LVBLPractices

attempts at reverse engineering LVBL with unique prompting methodologies (and eventually, real code) to see how they have improved Claude's design output, creates integrations and optimizes the wrapper for better code generation.

i'll eventually organize this repo, but first, you should read (this, and then) dev-notes.md. those are some notes on how i figured out that this was most likely a wrapper, not a fine tuned, specialized or truly trained Claude (or its own LLM).

then, read design-rules.md, additional-design.md following that and then final-design-output.md.

i am determined to fully reverse engineer and get down to the bottom of this service. after further cracking design and replicating it in bolt.diy, i have a few more fixes and integrations in store for that open source project. but by the end of it, i hope to create the best no-code ai software generator on earth.

To Do

• finish cracking the design wrapper and replicate it in cursor, windsurf or my own environment with
• compile my research on these types of wrappers and dump it here.
• take apart bolt.diy to get a better understanding of the inner workings; tweak it as i can.
• explore integrations (supabase, stripe, apis, etc)
• explore optimized code generation
• explore other facets of these wrappers like memory, context, multi-modal understanding, vector shit, dynamic visual hierarchy management, injection, prompt manipulation, etc. in my notes and chats with claude, perplexity and chatgpt i have managed to identify a lot of these things.

End Goals

• create my own more refined version of the large AI no-code platforms with many, many more moving parts (and for cheaper).
• improve on this interesting system
• contribute to the open source bolt.diy project
• fully understand how this stuff works without finetuning, model creation and training on codebases
• POTENTIALLY reverse engineer their api for claude entirely
• breed innovation >:)

Note

between job stuff, my startups, the markets and hectic life, i really don't have a lot of time for this. i'm just going to dick around with it while i can. i really just don't want to pay for this service anymore, it's AWESOME but holy moly it is so expensive. i def think i'll be able to at least massively improve bolt.diy, the end goal there is to heavily improve design using the methods i've learned of, give it mcp support, better chat and codebase context, stripe and supabase integration, better rules and memory. i might add additional stuff later but that's all for now.

enjoy your night.

love, ~ keygen / ra <3