Claude Artifact for Underwriting: Moving a Lender From Chat Prompts to 10 Skill Files

When the same deal can score two different ways depending on how a chat was started, you do not have an underwriting system. You have a conversation. That was the problem at a private lender whose underwriting ran as ad-hoc chat prompts that behaved like generic Claude.

The client is an executive at a private real estate lender. We had already built a set of custom underwriting skills for the lender, but the day-to-day workflow was not using them. If you run lending for investors, our page for mortgage loan officers and private lenders covers the wider work.

The situation

The underwriting build is made of markdown skill files. Each one holds a piece of the lender's scoring logic. For the build to matter, every run has to load all of those files before it says anything about a deal.

In practice, the executive was working in plain chat windows. A chat that never loads the skill files answers from general knowledge. It sounds confident and it ignores the lender's custom scoring, which is why results were inconsistent.

What was broken

Two gaps showed up on the monthly strategy call.

• The chat-based workflow behaved like generic Claude. It did not use the custom underwriting build, so scoring varied from run to run.
• The inputs for a run were not fixed in the workflow. The executive uses the same core inputs on every deal, but nothing required them.

What we built

We moved the workflow out of chat and into a dedicated Claude Artifact, then pinned down what goes in and what runs on a schedule.

• One Artifact, 10 skill files. All 10 markdown skill files are attached to the Artifact, so every run pulls the same scoring logic. We sent them by email for the executive to drag in.
• Three standard inputs. The executive confirmed the 3 core inputs used on every deal, and we standardized them as the required per-deal inputs for each underwriting run.
• Scheduled task files for comps. A second email carried the scheduled task files for the weekly market comp refresh. Those tasks run only through Claude Desktop with Cowork, so the plan included installing it on the executive's PC and enabling the virtualization features it needs. The files were set to be uploaded at a follow-up session after the machine restarted.
• A bulk validation run. We scheduled a run across multiple deals for the next Monday, to confirm accuracy and that the comps are dated correctly.

What changed

The call did not report accuracy results, and the validation run had not happened yet. What changed is how the work is structured. A deal now enters through the same three inputs, and every run loads the same 10 skill files. Two people running the same deal should get the same scoring path, because the logic no longer depends on how a chat opened.

That is a design improvement, not a proven accuracy gain. The bulk run on multiple deals is the test that tells us whether the outputs are right.

Check whether your AI tool is really using your build

You can test this on your own setup in a few minutes. Run one deal twice, in two fresh sessions, and compare the scoring. If the results differ in ways your rules would not allow, the tool is not loading your logic every time.

Next, ask the tool which files or rules it used. A tool that is loading your skill files can point to them. Finally, write down the few inputs you use on every deal. If you cannot name them, the tool cannot be consistent either.

Frequently asked questions

What is a Claude Artifact for underwriting?

It is a dedicated workspace where the lender's skill files are attached, so each run loads the same scoring logic. In this build, 10 markdown skill files sit inside one Artifact, which replaced loose chat prompts that behaved like generic Claude.

Why do chat prompts give inconsistent underwriting results?

A chat only knows what you put in it. If it does not load your skill files, it answers from general knowledge and ignores your scoring. The same deal can then score differently depending on the wording and attachments of that one conversation.

Why does the weekly comp refresh need Claude Desktop?

On this project the scheduled tasks for the weekly market comp refresh could only run through Claude Desktop with Cowork. That meant installing it on the executive's PC and enabling the virtualization features it relies on before the task files could be uploaded.

What are the standard inputs for each deal?

There were 3 core inputs the executive uses on every deal, and we standardized them for each underwriting run. The exact list is specific to the lender. The principle is to fix the inputs so two runs on the same deal start from the same facts.

If your AI underwriting still lives in chat windows, book a 30-minute AI follow-up and reporting audit. We review how deals enter your process and where your scoring rules should be pinned.