OpenAI just published a case study on Balyasny Asset Management.

Balyasny runs $32 billion across roughly 180 investment teams. They built an internal AI agent called the Merger Arbitrage Superforecaster

Here’s what it does: it monitors every regulatory filing, press release, and policy update on the deals their merger arb desk is tracking. Then it automatically updates the probability that each deal closes.

Before this agent, their analysts were doing it manually....

Tracking 10 to 30 active deals across SEC filings, FTC statements, press releases, shareholder votes, financing conditions, and activist investor positions. Updating a spreadsheet by hand. Every day.

That workflow is dead.

The agent replaced it with real-time probabilistic monitoring + structured outputs + traceable reasoning.

I wanted to reverse-engineer this and give you

For the less technical readers: The DIY version you can start automation ASAP using just 1 prompt
For the more technical readers: The roadmap if you're comfortable with APIs and Claude Code to take this workflow to take it to the next level ... and engineer this system to institutional grade quality.

Here’s what I built for you:

Designed a 6-factor deal scoring framework from merger arb practitioner literature
Wrote a single system prompt that acts as an autonomous deal monitor
Tested it on Manus so anyone can run it without writing code
Mapped out how to upgrade it to a production system

The Prompt

The full prompt is long, so I'm linking the PDF below. I'm running this on Manus which you can also set "Scheduled Tasks" to make this recurring (e.g., get a daily report on this merger).

Use this as the foundation prompt to build into your own framework.

If there are parts of this you want to customize, adjust the factor weights or scoring ranges for your own style.

Prompt: merger-arb-superforecaster-prompt.pdf

The Result

Full output here: Link

I tested this on the Union Pacific / Norfolk Southern mega-merger. An $85 billion railroad deal with real regulatory uncertainty.

Here’s what the agent came back with.

The factor scoring table is where the value is. You can see exactly which risks are dragging the probability down, with a specific source and date for each score. That’s the “traceable reasoning” piece that Balyasny’s system emphasizes.

I want to note this is a simplified version. The agent is using web search for its data, which means it’s working with whatever it can find publicly. To make it more systematic you should connect the right MCP servers (see more below)

The bigger picture: merger arb has always been about information processing speed. Who reads the filing first. Who catches the FTC signal first. Who updates their model first.

This prompt compresses that workflow from hours to minutes.

How to Take This to the Next Level...

What I showed you is the accessible version. Paste a prompt into Manus, give it your deals, get a report.

But if you want the real thing, here’s how I’d turn this into a production system:

Connect to SEC EDGAR directly. Using Claude Code with an EDGAR MCP server, the agent pulls filings in structured format. (Benefit: No web scraping. No wasted context window on HTML noise.)
Add options flow data. Services like Massive.com provide real-time options flow MCPs. Unusual put activity on the target, or heavy call buying, gives you a signal the market isn’t showing in the spread alone.
Schedule it as a recurring task. Claude Code lets you set up cron jobs that run the agent every weekday morning before market open. Your deal monitor updates automatically without you touching it.
Diff against yesterday’s run. Store each day’s probability snapshot. The agent compares today’s scores to yesterday’s and flags any deal where probability moved more than 5 percentage points. You only look at what changed.
Package it as a skills zip. If you’re deep in Claude Code, you can bundle this into a portable skill with the system prompt, a deal config file, and a report template. Share it with your team or run it across multiple portfolios.
Multi-agent architecture. Split the work across specialized agents. One agent monitors filings. Another tracks regulatory news. A third scores and synthesizes. Each agent is scoped to its data source, which means better accuracy and cleaner reasoning chains.

A benefit of the above architecture in Claude Code is you can go way deeper on customizing to how you frame your analyses.

Try this prompt on a deal you’re tracking and see where the agent’s probability diverges from the spread. That’s where the edge is.

Feel free to hit the reply button and let me know what you find.

P.S. If you're keen on getting started learning Claude Code for Investing use cases, I'm launching a 1 week bootcamp on this very soon. Sign up to my waitlist to get details as soon as it drops: Link

Personal

Quick thought on something I’ve been thinking about.

Open source AI is going to matter way more than most people realize.

Right now the frontier labs are burning tens of billions to win the AI race. That money has to come back somehow. And once these companies capture enough market share, the models will get more restricted, not less.

We’re already seeing it. Anthropic just cut off third-party harnesses like OpenClaw from using Claude...

The pricing and access you have today is subsidized. It won’t last forever.

(Remember the good ol' days of $5 Ubers in 2014?)

Found this old pic of me in my first Uber ride lol ... I miss zero interest rate days

The more you can run locally, the less you depend on a company’s willingness to keep the lights on for you at the current price.

There’s another reason too that I've been thinking about beyond finance applications of LLMs...

If something ever cuts us off from the internet (eg a military conflict, infrastructure attack, or something else), a local LLM running on an off-grid solar and battery setup gives you unlimited access to medical information, survival knowledge, and general reasoning.

I wouldn’t be surprised if the guys building bunkers in New Zealand / Hawaii already have this exact setup running.

Anyway. Something to think about.

Get one high impact AI prompt every week