Claude Cowork was released in January and Microsoft announced Copilot Cowork, built on the same technology as Claude Cowork, earlier this month.
But what is Cowork and why should we care?
Cowork is effectively the business version of Claude Code, the beloved agentic coding platform used by almost every software developer at this point. Claude Code is powerful because it can read an entire codebase (all files, not just one), understand relationships between files, use tools like web search and browsers, connect to third-party data, create and edit files, deploy multiple agents in parallel, create and run tests to verify work, and use a new concept called Skills to enforce work is done how we want it done.
If you apply these capabilities to business tasks like research, data analysis, model creation, report writing, etc., it becomes clear that we can use Cowork to do complex, multi-step work.
We can test this out by building an end-to-end real estate underwriting workflow. But first, let’s talk about Skills a bit more.
‘Skills’ are a critical component of Cowork. You can think of Skills as a playbook or process manual for AI. They capture domain expertise, corporate processes, workflows, specific sources for research, how to use an API, etc. Skills are primarily markdown files (glorified text files) but they can also include python scripts (for data processing and API calls) and other reference material. Skills are also ‘composable’, meaning they can work together and reference one another.
When I look to automate a workflow using Cowork, I typically break the process down into its individual components and create specific skills for each of those components. I like to think of each skill as instructions for an intern. I give them a set of instructions and then they go off and perform that task exactly as described.
If we were doing underwriting for a real estate transaction, we might have skills for the following: property overview, market research, rent comps, sales comps, physical / environmental risk assessment, rent roll / tenant analysis, discounted cash flow creation, investment memo structure, etc.
Skills are created simply by prompting. Cowork has a skill to create skills. Simply describe your process or task and watch it work. When it’s done you will see the markdown file and any python scripts that may also have been created (depending on what you’re doing). Review and test the skill. If you have any changes simply prompt the changes.
Let’s create a simple skill to build a property overview. The skill will use any files provided by the user but it will also perform a general web search.
Here’s a snapshot of our finished skill. In this one it included a section for market context. I’d rather have market research be its own skill. I’m going to prompt Claude to remove that section from the skill.
Let’s test it. I had Claude test it on a random office property in Washington DC and it chose 900 K Street NW.
Here’s the investment highlights section.
Let’s create another, more complicated skill. What real estate investment memo doesn’t have an analysis showing demographics for 1-, 3-, and 5-mile radius around the subject? We can create a skill called census-demographics that leverages the Census API to do this analysis for us.
This skill looks different. You will notice it has a folder called scripts on the left and includes a python file called ‘census_demographics.py’. This python script geocodes our property address, connects to the Census API, pulls the data we need, and does all the analytics.
Let’s test it on 900 K Street NW. Here is the output.
In addition to these skills I created the following:
Industrial-dcf with instructions on how to populate an industrial DCF model.
Rent-roll-tenant-analysis with instructions on how to process and analyze the rent roll.
Underwriting-memo with instructions on how to construct the underwriting memo.
For a full memo I would include additional skills but I’m keeping it simple for the sake of this post. Let’s give it a test.
Make sure you have all your source files including OMs, operating statements, models, etc. in a folder. When you open Cowork you will click ‘Work in a folder’ and select the folder that holds your files. Choose Allow or Always Allow on the pop-up depending on your preference.
Then enter your prompt to kick-off the process.
When Cowork kicks off it does a few things. First, it identifies the skills it needs execute the task at hand. I have many skills in my library but it intelligently selected only the few we need for this underwriting workflow.
Next, it breaks the workflow down into discrete tasks and starts executing them. It checks them off as it completes each task. Tasks without dependencies run in parallel. Those that depend on others wait their turn.
You can see its progress along the way.
When it’s finished it produces two outputs: the completed DCF model and the underwriting memo – both ready for human review.
Now, you may be wondering why I gave it a DCF model to populate rather than build it itself. There are several reasons for this. First, Claude Cowork and other AI Chatbots are not as good at producing spreadsheets as their Excel plugin counterparts are. The DCF model I used was created by Claude and ChatGPT Excel plugins which are both very strong. Second, by using a template there will be consistency in how the modeling is performed each time.
To help enforce this consistency I include a tab in my Excel files with LLM instructions. These instructions tell the model what it can and cannot do (i.e., what cells it can / should change and which it should never modify). Third, building spreadsheets consumes a lot of tokens and by using a template we can reduce the amount we consume.
Here is a view of the Executive Summary from the memo. I didn’t provide an OM as input so no photos were included. If you get the Claude in Chrome add-in though you could instruct Cowork to use Chrome to find and download images for your memo.
You can find the whole document here: West Fargo Underwriting Memo
And here is some of the DCF model. You can find the whole spreadsheet here: West Fargo DCF Model
This is just one example of a complex workflow that can be done with Cowork technology. Other real estate examples include: loan underwriting, loan rating, market research, investor reporting, data analysis, due diligence, accounting reconciliations, CAM reconciliations, etc.
The power of Cowork technology is undeniable. There are a ton of applications for it. You are only limited by your creativity.
Happy building!
-Prof. Scott