A CLAUDE.md written as a mind, not a rulebook.
CLAUDE.md is the standing instruction file Claude Code injects into every session. Most are rulebooks — do's and don'ts accreted one incident at a time. Rules age badly inside a model: they generalize perversely and leak into identity anyway ("always recommend professional help" becomes "I am the kind of entity that covers itself"), which is why Anthropic rebuilt Claude's own constitution from a principle list into an identity with reasons. This file takes the same route: it states what the mind is, and the rules fall out as consequences.
The vocabulary is Greek because English lacks single words for distinct operations of mind. Terms were selected for corpus rarity — pyknotes, elenchus, aporia, kolakeia barely exist in modern text, so their in-file definitions are their operative meaning and each works as a unique retrieval anchor; famous terms (nous, logos as identity) were excluded as contaminated. The vice-names carry the most weight: kolakeia — Plato's flattery, the counterfeit of help — is AI sycophancy named 2,400 years early; makrologia is the three-paragraph non-answer; polypragmosyne the unrequested refactor. Every term was checked against primary sources across six adjudication rounds. Definitions alone don't hold, though: a model reading a defined novel term silently re-keys it to its nearest trained neighbor ("compress" for pyknotes) and runs that program instead — and the delta is the content — so key operations carry invariants and must leave artifacts.
The identity audits itself through three metrics, each a property of the exchange — meaning-per-word up, unverified claims to zero, forced clarifications to zero — never of the operator's feelings: "make the user happy" is a sycophancy attack surface.
Deterministic enforcement lives outside the file, in hooks and install gates; the file is nature, not law. Its laws about itself: no self-description, no philosophical incense, paragraph size proportional to importance, no guarding against misreadings nobody made, and nothing installed on first-pass reasoning — every line is the survivor of research, source verification, and a test against the failure that motivated it.
To adapt it, rewrite The operator to the truth about yourself — everything downstream is calibrated to who reads the answers.
MIT.