ai-pref-nsync

[public]

git-backed personal preferences for ai coding assistants

what it does

ai-pref-nsync keeps CLAUDE.md, Codex instructions, and similar agent-config files in one versioned store. Preferences are JSON — shared profiles plus per-machine overrides — rendered into each assistant’s instruction file as a managed sentinel block, leaving hand edits outside the block untouched.

why

Every AI assistant wants its own config file. Without a shared store, preferences drift across machines and get out of sync. pref-nsync makes the store a first-class thing: one remote, one format, one command to sync.

# features

one canonical preference store versioned in git — no cloud
shared profiles with per-machine local overrides
managed sentinel block in CLAUDE.md / codex instructions — hand edits outside are preserved
four rule types: hard, conditional, repeatable-actions, conflict-resolutions
memorize script — agent appends new prefs mid-session and re-renders

example: memorize a preference mid-session

When the agent sees you giving the same correction twice, it can call the memorize script itself:

node ~/.ai-pref-nsync/scripts/memorize-preference.mjs \
  --kind hard \
  --id reproduce-bug-with-test \
  --category debugging \
  --rule "Start bug fixes by writing a failing test."

The script appends to your local store, re-renders the sentinel block, and your next prompt already obeys the new rule — no manual edit of CLAUDE.md.

rule types

hard — always-on rules
conditional — apply only when specific conditions match (repo path, category, etc.)
repeatable-actions — default commands for common tasks per repo
conflict-resolutions — memorized decisions for personal-vs-project rule clashes

# install

git clone git@github.com:ai-creed/ai-pref-nsync.git ~/.ai-pref-nsync
cd ~/.ai-pref-nsync
./scripts/bootstrap-machine.sh
node ./scripts/render-preferences.mjs