Claude Code: The Easiest Way to Build Financial AI Workflows (No Coding Required)

Claude Code, when combined with Claude Skills, is one of the easiest and most powerful ways to build agentic AI workflows. Claude Code was originally built as an agentic coding tool, but despite the name, it's far more flexible and far more powerful than a traditional coding assistant.

In this article, I will show you how to use the Claude desktop application to build a Loan Manager Agent.

To begin, let’s outline what our loan manager agent needs to do. Its key responsibilities are 1) standardizing borrower financial statements into the CREFC Operating Statement Analysis Report (OSAR) format, 2) standardizing rent rolls into a consistent tabular format, and 3) identifying key changes from one period to another (if more than one rent roll exists).

We will create a corresponding subagent for each of these responsibilities.

Let’s get started. Open your Claude desktop application and click on ‘Code’ in the upper left corner next to ‘Chat’.

You’ll be taken to this screen. Make sure the dropdown menu on the right says ‘Local’. This means it is referencing your local computer and files rather than GitHub.

When you select a folder, Claude will have the ability to read, edit, create, and execute files located in that folder. When you do this you are giving Claude a lot of control over your computer so be sure you’re using folders you trust. Claude itself gives you this warning at the bottom.

It’s this same control though that makes Claude Code’s agentic abilities so powerful. I’ve given Claude Code access to a folder called loan_manager_agent that will house everything we need for our AI workflow.

To start, we instruct Claude Code to create our CLAUDE.md file. The CLAUDE.md file is a configuration file that Claude Code reads automatically at the start of every conversation and effectively serves as your instruction manual.

Claude Code will get to work creating your CLAUDE.md file.

When it’s done, you should see it in your loan_manager_agent folder.

Next, we need to create a few folders. We can either do this manually or ask Claude Code to do it for us.

I’ve dropped our OSAR template from CREFC into the Templates folder.

Claude and other LLMs are trained on a significant amount of data and have general context about almost every industry, but they don’t always have detailed domain context about an industry, a company, or a company’s specific workflows and tasks. This is where Skills come in.

Skills give Claude Code the domain specific information it needs to perform complex workflows. This means giving it specialized domain knowledge, detailed instructions on how a task or workflow should be performed, instructions on how to use tools like APIs or code, files that should be referenced (e.g. our OSAR template), and example outputs.

Skills are only called and used by Claude when they are needed, which reduces the amount of instructions and text being sent to the LLM. This prevents you from quickly hitting context window limits. We call this context management.

Skills follow a very specific format but you don’t need to know the format. Anthropic very cleverly created a skill to create skills. Let’s ask Claude Code to create a skill to standardize operating statements into the CREFC format.

Claude will ask you to approve what it’s about to do. I’m going to choose ‘Always allow for project (local)’ but you may want to choose ‘Allow once’.

It creates the skill, called SKILL.md, and places it in a subfolder of our skills folder called osar-standardization.

Note: all skills are called SKILL.md and placed in a directory with a name that describes what the skill does (see image below). That descriptive name helps Claude find it when it needs to use it. Skills are always included as a subdirectory of the .claude directory (this may differ in Claude for Enterprise).

You should review the content of the skill and see if you want to make any changes. Click ‘Allow once’ to create the skill. If you want to make changes, prompt Claude Code with the changes after it creates the file. I’m happy with it for now but we may need to change it after we start testing the agent. Agentic workflows are all about iterating and refining.

In addition to creating our skill, Claude Code created helper documents that the subagent will be able to reference if and when it needs to.

Let’s create a second skill for processing our rent rolls.

I realized while creating the rent roll skill that in the osar-standardization skill I forgot to include instructions on how to name the final file and where to place it. I updated the skill accordingly.

Now that we have our skills we can create our subagents.

The agents are created in an ‘agents’ subfolder of .claude (similar to skills).

Time to test! I placed an operating statement and three rent rolls for a fake property called ‘Park View Towers’ in the Financials and Rent Roll folders. You can download and view these here: Loan Manager Google Drive

My only command is ‘Process the Park View Tower documents’.

Claude Code creates its ToDo list (seems reasonable) and gets to work. This may take some time.

It summarizes the results when it’s done.

The final files are placed in the Output folder. You can see mine here: Output Files on Google Drive

We’re done! From here I would test different properties and property types and enhance our skills and agents (we definitely didn’t get it right the first time). I would iterate and refine until the workflow is working exactly as needed.

These are just two tasks a loan manager is responsible for. We could expand this agent to perform additional functions, such as checking for missing documents and automatically emailing the borrower, performing data validation checks and following up with questions, running period-over-period variance analysis on operating statements, uploading data to a warehouse, building a DCF based on the latest OSAR, performing credit analysis, etc.

You are mostly limited by your imagination.

Happy building!

-Prof. Scott