I intentionally took a very simplistic approach: no Copilot in VSCode, Claude Code, OpenCode, etc. I wanted to see just how much a frontier LLM could accomplish without a sophisticated harness. What you see in the video above is the answer: it only took about half an hour of human effort to put this together.
Note that this is by no means a complete component! I look at it as an experiment: “can such a simplistic approach yield something useful?” The answer is a resounding “yes,” which raises my baseline confidence in the viability of vibe-coded components.
Nevertheless, I strongly encourage studying Richard’s post if you want to produce a more complete component.
Thanks for sharing. This confirms my recent experimentations with LLMs building DX Components. They are doing great, especially if you give them context of DX API, Constellation JS and DX Component builder.
This is very different to what I experienced only half year ago where I was not so successful with building DX Components with LLMs. Rapid progress.
@Kamil_Janeczek that’s been my experience as well. In particular, the recent-ish expansion of context windows appears to have introduced a step-change improvement in the LLM’s ability to consume complex reference material. The last time I looked into this was about a year ago, and it required much more hand holding: I could not even very approximately “one shot” it and I found myself going down the path of first “vibing” a generic React Component with the intention of having an LLM translate it into a Constellation DX Component… but it was requiring entirely too much effort and I was clearly in “this would be faster to write by hand” territory. The improvement since then is meaningful.
I’m not sure why the embedded video stopped working (it played correctly upon publication), but I switched it over to an animated GIF and that seems to resolve the matter.
