Julie Private Assistant

Problem Addressed

When a personal-assistant model must continuously remember preferences, health reminders, schedules, and long-term notes, the failure modes are usually not “it cannot write,” but writing to the wrong file, skipping write verification, scanning full files instead of retrieving relevant entries, relying on hallucinated time, and mixing todos with knowledge. julie turns private affairs into a fixed set of profile files and a retrieval index, so each run starts from explicit state.

How It Works

The skill follows read profile → infer intent → execute, splitting data across users.md, doctor.md, knowledge.md, and todo.md, with vector_index.json maintaining a BM25 index. Before querying, editing, or deleting, it prefers scripts/bm25_search.py to locate entries instead of reading whole files; after writes, it re-reads the file to verify persistence and updates updateDate in version.json. For dated schedules, it can create a cron reminder set 10 minutes before the event. At conversation start, it checks the data directory, version consistency, and today’s tasks, forming a real-time inspection.

Boundaries

It is best for one user’s private profile, health reminders, knowledge notes, and task lists, not for multi-user collaboration or complex project workflows. The skill is intentionally restrained: it asks when intent is unclear, avoids inventing data, does not automatically delete history, and does not proactively nag with habit reminders. If the runtime lacks Python, jieba, or cron, retrieval and reminder behavior may be limited.