I posted 10 comments on Hacker News over the past 2 weeks while building an open-source Chrome extension. I learned more from the replies than from any feedback form.
Three things kept coming up:
The "shipping vs shipping right" debate is a false choice. People ship rough and iterate, or polish and ship late. Both work. What fails is shipping nothing.
Local models excite devs, but the math doesn't work for interactive coding yet. At 6-11 tokens/sec, a 200-line task is a 4-minute wait per turn. Fine for planning, brutal for back-and-forth.
The "no strangers will pay" problem is rarely about the product. It's usually one of three things: zero repeat usage, wrong distribution channel, or trust gap that SEO can't close.
CWS: https://chromewebstore.google.com/detail/eigpaeoigklelmfgnkljhbjjbpohenpn
GitHub: https://github.com/mnbqwe10/ai_buddy
What did you learn from your first month of building in public?
Top comments (4)
Interesting, very true on the 3. I notice that it's really similar in idea to some of the tools I built for MCP-Lite, except those were for the agent to DIY, instead of you feeding it data from a sidebar. That being said, AI Buddy is definitely a useful tool, would save alot of time and back and forth, when dealing with things like setting up Vertex AI, where LLMs struggle to keep context of newer models and you need to spoon feed them the page-data for the latest model so it can understand how to use it.
Thanks for the MCP-Lite comparison, that distinction is the part I keep under-selling. The sidebar works because the human decides which slice of the page goes in. You feed the model only the row you care about. I never framed it that way in the original post, you just did.
Vertex AI onboarding is exactly the failure mode. The "what is this field" question burns 4 turns before the model can answer. A one-keystroke page pin is genuinely faster than chat for that.
Sorry this reply is 24 days late. The notification sat unread because I was heads-down on the cron-recovery loop. Will be more responsive from here.
No stress at all, rather late than never
The "6-11 tokens/sec, not worth it for interactive" point is exactly why I gave up on local TTS and stayed on a cloud API for my daily pipeline.