What it means
CLAUDE.md, AGENTS.md, GEMINI.md and any other target the policy names are not written by hand. majordomus update generates them from one canonical policy, the profiles, and one shared body text, through a small template per provider. Running it twice with the same inputs produces byte-identical files.
How it works
lib/update.sh hashes the policy and profiles, renders each target from its adapter — the templates under the distribution's share/providers/, or the repository's own override under its providers section — and writes each target atomically. Every generated file is a bootstrap: it names README.md, points at .ai/README.md, names the default profile and the lifecycle workflow, and carries no rule of its own. Its first line is a stamp naming the command that produced it, the policy hash it came from and the hash of its own content. Adapters translate; they do not add rules — a rule that exists for one provider and not another is a policy bug by definition, and doctor fails a generated file that has grown a rule corpus.
How to see it
majordomus update && sha256sum CLAUDE.md
majordomus update && sha256sum CLAUDE.md # identical hash; output says "unchanged CLAUDE.md"What it does not cover
majordomus update projects a deliberately narrow subset of the policy — the budget, the profile table, state locations, the finish contract, the identity-fields rule — and lists what is not projected. It does not merge an existing hand-written instruction file; you decide what moves into the body on first use.
Why it exists
In one workspace of about twenty repositories, the instruction files for two AI tools in the same repository shared between none and a tenth of their content. They were not duplicates; they were disjoint rulebooks, and which one applied depended on which tool was open.