Sharing an Atlas

Git is how an Atlas is shared. When several people work in the same Atlas, its own files live in a git repo of their own; a personal Atlas can skip it. Either way the sources keep their own git — the Atlas never folds them into its history.

What is versioned, and what is not

• Versioned: the Atlas's own files — the .md files, every _ folder (_archived/, _settings/, …), .vscode/ and .gitignore.
• Never versioned: the symlinks — their targets are paths on one machine, so a committed symlink is a dangling pointer on any other — and .claude/settings.local.json, which holds those same paths.

What the Atlas shares about its sources is their declaration in _settings/atlas.toml — their remote without credentials, never a user:token@ an origin may carry; where each one lives is up to each person. A token that reached the Atlas's history is compromised: rotate it, cleaning the history is not enough.

The .gitignore — deny everything, allow the Atlas's files

/*
!/*.md
!/_*/
!/.vscode/
!/.gitignore

Every current or future symlink is ignored automatically, with no list of sources to maintain. .gitignore can't select by file type (no "ignore symlinks"), so allowing the known set of Atlas files is the form that never needs editing when a source is added.

Cloning a shared Atlas

Clone it anywhere, then rebuild the sources on your machine: atlas doctor lists what is missing or broken, with the exact commands to get each source, and atlas link wires it in. atlas doctor also flags any orientation block still unwritten or with placeholders; the full list of its findings is in 03-atlas-anatomy.md.