Stop putting everything in CLAUDE.md
Claude Code has four mechanisms for shaping agent behavior. Most people use one. I researched how they actually work and validated my findings with TDD.
Translunar LLC
Juno Woods is an aerospace engineer, AI researcher, and computational biologist. Spacecraft navigation, AI agent systems, and inference under uncertainty.
Latest posts
Anthropic quietly removed MagicDocs from Claude Code
MagicDocs was the least-modified of Claude Code's five oldest system prompts. It and /pr-comments — both legacy approaches from before the skills era — were removed in v2.1.91.
Gender-affirming health insurance coverage how-to
A practical guide to getting gender-affirming care covered by health insurance in the US — prior authorization, network exceptions, letters of agreement, and how to file a complaint when your insurer stonewalls you.
A call for a quantitative report card for AI bioterrorism threat models
A proposal for a quantitative report card to evaluate how much LLMs like ChatGPT and Claude actually elevate bioterrorism risk across three threat models: technical, social engineering, and disinformation.
Four types of verbal consent
A framework for the four main types of verbal consent — enthusiastic, affirmative, no means no, and token no — and why understanding which one you're using (and which one the other person expects) actually matters.