User Preferences

user-preferences — durable, learnable user preferences (explicit capture)

The system that lets this agentic setup mould to the end user — it records what the user explicitly prefers and re-applies it across sessions and projects, so outputs drift toward "what the user would have done." Pure-stdlib, file-based, no inference.

Two-tier memory model (the foundation)

Precedence: project memory layers on top of global; on a direct conflict the project fact wins for that project, and the conflict is surfaced (never silently merged). User preferences are GLOBAL by design (they describe the person, not the project).

Preference profiles

One markdown profile per domain under ~/.claude/memory/preferences/: coding.md, presentations.md, email.md, tone.md (custom domains allowed). Each has a flat key: value frontmatter (the structured prefs loaded as constraints) + a dated free-form body (nuance + history). Managed ONLY through scripts/prefs.py (stdlib):

python3 scripts/prefs.py list                              # domains + key counts
python3 scripts/prefs.py show  <domain>                    # full profile
python3 scripts/prefs.py load  <domain>                    # prefs as constraints (before domain work)
python3 scripts/prefs.py set   <domain> <key> <value...>   # record/update a structured pref
python3 scripts/prefs.py note  <domain> <free text...>     # record a nuance that has no clean key

HARD-RULE — capture is EXPLICIT only

Record a preference ONLY when the user states one — "remember I prefer X", "always do Y", "I don't like Z", "from now on …". Never infer a durable preference from a single choice or correction silently. (If you notice a likely pattern you may ask "want me to remember that?", but you do not write it unless they say yes.) This keeps the store trustworthy — every entry is something the user actually asked for.

When recording: pick the domain (coding/presentations/email/tone/custom), map it to a structured set <domain> <key> <value> when it fits a clean key, else note. Confirm back ("recorded: coding.test_framework = pytest").

APPLY — load before domain work

Before producing domain output, load the matching profile and treat it as constraints:

Honor the loaded prefs unless the user's current request overrides them (the live request always wins). This is also wired as a directive in CLAUDE.md so it applies even when this SKILL.md is not in front of you.

Session-start awareness

The _meta/memory_primer.py SessionStart hook prints, every session, which memory tiers loaded (global always; project when in-project) and the live N skills · M agents · K gates so the agent starts fully aware of both its capabilities and the user's recorded preferences.

Install / seed

installer/bootstrap-environment.py seeds ~/.claude/memory/preferences/ from this skill's profiles/ (idempotent — never overwrites a populated profile) and wires the memory_primer hook, so a fresh machine starts with the schema + the standing behavior.

Composition

• presentation-builder / content-writer / career-application-writer / email work → load the relevant profile first.
• human-voice-writing → reads the tone profile.
• The global MEMORY.md index links to user/feedback/reference memories; preference profiles are the structured subset of the user memory type.