Deep dive · wiki
How to build agentic systems for real
Not a post: a 16-module course-wiki, from the substrate to the frontier, every claim sourced and graded for credibility. Read it end to end or drop into any topic.
- 16
- modules
- 8
- phases
- ~32k
- words
- ~3h
- read
- 145
- sources
Live components · lab
Every component, running for real
Not screenshots: the live gallery of every interactive piece on the blog — charts you scrub, diagrams that play, small simulators — running in their real environment, each with the reason it was built.
- 31
- components
- 5
- categories
- 26
- stable
- 100%
- live
Shipping a high-accuracy AI agent: what to prioritize, what to ignore
The gap between a 95% agent and a flaky demo is the harness, not the model. What actually raises an AI agent's hit rate, and the distractions to skip.
Read moreInteractive Blog Posts: Why I Built 30 Components Instead of Pasting Screenshots
Why this blog renders live simulators, draggable easing curves and step-through diagrams instead of static screenshots: the idea behind its 30 MDX components, what they cost to build, and the research — from Bret Victor's explorable explanations to Distill — on why a claim you can drag beats a claim you can only read.
Game feel on the web: squash, shake, and the art of juice
A button that just works is dead. I spent an evening teaching a side-project to punch back, and these are the six techniques that did it. Each one playable, dependency-free, and respectful of your settings.
Git worktrees let me run three Claude agents in parallel
Three Claude agents on three bugs in one afternoon, only because each ran in its own git worktree. The file-isolation discipline that makes parallel AI agents safe, and the mistake that taught it to me.
AI gave me speed. Tests gave me control.
I ran /insights after a month with Claude Code in a side project. The numbers I expected weren't the interesting ones. Here's what was.
Learn by building: why makers learn faster than students
The most effective way to master a technology isn't reading docs or watching tutorials, it's shipping something real. Here's the case for learning by building.
Custom slash commands for AI coding: build your own spec-driven workflow
How 6 markdown files turned my AI coding assistant into a structured co-architect and why you should build your own command system instead of relying on raw prompts.