An iPhone pointed at a computer monitor. The person who does the work, walking us through it and explaining it in her own words.
That is the part of our Dive Bar Restaurant project I keep coming back to.
We were building automation for real back-office work. But the most important input wasn't a technical document. It was being able to see the process and hear the reasoning behind it.
She showed us how the work actually happens
When someone has done a job for years, a lot of the knowledge never makes it into a manual.
They know which report to open. Which result to compare. Which difference deserves a second look. They move between screens almost without thinking about it.
Ask them to write a software specification and you're giving them another job. Ask them to show you their work and explain it as they go, and you get something much closer to the truth.
We used an iPhone to record the screen while the manager walked us through her process. Her narration mattered as much as the pictures: it explained what she was doing and why.
Those recordings became a practical starting point for building two workflows: payroll preparation and the morning bank reconciliation check.
From her explanation to working software
Working with Claude, we used the walkthroughs to map the steps, understand the checks, and build automation around the existing process.
For payroll preparation, that meant understanding how hours, overtime, and tips come together—and how the prepared information should compare with the source reports. The walkthrough also exposed the checks a person makes before payroll is ready for submission.
For the morning bank check, it meant organizing the deposits, payments, cleared checks, and balance into a useful review for the person doing the reconciliation.
These became two workflows within Dewey, the restaurant's robot assistant. They were built around the manager's knowledge, not a generic guess about how a restaurant ought to work.
We could follow the process from its inputs to the output she needed, then compare what the automation produced with the work it was meant to support.
The results were remarkably close
The first real bank file matched exactly. Payroll preparation also validated closely against the source reports, and the remaining differences were investigated rather than treated as good enough.
That was the impressive part: how faithfully we could translate a real person's working process into useful automation from a narrated walkthrough.
Payroll preparation was being validated, with payroll-system entry still a separate next step. The bank check supported the manager's reconciliation work.
We built, tested, compared, and refined. The recordings gave us the right process to build from.
The expertise was already there
The manager didn't need to learn to write code. She brought something more important to this project: an understanding of the work.
Our job was to capture that understanding and turn the repeatable parts into software, with clear checks and human decisions where they belong.
There is a larger technology story here too. The restaurant uses a mixture of modern systems and established desktop software. We didn't need to replace all of it before making the work between those systems easier.
But the connection I find most interesting isn't between two applications. It's between someone's experience and the software we can now build from it.
Start by showing the work
If you're wondering where AI might help in your business, start with a recurring task that someone knows inside out.
Have them show the process and explain the decisions. Capture it with permission, protect the private information on screen, and use the walkthrough to understand what a successful result looks like.
Then build against that reality and check the output.
An iPhone captured the process. Her voice supplied the context. Working with Claude, we turned that knowledge into useful automation.
The expertise was already in the business. We found a better way to build around it.
Read the full Dive Bar Restaurant case study: Dive Bar Restaurant case study