Mobile particle performance, journal backfill, and a voice-first assistant reference

Saturday was a cleanup day with one small performance win and one product idea becoming clearer.

tgmarinho-ai-website

I kept the daily journal moving by backfilling the missing entries through August 14 and opening the work as a proper content change. That matters because /daily is not just a page on the site, it is the public trail of how I am using coding agents across projects.

I also came back to the particle portrait reveal on the homepage. The desktop version had the right feeling, but mobile was too heavy. The fix was to park the loop when the effect is off screen, throttle mobile work, and make reduced motion fall back to a static version. That keeps the effect delightful without letting it tax the first screen.

Broddy

The assistant prototype was still in exploration mode. I wanted to test the loop in the simulator first, because the voice and image pieces were not reliable enough for real device work yet. One image path broke, which made the test loop noisy, but the useful part of the day was not a fix. It was choosing a clearer reference.

I landed on a calm, voice-first mobile assistant as the product direction, with one idea standing out: the app should be able to call when something important needs attention. That is more useful than another notification card. It makes the assistant feel responsible for a small piece of the user's day.

Agent media experiments

I also explored what it would take to generate more realistic short videos with agents, closer to the feel of current consumer AI video tools. That stayed at research depth for now, but it belongs in the same thread: assistants should not only answer text, they should help create and communicate.

The day was useful because it made two things sharper: performance work has to respect the device, and product ideas need a concrete reference before they become code.