Ok, bye bye

> In a 100% AI-generated project, the preferred form of the work for making modifications to it is tweaking the AI prompt.

Not really. The prompt is not what an AI uses to generate code, it is its context - of which the prompt(s) are just a part. Potentially a small one.
The context also contains documentation the agent read, specifications you provided, web search results, sample code you pointed it to, whatever tool calls your agent did, ... and most importantly: also the results of the AIs internal reasoning, which are in turn based on the previous context.

In other words: the pure prompt is essentially worthless for any non-trivial software.

To give you a recent example: I wrote a prompt that said "reimplent the ancient, unmaintainable Java micro-service at ../here_be_dragons in Perl Mojolicious, drop feature A and B, only keep C and D, preserve the exact API and get the config from a YAML file instead of properties and a database; see dump.sql". The prompt was maybe 1% of the context the AI used to generate the code, the rest was the old code, including its tests and the DB dump.

For more serious development, you'd typically use something like "spec-driven development" (not to be confused with the pre-AI thing of exactly the same name). This has a set of rules the AI follows and you use this to create a specification, in multiple steps and iterations. In the end, you get a list of requirements, rationale and tasks that contain code generation and testing; usually in the form of a large tree of markdown files.

This can be committed, and iterated on, on its own, but typically it is committed with any code that was already created. It is simply far to expensive and slow to regenerate the code from spec, and even then it is not deterministic; especially if the spec was not clean (i.e. there were additional, significant prompts that are not part of it) or if the model you use to recreate it is significantly weaker. This arguably is the preferred form, and it works exceptionally well even for large codebases (but there still is a limit where it gets tricky and/or too expensive; figuring out how to push this is ongoing work).