Structure

A conformant file uses these section titles, in this order. Levels follow RFC 2119. The spec is the source of truth. Voice fingerprint samples SHOULD come from real sent or published writing; see How do I make the file accurate?

LevelSectionIn the file
MUSTMetaname, kind (person or organization), language, version, updated, license
MUSTIdentityWho is speaking, and what they do not do
MUSTVoice fingerprint3 to 7 writing samples
MUSTTone rulesConcrete instructions, not adjectives
MUSTHard bansWords, phrases, and moves the model must not use
MUSTSafetyNo secrets. No invented facts, metrics, or clients
SHOULDRegister shiftsHow email, social, website, and proposal differ
SHOULDBefore and afterGeneric AI copy rewritten in this voice
SHOULDFacts and claimsFacts the model may treat as true
MAYHow to useShort pointer to the load guide

Example Spec 0.1