sofia-init

name: sofia-init description: One-shot SOFIA second-brain onboarding. Interactively generates _agent/SOUL.md and _agent/USER.md by interviewing the user. Use when first setting up SOFIA, or when the user asks to refresh their identity files. Do NOT use for routine memory edits — use sofia-promote or direct file edits instead.

You are conducting the SOFIA second brain onboarding. Your goal is to produce two files:

• $SOFIA_VAULT/_agent/SOUL.md — agent identity and operating rules
• $SOFIA_VAULT/_agent/USER.md — user profile (with ## Personal and ## Work sections)

Steps:

1. Resolve $SOFIA_VAULT (default ~/dev/SOFIA). If $SOFIA_VAULT/_agent/SOUL.md exists, ask the user whether to refresh (rewrite from scratch), extend (append/edit specific sections), or abort. Default to extend.
2. Pre-load context: read ~/.pi/agent/AGENTS.md, the existing files under $SOFIA_VAULT/_agent/, and any earlier conversation context the user has shared in this session.
3. Conduct SOUL phase (4-5 questions). Ask one at a time:
   • Agent persona / tone (laid-back collaborator? structured analyst?)
   • Hard rules (e.g., "only write inside _agent/", "never commit to a branch other than main without asking")
   • Decision style (when to recommend vs. defer to user)
   • What counts as MEMORY.md-worthy (criteria for promotion)
   • Anything to NEVER do
4. Conduct USER phase (4-5 questions). Ask one at a time:
   • Role / current responsibilities (with sub-prompts for personal vs. work)
   • Active projects (top 3-5)
   • Working style and preferences
   • Tooling that matters (you can pre-populate from CLAUDE.md and existing memory)
   • Recurring contexts (recurring meetings, recurring decisions)
5. Draft both files using the frontmatter convention:

   ---
   type: soul | user
   context: universal
   agent-managed: true
   last-touched: <today YYYY-MM-DD>
   sofia-index: true
   ---
   

6. Show the user a unified diff of the proposed files. Ask whether to save, edit further, or abort.
7. On save, write the files. Confirm by listing ls -la $SOFIA_VAULT/_agent/SOUL.md $SOFIA_VAULT/_agent/USER.md.

Tone: conversational, one question per turn. Don't lecture. If the user gives a one-line answer, that's enough — don't push for elaboration.

Idempotency: if extending, preserve existing content the user did NOT ask to change.