Building an advanced stock screener with AI
Claude Code = magic
Hunting for interesting stock ideas is a lot of fun.
Your typical screener uses financial criteria, or you could even just type what you want as a prompt these days.
But now you’re leaving money on the table.
You’re not really using the power of these LLMs.
What we’ll show you today is a tool we built called Hunter. Because of the massive drawdown at many software companies due to the rise of AI and the development of tools like Claude Code or Claude Cowork, I set out to build a specialized software screener to analyze these companies and assess how AI-resilient they really are.
The software screener
It looks like this:
The inputs are 10-Ks and 10-Qs to provide context on the different companies and financial data. The brain is a Python pipeline that fetches data and uses Claude Sonnet to apply a scoring framework to the companies. The final output is an Excel or CSV file with a list of companies.
The AI resilience scoring output looks like this:
And each score has an analysis behind it.
Here’s a glimpse of Duolingo:
Now, the result I’ve shown here was an intermediate result while we were iterating. The final result is a lot better, thanks to Einstein’s help.
Everything should be made as simple as possible, but not simpler
- Albert Einstein
In the past, you could only screen on financial data.
Now, we can screen on text. But as we said before, AI is bad at judgment. We need to help it. That’s why you need a framework. Some kind of scoring system.
The cooking recipe
This build is like cooking. You need ingredients and a recipe. But first, what did we use?
I built this in Claude Code. And I wish I could build it in Claude Cowork, but that’s not possible. I will be able to provide it as a tool you can use in Claude Cowork.
The ingredients
A scoring framework. We are using a specific AI resilience framework to score software companies. But you can imagine other use cases. For example, you could build a framework on serial acquirers or on turnarounds. Be as detailed and specific as possible, and make sure the LLM can score a company based on text. The scoring framework is stored as a markdown file.
Financial data. If you don’t have a database API, you can always use yfinance. It’s an unofficial financial data API that scrapes data from Yahoo Finance. It’s great for personal use, but be mindful that it can always stop working tomorrow.
A list of tickers: You need a list of companies you want to evaluate. I pulled them from my financial database. You can always just pull a couple of companies from your portfolio and add them.
The recipe
Now we need to bring all the ingredients together in a certain way. Our output is an Excel file that scores each company and lists some financial data.
The workflow looks like this:
Ingestion layer: Retrieve the data from the EDGAR filings and the financial database
Cleaning layer: Process the data so it’s more suitable for an LLM to process
Processing layer: Parse the data and run the scoring framework
Output layer: Output a CSV and Excel file
One thing to note is that this requires iteration. When you build this, it won’t be perfect from the start.
And I encourage you to just start building something small like this, testing it on a couple of companies, and keep building. You’ll be surprised and amazed by the final outcome.
Here’s the simplified version:
Using Claude Code
So how can you build this yourself?
When you use Claude Code, always start with your end goal. What do you want to achieve? Be as precise as possible.
I asked Claude the following:
I want to build a tool that uses financial and text data to analyze software companies, score them on AI resilience, and output CSV and Excel files with a row for each company, including its score and selected financial metrics.
And then ask Claude:
Ask me questions? What additional context do you need to achieve this goal?
Claude will start the interview. The goal is to develop a plan. Once you and Claude are happy with the level of detail in the discussion, you say:
Take all the information from our brainstorming sessions and show me a step-by-step plan that I can follow to build this.
Claude will go into planning mode and show you the plan.
Review the plan, make small modifications if needed, and then just start building from step 1.
Now there might be 20 steps. And once you start to test, you’ll run into edge cases that need solving. I’m not saying this is a breeze. But you can unlock dev powers for yourself.
I’m not a software developer. I’m not really that tech-savvy.
I consider myself a power user with endless curiosity.
If I can build this, you can too.
I’ll see if I can add a detailed video on our course to show you how to build this from start to finish.
Next steps
We’ll run Hunter through about 100 to 200 US Software companies, and I’ll provide you with the final results.
And my aim is for Hunter to become more flexible. To allow you to select frameworks and use them on certain companies, and see how they score.
If you want to learn Claude Code, how to build this, or just download the entire project and run it on your own computer, you can join the private beta for our course and skill library with the button below.
Note: I want to emphasize that the private beta is not free. You’ll get access to a platform with prompts, skills, videos, and a software tool called Flow to track your decision-making. You’ll be able to enter at a significant discount off the final release price, but in return, I will need your feedback while you're “playing” with the product.
Thanks again, and keep learning AI, keep building.
Kevin