syntax cheatsheet

[dataf3l]

if https://ryelang.org/blog/posts/programming-language-in-age-of-llms/ then let this project build a document of 2000 tokens with all the syntax and what makes the language different, we can put at the beginning of each prompt for the llm so that the LLM can be effective, it won't replace the millions of lines of python but it will help get the syntax right so that the llm can do the job of helping build a standard library, let the library be inspired by another language so libraries can be ported, and let an unattended llm agent port the most important libraries, in the space of the web, we need http clients, servers and db stuff, disk access and of course all the crypto stuff, let the llm write the tests, these are just suggestions, but a short file with the syntax can help catapult the language forward, maybe adding that, plus instructions, can help newcomers, SO THAT the language can compete with the other languages, if it is not made competitive, it is hopeless, have fun in the journey.

[refaktor]

I can experiment with such file and adding it to the prompth and seeing the resulting code. I've seen that LLMs can generate Rye code (giving them examples), but some language specifics (like mandatory spacing between tokens), op/pipe words, returning functions, injected blocks do cause it trouble. I will see if I provide an well prepared explanation and maybe an example, if it will help.

Thanks!

[haydenflinner]

+1 to this, I don't often write code char by char anymore, instead throwing a bunch of prose at an LLM and letting it write the syntax. This works fine in Rust and JS but I have had trouble with it in Rye, haven't yet found the right one-page doc to guide it. For Rescript I had no such issue since they have a 1-page overview.

[refaktor]

I will prepare such 1 page overview. Rye has a slight problem with LLM-s because some of it's behaviour is quite new, but I've gotten OK results in some cases. What LLM are you using?

[dataf3l]

@refaktor I use AiStudio, at the time of the writing, GEMINI PRO 2.5 and I also use OpenAI's GPT family of models, specifically the free one. Others I have used and have found useful are deepseek, qwen, glm, grok, among others, but I'd say it doesn't matter that much, what matters is that the document is short, contains the full syntax (similar to learn X in Y seconds) and also has examples for input, output, byte manipulation with explicit semantics, string, numbers, dates, etc. this is possibly something the LLM can construct itself, but I say let's help it.
In 5 days, we'll see gemini 3, which will be 5% smarter, and none of this will matter, but this matters now, and we should refrain from making assumptions about the future performance of autoregressive models, until evidence is presented.

if we can make gpt 3.5 behave at gpt 4 levels by providing context, this will help models like github's copilot if there is a pre-comment at the beginning of the file with the syntax, also, it may be worth exploring making a openai bnf syntax validator, so it can act as an oracle as well like this or possibly using openai's CFG checker idk just throwing some ideas out there to see what sticks. I think every newlang needs this

[refaktor]

I am thinking that such document would need a section about syntax rules and specifics and a section about Rye's common functions and patterns. I assume for both providing multiple examples of each syntax mechanism and for each function as well as short explanation. We could generate the second as there is already system in place for function reference based on code comments that includes description, arguments description and unit tests / examples. For the first one, I would have to try to create a document that explains all syntax elements and mechanisms in one place.

I will start making it and we can see how it works and improve.

[refaktor]

I am working on this , and I've also made all tests pass after changes to the language so autogenerated function reference could be made.

How large should such document be to be useful?

[haydenflinner]

Keeping the doc short doesn't matter too much, it's best if it comprehensively covers gotchas and surprises for, say, a Python expert switching over or a Javascript expert switching over. Cookbook-style examples for common patterns or idioms are great. It's even better if there's a listing of every built in function along with brief doc or example usage of each. If it's too long for a particular AI we can always summarize it, but it's super nice to get the one-link complete reference. Lots of the AIs support linking to this as a sort of always-present background reference, to be looked through for info whenever a question is asked of it.

[dataf3l]

if the file is over 500 lines i will run into issues