FWIW Docs sources are already available, e.g.:
https://docs.djangoproject.com/en/6.1/_sources/ref/models/fields.txt
I’m not sure LLMs really mind the rst vs md difference.
As I’m also developing a Django skill for setting up new projects—Django Seedkit—I can share some experience from working on it.
The biggest challenge is determining whether a skill is actually needed. For example, LLMs are already quite good at template inheritance and choosing the correct template directory. Even when they make a mistake, they usually spot and fix it quickly.
The biggest gap is in newer framework features that are missing from the LLM’s training data. Django template partials are one example: you need to mention them explicitly if you want the LLM to use them.
So I recommend setting up baseline tests that generate code without the skill. Then include only the information needed to address the failures identified by those tests.
That makes a lot of sense. Baseline tests seem like a good way to find where the LLM actually needs help instead of adding unnecessary instructions. Focusing the skill on newer Django features that aren’t well represented in the training data sounds especially useful.
I think the idea of keeping the skill lightweight is important. The discussion makes a good distinction between repeating framework documentation and capturing the community’s preferred patterns, common Django footguns, and guidance that may otherwise be scattered across different sources.