5 posts · archived by year
Vibe-coding a ReAct agent is easy; getting it to rewrite its own prompts, tools, and memory from failure feedback is not. This post documents one full attempt — and the methodology that survived.
From the ReAct paper to a working agent skeleton: the Thought → Action → Observation loop, why it reduces hallucinations, and the engineering caveats that matter. Written in Chinese and English.
A deep dive into Astro's Islands architecture — start with zero JavaScript and hydrate only the interactive components on demand, balancing first-paint performance with developer experience.
A clear map: Flexbox solves one dimension, Grid solves two, Subgrid pierces nesting, Container Queries respond to the container. A single card component walks through each technique's proper boundary.
A real performance optimization log: measure the baseline, localize the bottleneck, optimize each item — images, fonts, JS execution, layout shift — with a reusable performance budget checklist.