Potential New Pattern

[mlevison]

I mentioned this on LinkedIn, happy to iterate on the idea.

Version control prompts and responses, with the code they generate. When I find a bug or code I didn't expect in the generated system, I want to be able to see what prompt helped generate that code.

I also like to review my prompts and responses, to see what I glean from them over time.

[lexler]

@mlevison I like it. Do you think you'd be able to do a PR on this?
We have a CONTRIBUTE.md that should make it really easy to come up with the first draft with a genie:
https://github.com/lexler/augmented-coding-patterns/blob/main/CONTRIBUTE.md

[mlevison]

I'm happy to contribute. The only challenge I see is that this idea is fragmentary for now. Particularly since theres is one obstacle that I've run into. When I store the prompts in a folder inside the repo, the model insists on reading the files in that folder as part of the project. That rapidly gets out of hand. Ideally I would like a .gitignore file for the coding LLMs.

[mlevison]

Small discovery, with ClaudeCode you can use settings.json to deny permissions to read a directory and its contents.
https://code.claude.com/docs/en/settings

So I will a common directory llm_history for all my repos. The problem is that solution won't generalize to other LLMs