Add much shorter cheatsheets to Docs for LLMs

[einarpersson]

Describe the problem

One pain point that many Svelte users experience today (I imagine) is the fact that current llm models have a limited ability to produce accurate Svelte 5 code. This is of course expected to get better over time but it remains a problem for the near/medium term.

One suggested and implemented solution for this is the docs for llms.

However, even the small size is actually quite big, too big to be of any practical use for me at least. Perhaps it works for youtubers to show-off cool demos, but how would you use THAT embedded in your system prompt without seeing your token usage go through the roof?

It also contains a lot of information that the llm probably already has captured in its training data, for example what Svelte is and how Svelte relates to Sveltekit.

I would like something much shorter and focused that I can use in my daily work.

Describe the proposed solution

I propose a new llm txt file that is more of a Svelte 5 cheatsheet than anything. It should only contain the bits that are expected to present problems for llms due to being newly added.

It should laser focused and be clean of any redundant statements like Components are the building blocks of Svelte applications. (taken from the current llm small docs)

I suppose there already are these cheatsheets (and I can of course create one myself) but it would be nice to have a official and up-to-date source for this.

One alternative intepretation of this request (perhaps better now that I think of it) is to create a kind of llm-recent-changes.txt that is a cheatsheet variant of a release log, not only specific to Svelte 5. The same thought model could be used for handling recent changes in Sveltekit also (it is also evolving!).

Importance

would make my life easier

[Thiagolino8]

The files are divided into sections, you can delete the sections that are not relevant to your project

[dummdidumm]

I also think a much shorter, focused llms-small is necessary. We probably have to write it by hand.

[einarpersson]

I also think a much shorter, focused llms-small is necessary. We probably have to write it by hand.

👍 I'd also encourage to think about the fact that the problem is not Svelte itself, but the changing nature of Svelte. And this of course is true for all software under active development. Therefore, perhaps the more precise solution would be something along the lines I suggested last in my post i.e. having a svelte-llm-recent-changes.txt or something like that. It is about closing the gap between what is expected to already be in the model (training data) and the latest version.

A human written / curated to-the-point svelte-llm-cheatsheet.txt would be valuable right now, but I just want to emphasize that the <training data> --- <latest version>-gap is a general problem (which also affects sveltekit).