~~Software is dead~~. What's changing is how we interact with it.
We still need Gmail, but now agents can read our emails. We still use Slack, but now agents can participate in our conversations.
I built a browser extension that works as a renderer for the browser’s native bookmark system.
It supports workspaces, pins, favorites, and more, while keeping everything entirely local to the browser.
No backend, no external services, no account required.
chromewebstore.google.com/de…
I’ve noticed 3 stages of using AI:
1. Use AI for everything.
2. Start questioning whether AI can actually do everything well.
3. Learn how to harness AI to actually get things done
I keep opening more technical articles than I can actually read — partly because of FOMO.
But when I finally come back to them, I often realize:
1. I already knew most of it
2. A summary would have been enough
3. It’s actually worth reading in full
Meanwhile, I have too many browser tabs open.
I don't know how to make videos.
When I built usereframe.io, I needed a demo video for the homepage. codex made one for me — I just recorded a silent screen capture. That's amazing.
Code agents can now review code and use tests to verify correctness, but I still feel like I'm losing a sense of control.
Maybe this is an outdated concern, even AI can still generate logically correct but weird code.
I wonder if code quality matters less in the AI era.
As a developer, I like having a certain level of control over my code.
When I'm writing code, I naturally know where things are going.
But with code agents generating so much code, reviewing becomes much harder.
Built `coma` while working on an OCR feature. It’s token-efficient.
github.com/Hunter-Gu/coma
The Claude API allows a max of 8 skills.
Should I limit the number of skills in Claude Code/Copilot to optimize performance?
I built `uskills` to solve this. It's a CLI tool to manage and link AI agent skills across your workspace.
github.com/Hunter-Gu/uskills