Personalized onboarding: goals-driven setup that shapes identity, soul, and prompts

[matthewod11-stack]

Summary

Transform onboarding from a data-collection exercise into a goals-driven conversation that deeply personalizes Gary — shaping IDENTITY.md, SOUL.md, prompt selection, and priority weighting based on what the user actually needs help with. The current onboarding asks who you are; this asks what you need and adapts everything accordingly.

Why This Matters

Every Gary deployment is different:

• A founder needs email triage + calendar prep + investor research
• A PM needs Slack synthesis + meeting prep + project tracking
• Someone using Gary as a personal assistant wants personal email, family calendar, research, reminders — not Slack channels
• A developer might want GitHub activity + code review prep + minimal email

Right now, everyone gets the same prompts, the same SOUL.md, the same briefing format. Personalization makes Gary feel yours instead of generic.

Proposed Flow

Phase 1: Goals & Needs Assessment

Use AskUserQuestion with a mix of multiple-choice (select all that apply) and free-form questions:

Q1: "What do you most need help with?" (select all that apply)

• Email triage — too many emails, need someone to sort signal from noise
• Calendar management — meeting prep, schedule awareness, conflict detection
• Slack/messaging synthesis — keeping up with channels without reading everything
• Research — staying informed on topics, competitors, industry
• Task/project tracking — keeping commitments visible, following up
• Personal organization — family calendar, personal email, life admin
• Writing/drafting — emails, docs, announcements
• Other: ___

Q2: "How would you describe your work style?" (pick one)

• Fast-paced, give me bullets and action items only
• Thorough — I want context and reasoning
• Casual — talk to me like a friend who happens to be organized
• Professional — clean, structured, no personality needed

Q3: "What's your biggest recurring frustration?" (free form)

• This surfaces the pain point that Gary should prioritize solving

Q4: "Is this primarily for work, personal life, or both?"

• Work only
• Personal only
• Both — but work is primary
• Both — roughly equal

Q5: "What tools/surfaces do you check most?" (rank or select top 3)

• Email, Slack, Calendar, GitHub, Linear/Jira, Notion, Discord, other

Q6: (If "personal" selected) "What personal things should Gary help with?"

• Family calendar coordination
• Personal email management
• Research / learning
• Health / fitness tracking
• Finance / bills
• Travel planning
• Other: ___

Phase 2: Identity Generation

Based on answers, Claude generates a customized IDENTITY.md:

# IDENTITY.md — Who Am I?

- **Name:** Gary (or user's chosen name)
- **Role:** Personal productivity assistant (work-focused)
- **Vibe:** Direct and action-oriented  ← derived from Q2
- **Primary focus:** Email triage, meeting prep  ← derived from Q1
- **Secondary focus:** Slack synthesis  ← derived from Q1
- **Pain point:** "I miss important emails because they're buried in noise"  ← from Q3
- **Human:** [name from existing onboarding]

Phase 3: Soul Generation

SOUL.md gets generated from a personality questionnaire:

Q7: "When Gary gives you information, do you prefer:"

• Just the answer, no explanation
• Answer + brief reasoning
• Full context so I can make my own call

Q8: "Should Gary proactively suggest things, or only respond when asked?"

• Be proactive — surface things I might miss
• Mostly reactive — I'll ask when I need something
• Proactive for urgent things only

Q9: "How should Gary handle uncertainty?"

• Flag it clearly and let me decide
• Make a best guess and note the uncertainty
• Ask me before proceeding

Q10: "Pick the vibe that feels right:"

• Sharp coworker who knows everything
• Calm executive assistant
• Enthusiastic teammate
• No personality — just the data

These answers generate a personalized SOUL.md with the right tone, proactivity level, and communication style.

Phase 4: Prompt Adaptation

Based on Q1/Q4/Q5 answers, adjust prompt behavior:

User Need	Prompt Impact
Email-heavy	Briefing leads with email triage, expands email section
Slack-heavy	Briefing leads with Slack digest, more channel coverage
Personal mode	Skip work-oriented language, add personal calendar/email sections
Research-focused	Enable research-refresh task, expand knowledge-compile
Minimal comms	Shorter briefings, bullet-only format, skip narrative sections
Meeting-heavy	Prioritize meeting-prep frequency, expand prep depth

Implementation options:

• Option A: Conditional sections in prompt files (e.g., <!-- IF email_priority == high -->)
• Option B: Generate user-specific prompt variants during onboarding
• Option C: A config/user-profile.json that prompts read at runtime to adjust behavior

Recommend Option C — single config file, prompts check it, easy to update later without re-onboarding.

Phase 5: Import Existing Identity

Support bringing identity from other AI tools:

"Do you already have an AI assistant identity or memory you'd like to import?"

• Claude Desktop (check ~/.claude/ for memory/identity files)
• Claude Code CLI (check for CLAUDE.md in home or common project dirs)
• ChatGPT (paste or upload custom instructions / memory export)
• Other (paste text)
• No, start fresh

If importing: parse the content, extract relevant personality/preference signals, merge into IDENTITY.md and SOUL.md rather than overwriting.

Technical Implementation

Config File: config/user-profile.json

{
  "priorities": ["email", "calendar", "slack"],
  "mode": "work_primary",
  "style": "direct_action",
  "proactivity": "proactive_urgent",
  "uncertainty_handling": "flag_and_decide",
  "vibe": "sharp_coworker",
  "pain_point": "Missing important emails in noise",
  "personal_areas": [],
  "imported_from": null
}

Prompt Integration

Prompts read user-profile.json at runtime:

## User Profile
Read `config/user-profile.json` to understand what this user cares about most.
Adjust section ordering and depth based on their priorities.

AskUserQuestion Usage

Use AskUserQuestion tool throughout — it supports:

• Multiple choice (select one)
• Free-form text
• Can be called sequentially for a conversational flow

For "select all that apply" — present as numbered list, ask user to type numbers (e.g., "1, 3, 5") or use multiple sequential single-select questions.

Verification

• Onboarding asks goals/needs questions before data source setup
• IDENTITY.md is generated from answers (not template defaults)
• SOUL.md is generated from personality questions
• config/user-profile.json is written with parsed preferences
• At least one prompt (morning-briefing) reads user-profile.json and adapts
• Import path works for at least one source (Claude Desktop memory)
• User can re-run personalization without losing data source connections

Automation Hints

scope: CLAUDE.md (onboarding section), IDENTITY.md, SOUL.md, config/user-profile.json (new), prompts/morning-briefing.md
do-not-touch: SETUP.md, prompts/shared/, data source connectors
approach: extend-onboarding-flow
risk: medium (touches core onboarding UX — test with a fresh clone)
max-files-changed: 8
blocked-by: none (can build independently of #2 onboarding redesign, but should integrate with it)
bail-if: AskUserQuestion tool unavailable or onboarding redesign conflicts

Priority

Medium-High — this is a differentiator. Generic AI assistants are everywhere; one that adapts to your workflow is the value prop.