Hello World

For the past few months, I've been using LLM agents to help run real work — legal research, board governance, household coordination, college planning, writing, and the everyday mess that doesn't fit neatly into a productivity app.

I'd built what I thought was an architecture — workflows, configurations, structural choices about how to organize agent context, how to shape the data the agents work with, how to route work between models.

A lot of it was wrong.

The system worked enough that I kept using it. But most of the structural choices were somewhat right in a way that kept me from seeing how the operating model itself was broken. The system did real work. It was also just sluggish and frustrating enough to force me to keep rethinking and rebuilding.

That journey, and what came out of it, is what this publication is about.

This is NeuraLiminal

Liminal is from the Latin limen — threshold. NeuraLiminal is the space between human cognition and artificial intelligence, where the interesting work happens. Not what AI does on its own, not what humans do without it. The interaction layer where the two meet, and where almost all the design decisions worth making live.

The interesting questions in AI right now aren't about the models. They're about the system humans need to use them well, which is mostly underbuilt.

This publication is about how that interaction layer should be structured — not which prompts to copy, not which model just released. It's a record of that work: what breaks, what holds, and what it takes to make AI useful at the threshold between human judgment and machine assistance.

About me

I'm Andy Cho. I'm not approaching this as a coder or an AI researcher. I'm trying to make agentic systems useful across messy, high-context human work — using whatever combination of Claude Code, Codex CLI, and Gemini CLI fits the task. The architecture I'll publish is the second or third version, because the first one was wrong, the second is mostly working, and the third is what I'm building toward.

I don't know what I don't know. But I'm passionate about AI as a force multiplier for human potential, not a replacement for it. If you scrolled to the bottom of the archive and found this, you probably already get the vibe.

Welcome.

Tutorial Part 5: When Configuration Becomes Architecture

Tutorial Part 5: When Configuration Becomes Architecture Part 5 — the final part — of the Command-Line LLM Tools tutorial series. Previously: Part 4 — When the Agent Doesn't Know Enough — extending capability with skills and sub-agents. Every fix in this tutorial series scales until it doesn't. The instructions

Tutorial Part 4: When the Agent Doesn't Know Enough

Tutorial Part 4: When the Agent Doesn't Know Enough Part 4 of the Command-Line LLM Tools tutorial series. Previously: Part 3 — When the Agent Goes Off the Rails — git as time machine, approval modes, hooks for hard guarantees. Parts 1, 2, and 3 made the agent safe and

Tutorial Part 3: When the Agent Goes Off the Rails

Tutorial Part 3: When the Agent Goes Off the Rails Part 3 of the Command-Line LLM Tools tutorial series. Previously: Part 2 — When the Agent Forgets — managing context, compaction patterns, files-as-memory. Part 2 solved memory. The agent stays sharp now — context doesn't rot, instructions don't get

Tutorial Part 2: When the Agent Forgets

Tutorial Part 2: When the Agent Forgets Part 2 of the Command-Line LLM Tools tutorial series. Previously: Part 1 — Getting Started with Claude Code & Codex CLI — install, instructions file, the five rules that prevent day-one mistakes. In Part 1, we set up a working CLI agent with an instructions