Remix.run Logo
GuB-42 5 hours ago

AI works well for one kind of documentation.

The kind of documentation no one reads, that is just here to please some manager, or meet some compliance requirement. These are, unfortunately, the most common kind I see, by volume. Usually, they are named something like QQF-FFT-44388-IssueD.doc and they are completely outdated with regard to the thing they document despite having seen several revisions, as evidenced by the inconsistent style.

Common features are:

- A glossary that describe terms that don't need describing, such as CPU or RAM, but not ambiguous and domain-specific terms, of which there are many

- References to documents you don't have access to

- UML diagrams, not matching the code of course

- Signatures by people who left the project long ago and are nowhere to be seen

- A bunch of screenshots, all with different UIs taken at different stages of development, would be of great value to archeologists

- Wildly inconsistent formatting, some people realize that Word has styles and can generate a table of contents, others don't, and few care

Of course, no one reads them, besides maybe a depressive QA manager.

asah 5 hours ago | parent [-]

not true! it's read by other LLMs! /s

pgwhalen 3 hours ago | parent [-]

Not sure why the /s here, it feels like documentation being read by LLMs is an important part of AI assisted dev, and it's entirely valid for that documentation to be in part generated by the LLM too.