AI Workflow
Front-end code in hours instead of a week, with a human review gate.
The Challenge
As part of our major report builder project, we needed a facade for every business class. A facade is basically the menu for a report: the fields, filters, and descriptions people see when they build one. Someone in Payables might need Vendor Name, Invoice Number, and Posting Date, plus a filter that only shows invoices on hold, and each of those needs a description that tells them what they're getting. These had to be written from scratch, and our developers were stretched thin. So the job landed on my team and me, with about 6 months until we shipped.
My Approach
Each business class came with two files written in our proprietary Landmark Programming Language: UI code and business-logic code. My starting point was pretty straightforward. If something showed up in the interface, it was probably important for a report. So I pointed the model at the UI code first and had it use the business logic to fill in the details.
The Guardrails
The first output was solid, and it kept getting better as I refined the steering file. I added two checks along the way. Before adding a field, the workflow confirmed it was actually used in customers' tenants, so nothing made it in just because it existed in the code. Then it checked each facade against a persona I'd built from the business-logic files, to flag fields that person would never need.
The Human in the Loop
A model checking its own work is a bit like grading your own homework, so it never had the last word. I kept final say on every field name and every description. Engineering then tested the generated files before anything shipped. The most common issue I ran into was poor formatting that resulted in errors during import. This was easy enough to continue to add to the steering file, ensuring that the model fully adopted the patterns of a known working facade file. Selections of labels and descriptions were vetted with product owners who knew what made sense and what might not.
The Results
By the end, the workflow had generated descriptions for over 100 fields across Payables and Projects. A manual process that normally took a week came down to a few hours, and the descriptions read consistently across dozens of separate business classes. It was also a low-stakes way to prove LLMs could handle domain-specific writing, as long as someone who knows the users is checking the output.