I was testing ChatGPT 4.5 today and was surprised to see it can write perfectly serviceable FW/1 code. Anybody else been playing around with using models for CFML coding?

i have had good luck with claude, i use cline extension in vscode

Using copillot with gpt 4 in vscode It's helpful but you have to watch it

Nice! I normally am doing .Net, Sql, TS, etc in my day job; was surprised how good Claude Sonnet 3.7 was with CFML and FW1. Seemed to know what was up on casual inspection. Will keep playing with it. I’m using Claude Code through the command line btw. Really recommend it.

I've been using Replit to make a variety of tools that I would have otherwise had to cotract out on upwork for $700-2000 and it would take weeks. Now I can make them myself in hours. Pretty amazing what it can do, its like a JR devloper who keeps coming back to you with the latest update for inspection and feedback, but the feedback cycle happens in minutes. The only downside to Replit is it can't do CFML. I've made js-only programs and a python one

I know about Replit but haven’t used it. Have you compared it to Claude by chance?

I have not used claude but just heard about it

That’s what a large context window is for. You just send it the entire codebase instead of doing RAG, and give it the most information possible. Then you prompt it very specifically and review what it does. Claude is 200k tokens, or like 500 pages of code. Gemini is 2m tokens (!) Try getting like $5 of Claude tokens and use Claude Code and just try it… give it a very complete prompt.

so you..upload a whole directory of files?

Yes.

what aboute the database?

If you have that scripted, it knows the schema.

so you give it the schema for the DB and the whole code base, and then you can more or less ask it to generate new features in the style you currently make them?

I mean, in theory. In practice it’s more nuanced. But yes - that’s how it works. It can’t do things in “the style” of something unless you feed that thing in. Which requires RAG or large context windows.

@Shaunobi ahwat is "RAG"?

It’s where you use a vector based similarity search to find what code fragments or documents or whatever are most applicable to a given prompt, and then feed the results into the context window before your prompt to give the model more context.