The Regulator Knows

Regulators aren't fooled by governance theatre. They're co-authors of it — and that's not cynicism, it's institutional design.

Every Control Has an Attention Budget

More controls doesn't mean more safety. Attention is finite, and over-monitoring dilutes the signal.

900 LLM Tasks While I Slept

Running an autonomous coding factory on Chinese LLM providers for $5/day

The Architect-Implementer Split

Your most expensive model should design, not implement. The pattern that turned 14 parallel coding tasks into a Tuesday evening.

The Four Layers of Every AI Agent

Interaction, inference, orchestration, tooling. The boundaries between them must be enforcement points, not design principles.

my-post

One Skill, One Triggering Context

The right decomposition boundary for LLM agent skills is the trigger, not the action.

What Naming 1,100+ Files Using Only Cell Biology Taught Me About System Design

Renaming 1,100 files using only cell biology vocabulary. Not as a metaphor -- as a constraint. Here's the method and what it revealed.

The Constitution Eats Itself

How a ruleset for LLMs metabolises its own failure modes — converting constraints into programs, leaving only irreducible judgment behind.

The Vocabulary Trap

Framework nouns colonise your thinking within a week. The discipline is naming what you're building in the problem's language before touching the framework's vocabulary.