Tech notes, learnings & development experiences
After Claude Code introduced mods, I made four small plugins that put uneventful AI actions on single lines and gather scattered check reminders into one line. Only after finishing did I discover others were doing similar things, and Claude Code had a similar feature of its own.
Starting with a Short that compares the Wandance anime with the creator’s hand-drawn animation, I tested AI in October 2026 at four levels: generating from dance move names, filling in motion between manga panels, animating from the creator’s drawings, and fixing the anime’s uncanny motion capture.
A workflow I worked out with Opus 5.5: using ChatGPT, Tripo, Blender, and three.js to turn one design sheet into a 3D dress-up doll that can change clothes, dance, and speak four languages in a browser. Each step includes the method, acceptance criteria, and why I chose it.
I used AI to create presentation tutorials, whiteboard videos, self-introductions, character animations, timed shorts, and music videos, and I also tested three video generation models: LTX-2.5, MiniMax H3, and Seedance 2.5. I’ve compiled a list of what AI can and cannot do right now, a three-tiered approach to creating animated visuals, and the factors that determine cost and time.
I started by running Stable Diffusion WebUI on my own computer, subscribed to Midjourney for its polish and ease, and taught two AI art classes at Feng Chia University. Back then an image took a long string of English keywords, negative prompts and LoRA; in 2026 one plain Chinese sentence is enough.
Reflections after using Claude Code for a while: wanting to make better use of the tools and quota at hand, I tried delegating work to different CLIs and different machines. This is a record of piecing things together step by step: task delegation, communication, multi-machine setup, cross-CLI usage, and verification.
Starting from a few everyday pain points—manually splitting when opening a project, windows getting stuck at 80×24, AI developing amnesia upon restart—this introduces twelve plugins extracted from my own tmux configuration over the past six months. You don't need to install them all; just install whichever one hits your current scenario.
The Claude Code 20x quota runs out in about four days. Leave the planning to Claude Code, throw batch subtasks to cheap workers on LiteLLM, and use Codex for double-checking—this "swap the model, keep the interface" workflow is also a chance to see how other models actually perform when put to work.
In February 2026, inspired by openclaw, I started using tmux, gradually building up my workflow from the status bar, split-screen, and remote development all the way to multi-agent collaboration.
memvault Trilogy Omnibus — Weaving the three tracks of writing, organizing, and retrieval into a single map, complemented by three cross-track mental models: CLT Cognitive Load, 4 Event Flows, and RxJS Reactive
Cross-session memory retrieval: Moving from “being able to store” to “being able to retrieve.” Three scenarios, one pipeline, and 11 layers of sorting.
memvault’s “Background Track” Journey. The nightly 4 a.m. dream loop, knowledge check-ups, and interest profiles—it’s not enough to simply store them in a vault; true health comes from knowing how to let go of what’s outdated.
The Journey of memvault's Write Track: From an idea sketched on a napkin to a system featuring three gates and dual-track writing.
A colleague's AI CLI tool deleted an entire project folder. This incident prompted us to start building deterministic safeguards.
From the "Hand-Coding" School to the "Command-Line" School: My Journey in AI Development
199 tools, 40K tokens in schema overhead per round—until a five-layer architecture slashed that by 90%. Practical tool management in the era of AI agents.
The reasoning behind every SEO decision made while building a blog with Astro and PostgreSQL—and the pitfalls that nearly caused posts to disappear from search results.