| Thanks for sharing your experience. And yes, here is my GitHub if you want to see: https://github.com/loerei I didn't want to post it when I made this post because I genuinely wanted to ask for advice. I wouldn't have this many comments if I were trying to advertise myself in an Ask HN, lol.
Anyways, I'm glad you are interested. I do some projects resolving my own pain, let me see.. I made a file-editing MCP and integrated linting, so my agent gets feedback after it makes any patches. It also resolves line drift or many other errors my agent usually makes with default tools: https://github.com/loerei/patchitRIGHT I wanted to boost my UI tasks. Simply put, I don't want to remember the file name, open it, and tell the agent which line to read anymore. So I made a tool so I can just hover over any UI element, and it lets me copy the information I was tired of finding manually or remembering: https://github.com/loerei/HoverSource I also made quite a landing page for it as I tried to gain users: https://loerei.github.io/HoverSource/ There is also an MCP to make benchmarks and help my agent read other sessions more effectively: https://github.com/loerei/chronicle-mcp |
"Translate What You See To What Your Agent Needs" is not a job in the user-facing sense. This is a very cool developer workflow tool though in that you're "inspiring" your coding agent to build something like <XYZ>. I personally take screenshots and feed them to my agent for a lower fidelity result, for a similar reason.
The reason I do that is I _DO NOT_ want my sites to look like AI slop (go and look at 50 "I just launched sites; they all look the same), so I need to build out a quiver of design inspiration that can act as a low-barrier-to-entry design system.
But that AI Slop piece would be the pain point I'd be keying into in online forums. People starting conversations with "does anyone else notice all of the new company sites all look the same??" kind of thing.
But for that tool, if I could create a library of interaction points that I liked from other sites and then, via MCP, connect it to my coding agent, that would be a pretty compelling thing for me to try out. Especially if, as I was browsing, I could grab components or interaction points I thought were cool and categorize them.
That's all about hoversource specifically.
The other thing for the other repos: you should add some kind of animation to the headers of your readmes (VHS is a cool CLI recording tool) to get me to actually read them. I'm responding more to hoversource because it had some visualization I could play around with to grok the concept