Using AI in Conjunction With Business Rules

[gordon]

For those who don't know it, Luis, Mikhail, Chris and I have been quietly upgrading Business Rules. While we are not ready to announce a new release yet, I figured it would be useful to begin a discussion of how we can apply AI to the Business Rules compiler.

To start with let me define some terms in relation to AI:

• Model/LLM - A computer program that thinks and communicates in natural human language

• Coding Assistant - An application that utilizes an LLM to assist with programming

• MCP interface - Model Context Protocol - an extensible protocol that lets models and assistants discus complex topics

• MCP server - A server that exposes a model's capability through the MCP API

• Agent - A program that assist with AI prompting, typically aware of local applications and toolsets and with a persistent memory

In short, we are contemplating establishing an MCP server for the BR language.Nprmally such an endeavor would require a year or more of programming. But with AI we may be able to dramatically reduce the development time.

[snafoo972]

Thank you all for sharing your experience. Cursor was a real game changer for me. Microsoft had unfortunately restricted its use with .Net. I also use Claude Code a lot now and its terminal based and works more as an independent assistant. My process often looks like:

I'll often give it a few documents of context before giving it a task like

MyBrLanguageDefinition.md
SystemOverview.md
CurrentProjectGuidlines.md

then ask it to make a plan to complete a task like adding a feature.
We have a short conversation to refine the plan.
Then I give it permission to complete the plan.
It goes to work for a few minutes . I review it's work and and suggest changes or ask for explanation.
Continue this process until I'm satisfied with the results.
So for example I tried the example prompt:

> @notes\BR_Language_Grammar_Complete.md Please create a BR function to convert Celsius to Fahrenheit

Code: Select all

00100 ! Function: ConvertTemp
00110 ! Purpose: Convert Celsius to Fahrenheit
00120 ! Author: Claude
00130 ! Date: 2025-07-27
00140 !
00200 DEF FNFAHRENHEIT(CELSIUS)
00210     ! Formula: F = (C * 9/5) + 32
00220     LET FNFAHRENHEIT = (CELSIUS * 9 / 5) + 32
00230 FNEND
00240 !
00300 ! Example usage:
00310 ! LET TEMP_C = 25
00320 ! LET TEMP_F = FNFAHRENHEIT(TEMP_C)
00330 ! PRINT "Temperature: "; TEMP_C; "C = "; TEMP_F; "F"

[gordon]

Was the example you provided coded by Claude or by Cursor?

And have you integrated the two, or do you use them individually?

Is anyone else on this thread using AI presently (other than Luis - who uses it a lot) ?

[snafoo972]

Claude Code generated that. Its still pretty limited in BR but I'm making progress.

Claude also has a CLI that can also be scripted so you can execute prompts from a python or node script. I'll make it iterate through a list of things and execute a prompt template for each item.

There are really cool automation tools, too. Like this one called n8n. It lets you visually design automated workflows that link to the Claude API for AI tasks.

[gordon]

"Claude Code generated that."

I have the impression that Cursor is a convenience, but the real work is done best in claude. And VS Code is approaching Cursor's functionality. Cursor is a derivitive (fork) of VS Code. So I personally don't bother with Cursor.

I also find that the Anthropic PRO service is sufficient for all of the challenges I have given it so far, and that allows (practically) unlimited use for $20 per month.

I'm thinking of staying at the PRO level (which uses Sonnet) and using tokens for the occasional deep research that Opus can provide.

Cursor is nice though, and it offers connection to a variety of models (LLMs).

Luis .. Would you describe how you implemented "Ask Goldie" and what it does for your end users?

[mluchterhand]

You mention feeding in the following documents to AI:

MyBrLanguageDefinition.md
SystemOverview.md
CurrentProjectGuidlines.md

Can you provide a brief description of what each of the documents is doing?

I'm assuming one of those is providing the language syntax for BR. Did you build this yourself or is this something that is available to all of us?

[gordon]

Chis feels that the BR AI context that he has been experimenting with is not sufficient for meaningful application development. I will attempt to produce an AI context that improves on what Chris is working with. But that will be a stop-gap measure at best. What is needed is a fully BR trained MCP server that any LLM can utilize. And we are rapidly moving in that direction.

I HAVE DECIDED TO MOVE THE AI WEBINAR MEETINGS TO SATURDAY AFTERNOONS. Several people have requested that we record these sessions, and we certainly will be doing so. Based oon those requests, I have the impression that having the meetings during business hours would make attendance difficult for many, and in some cases, impossible. So Saturdays it is.

This Saturday ( August 16 - mark your calendars) we will meet at 4 PM Eastern time which is 1 PM Pacific time. Allow 30 minutes. If it appears we should meet more often than once per month, we will do so.

Here is the link to the meeting:

https://events.teams.microsoft.com/even ... 588cbe007b

You will need to register for the meeting when you use the link. Sorry for the inconvenience. Teams doesn't make it easy. They have some pretty onerous restrictions on their commercial versions. I am using that version so we can record. I am also sticking with Teams becuase I don't want people to have to download an app to join the meetings.

[snafoo972]

This approach is helpful for generating some sample code but isn't able to replicate the way AI agents are able to write fully functioning code for other languages. Here is my current BR language rules i use for BR related prompts like the one in my example.

https://gist.github.com/christopherroys ... 45af005ecd

It attempts to provide a concise description of BR language rules with syntax diagrams well known to common AI models. It's relatively small size is designed to prevent context overload which seems to 'distract' the model.

It does seem to improve the quality of responses significantly but I'm sure it can be improved so I welcome any feedback or suggestions.