in real time using the right data feeds.

It scans the entire U.S. options market, filters through institutional-grade criteria, cross-references insider and congressional trading data, and ranks plays by conviction score.

I didn’t write a single line of code.

Whether you’re an individual trader looking for idea generation or a sophisticated institutional investor running a multi-strategy book, you can build workflows like this on Claude Code right now. Without an engineering team!

That’s the point of today’s issue. I’m going to walk you through exactly how I built this, step by step, so you can build your own version with your own screening criteria.

But first, let me explain why options flow screening is worth building in the first place.

Why Options Flow Works

Options markets are one of the few places where informed money ("insiders" lol) leaves footprints in real time.

When someone with an informational edge wants to express a view, they often use options. A board member, an executive, a fund that just finished diligence on an acquisition target.

The leverage is better than stock. The position is harder to trace.

But this signal is buried in noise. Thousands of trades hit the tape every day.

Most of them are market makers hedging, institutions rolling positions, or retail buying lottery tickets.

Building a system that separates the signal from the noise automatically is the opportunity. That’s what I built.

Here’s the plan:

Connect Claude Code to a live options data feed using MCP servers
Define an 8-step screening funnel with institutional thresholds
Cross-reference survivors with insider trades, congressional activity, and social sentiment
Score and rank by conviction, then save every scan for backtesting

The Build

What Is Claude Code & Why Does This Matter

Claude Code is Anthropic’s AI coding agent that runs in your terminal. You talk to it in plain English, and it builds things for you.

What makes it different from ChatGPT or regular Claude is that it can use tools. It can read files, write files, and run commands.

But the big one for us is that it can connect to live data sources through something called MCP servers.

MCP stands for Model Context Protocol. Think of it like plugging your AI into Bloomberg.

Instead of copying and pasting data into a chat window, MCP lets the AI query real-time financial data directly. Options flow, insider filings, congressional trades, earnings calendars, dark pool prints, social sentiment.

All accessible through the same interface.

I connected two MCP servers for this build:

Unusual Whales MCP gives the AI access to live options flow alerts, the full options tape, a contract screener, insider transactions, congressional trades, open interest changes, earnings calendars, and dark pool prints
X MCP gives it access to X/Twitter for sentiment scanning and catalyst research

Once connected, the AI can query these sources dynamically. If it finds unusual flow on a ticker, it can pull that ticker’s insider history, check for upcoming earnings, and search Twitter for rumors on its own.

All in the same pipeline, without you doing anything.

You describe what you want the screener to do in plain English, and Claude Code builds it.

You run it whenever you want. The results save to disk so you can track performance over time.

P.S. If you want to learn how to use Claude Code for Finance, I'm setting up a LIVE bootcamp ... sign up to get on the waitlist:

Claude Code for Finance Bootcamp Waitlist

Step 1: Connect Your Data Sources

Setting up MCP servers on Claude Code takes about 2 minutes. You run a single command in your terminal for each server.

For Unusual Whales, you need an API key from their site. For X, same thing.

You paste one command per server, restart Claude Code, and you’re connected.

After that, the AI has access to everything those platforms offer. You don’t need to learn an API or read documentation.

You just tell the AI what you want, and it figures out which endpoints to call.

Step 2: Define Your Screening Criteria

This is where the real value lives. The screening criteria are yours to define.

Here are the filters I personally use. You can keep mine or swap in your own thresholds.

The hard filters that eliminate candidates:

Vol/OI above 2x means today’s volume exceeds existing open interest. Someone is opening new positions, not managing old ones. Above 5x I flag as a strong signal.
Premium above $250K filters out retail noise. Above $1M I call it a “golden sweep.”
DTE between 7 and 21 days is my sweet spot. Short enough that the buyer expects a near-term move. Long enough that it’s not a zero-day gamble.
Ask-side execution means the buyer paid the offer price to get filled fast. If they’re willing to cross the spread, they’re in a hurry.

False-positive exclusions:

Just as important as the filters themselves. A lot of what looks “unusual” on a flow screen is actually routine institutional behavior.

I built exclusion rules for earnings-week gamma hedging, married puts, index arbitrage spillover, dividend capture strategies, roll activity, and meme squeeze mechanics. Without these, the screener would flag dozens of trades that aren’t actually informational.

For example, if a stock has earnings in 3 days and you see heavy put volume at that week’s expiry, that’s almost certainly gamma hedging, not a directional bet. The screener strips it out automatically.

Step 3: Build the Conviction Scoring Rubric

After screening, every surviving candidate gets scored 0 to 10.

The weights reflect what I think actually predicts returns. Not just what looks impressive on a flow screen.

The highest-weighted signals (+2 points each):

An insider trade in the same direction within 30 days
A congressional trade in the same direction within 60 days
The same ticker appearing on consecutive scan days (repeat activity)
Vol/OI above 5x or premium above $1M

Moderate signals (+1 point each):

Sweep execution, ask-side fills, short DTE, deep OTM, floor trades, and confirmed opening positions.

Deductions (-1 to -2 points):

Earnings proximity, macro event alignment, bid-side execution, and meme stock flags all subtract points.

I weight insider and congressional corroboration highest because they represent independent confirmation from people with potential non-public information. When an insider buys stock and then institutional LEAPS sweeps follow a few weeks later, that convergence is what the screener is designed to find.

Step 4: Tell Claude Code to Build It

Here’s where it actually gets put into action.

You describe this entire workflow to Claude Code in plain English. The screening funnel, the false-positive exclusions, the scoring rubric, the output format, everything.

Claude Code creates what’s called a “skill.” A skill is a reusable tool you can invoke anytime with a single command.

Mine is called /options-flow. I type that, and the full pipeline runs.

The skill also saves every scan result to a file on disk.

Over time, this builds a proprietary dataset of what the screener flagged, when, and at what conviction level. You can backtest that dataset and refine your criteria based on real outcomes.

This is what I call the quantimental approach. Fundamental catalysts scored by quantitative signals, refined by a growing dataset.

The Result

I ran my screener on April 10.

Here’s what it found:

Top Play: $PANW

The screener flagged $PANW with the highest possible score.

CEO Nikesh Arora had bought ~68k shares at $146.87 on March 27. That’s roughly $10M in open-market purchases.

Rep. Gilbert Cisneros bought PANW stock on March 13.

Then $3.1M in LEAPS call sweeps targeting the $175 strike (March 2027 expiry) also hit the tape. All at the ask, across multiple exchanges.

Three independent data streams pointing the same direction. An insider purchase + a congressional trade + institutional options flow, all on the same stock within a 4-week window.

Bearish Flag: $MU

The screener also works in the bearish direction.

$MU scored 9 out of 10.

EVP Sumit Sadana sold 24k shares at $421 that same day. Discretionary, not 10b5-1.

Put open interest had been building for 9 consecutive days on the $345 strike. Over $4M in put premium across multiple near-term strikes.

Earnings aren’t until June 24. This isn’t routine earnings hedging.

Interestingly, the bullish headline scraper found on Micron that day was that it had passed Nvidia’s HBM3E qualification test. The options flow was going the opposite direction.

The Bigger Picture

This entire build, the screening funnel, the scoring rubric, the false-positive exclusions, the historical tracking, was created by describing what I wanted in plain English to Claude Code.

If you can describe what you want a screener to do, you can build it. That’s the real takeaway.

Hopefully this week's newsletter shows what you can do with Claude Code and why I'm so excited about it!