Profile Setup - Personal-Finance Context Capture

Purpose

Personalize every advice skill by storing one structured profile of life context (separate from Wealthsimple account data). Stored locally only at ~/.aifolimizer/personal_context.json with mode 0o600, never committed, never sent to external LLMs in raw form.

How to run

1. Call mcp__aifolimizer__get_personal_context to see what is already stored. Keep context_hash in mind - if the user updates anything, the hash flips and self-learning loops will know to re-weight.
2. Ask the user which mode they prefer:
   • Guided: walk through field groups conversationally. Recommended first time.
   • Paste template: user pastes a filled JSON object - tool validates and writes atomically.
3. Walk through groups (skip any group the user does not want to share - every field is optional):

Group A - Identity (★★★)

• age
• province (e.g. ON, BC, AB)

Group B - Income (★★★)

• gross_salary_cad
• employer_match_pct (RRSP match)
• bonus_target_cad

Group C - Cashflow (★★★)

• monthly_expenses_cad
• monthly_savings_target_cad

Group D - Goals (★★★)

• fthb_yes (first-time home buyer?)
• home_horizon_years
• mortgage_planned
• dependents_count
• spouse_present

Group E - Tax shelter room (★★★)

• room_rrsp_cad, room_tfsa_cad, room_fhsa_cad

Group F - Risk (★★)

• risk_tier_per_account - map per account-type, values ∈ {conservative, moderate, aggressive}

Group G - ESPP (★★)

• espp_employer_label (anonymized - "EmployerA", never the real name)
• espp_discount_pct, espp_cycle_months, espp_monthly_contrib_pct, espp_qualifying_hold_months

Group H - Liabilities (★★)

• iou_receivable_cad (sum only - never counterparty names)
• debts (list of {type, balance_cad, rate_pct} - type examples: student_loan, credit_card, line_of_credit)

Group I - Banking (★)

• institutions (list of anonymized labels: "BankA", "BankB" - never real names like RBC/TD)
• external_cash_cad (non-WS chequing + HISA total)

Group J - Insurance (★)

• insurance_tier ∈ {basic, middle, comprehensive}

Group K - External crypto (★★)

• external_crypto (list of {symbol, qty, cost_basis_cad, platform_label} - platform_label is anonymized: "PlatformA")

4. After each group, call mcp__aifolimizer__set_personal_context_field for each provided field (one call per field - atomic + idempotent).
5. For paste-template mode: ask user to paste a JSON object matching backend/personal_context.template.json. Then call mcp__aifolimizer__set_personal_context_bulk once with the parsed payload.
6. After capture, call get_personal_context again and read back the derived block (marginal_tax_rate_pct, emergency_fund_target_cad, account_waterfall, fhsa_priority_first) so the user sees what skills will now infer.

Output structure

1. Confirm what was set (field count, fields touched).
2. Show derived summary: marginal tax rate, emergency-fund target, account-funding order.
3. Print the new context_hash (16 hex chars) - used by track-record stratification.
4. List skills that will now produce more personalized output: cash-deployment, tax-loss-review, dividend-strategy, portfolio-health, risk-assessment, daily-briefing, pre-trade-check.
5. Reminder: re-run this skill any time a field changes (job change, dependents, FTHB completed, contribution room used).

Rules

• Never store real names, emails, account numbers, employer names, bank names, counterparty names. Use anonymized labels. If user provides a real name, replace before calling set_personal_context_field.
• Every field is optional. Do NOT pressure the user to provide a value - accept silence or "skip" and move on.
• Validate values before submitting: province must be 2-letter Canadian code; risk tier ∈ {conservative, moderate, aggressive}; balances and rates ≥ 0.
• Do NOT echo dollar values back in summaries beyond order-of-magnitude (use bands like ">90k") if the user later asks for a sharable summary.
• If user says "wipe my profile", call mcp__aifolimizer__clear_personal_context and confirm.

Gotchas

• Field validation happens server-side via Pydantic - if a set_personal_context_field call fails, surface the error to the user verbatim and re-prompt that field only (do not abort the whole flow).
• risk_tier_per_account expects a full dict, not a single key - collect all account-type tiers in one prompt before sending.
• debts and external_crypto are lists - sending an empty list [] is fine; do not send placeholder example items from the template.
• personal_context_hash is short (16 hex). It will change every time any of the canonical-keys fields is updated. That is expected and is what enables self-learning stratification.
• For the paste-template mode, the template ships with _INSTRUCTIONS and _help keys - strip those before validating.