Son

Babysitting AI. Exploring crypto × AI.
Former Blockchain Architect.
Built websites at Yahoo back in the day.

[Son @phuson]

Spent more time polishing Tomo. Added a goggle with a frame over the eye. When the cursor is nearby, it tilts slightly toward it to fake a 3D feel. Redid the eyelids so Tomo looks kinda terrified when dragged. Really like how it’s coming together, feels a lot more polished now.

[Son @phuson]

Here is Tomo's trying to get my attention.

[Son @phuson]

Spent a good part of last week polishing Tomo’s look and feel. Made it more 3D looking with subtle gradients, added just enough glow to make it a bit shiny. A bit softer and fluffier, less of a perfect circle so it feels more organic. Also added more depth to the sclera and pupil

[Son @phuson]

My current agentic coding workflow is basically: first half of the week having AI build a bunch of features, second half having AI clean up and fix all the slop it generated.

Not sure if this is actually better than before AI yet, but it does feel more productive.

[Son @phuson]

As much as people here hype up AI, working with agentic coding feels like a robot vacuum. You set boundaries and prep everything before letting it run. Even then it gets stuck, needs rescuing or eats something it shouldn’t. Sometimes it acts like it’s not even in the same house.

[Son @phuson]

Implemented the debug monitoring hub for Tomo over the weekend. Now I can actually see what’s going on inside. What mood it’s in, what it’s looking at, what expression state it's in. No more guessing blindly at what it’s doing. Getting this ready before integrating the LLM brain.

[Son @phuson]

Switched from a transparent SKView to using SKRenderer so shaders actually work over transparent chrome. Trying to see how far I can push the visual effects with this setup.

Maybe a bit much for a selection ring. 😁

[Son @phuson]

Time to start working on the brain and memory for my little friend. I have some interesting ideas I want to try with this little guy.

However, knowing me, I’ll probably spend another week polishing the visuals and animations. 😅

[Son @phuson]

Was going to learn Rive to animate my little desktop friend. Didn’t have the time for that.

So I’m trying a programmatic approach instead. Graphics and animations in code.

Faster to translate what’s in my head and have Codex generate the code for most of it.

Way more flexible.

[Son @phuson]

Using AI to decorate a room for the first time. Leaving the bedroom is more of a suggestion now.

[Son @phuson]

Started building my own little desktop friend. Next I need to figure out @rive_app so I can give it a new look. Also chatting with ChatGPT to architect its brain in TypeScript. Not really sure where this is going but it should be interesting.

[Son @phuson]

The last 2 weeks were mostly refining the synthetic cursor and trying a bunch of motion blur ideas for zoom and pan. Still more to do, but I will get back to it later when I have more time.

[Son @phuson]

Finally got my screen recorder app to a good stopping point. Started as a prompt in ChatGPT to see how hard it would be. 3 weeks later of prompting in Codex and zero coding, I can now use it to record and share things I am building. Here is a quick video of what it can do.

[Son @phuson]

It took about a weekend to get the basics working. Recording the screen and exporting. Then about a week to polish the UI and add live preview and basic editing. Took a lot of inspiration from Screen Studio.

[Son @phuson]

Redoing the motion blur for my little experimental screen recorder app. Maybe the zoom effect is a bit too Michael Bay for a screen recorder app. Time to tune it down a bit.

[Son @phuson]

Managing all the coordinators, agents, and workspaces adds cognitive overhead. IMO things could be simplified since it is easy to feel overwhelmed.

Intent is in beta and is still a bit buggy, but I like the direction they are heading.

[Son @phuson]

Been using Intent by @augmentcode this past weekend.

I like where it is going. Feels like they have the right intention, but parts of it seem overcomplicated/confusing.

The biggest one for me is workspaces. I think I understand them, but I still feel like I am using them wrong

[Son @phuson]

I do like that you can inspect each agent and see the conversations between them.

But after a few back and forth with the coordinator sending change requests, it gets hard to track which subagents are doing what. It is also not completely clear which subagent is still active.