Private Wealth Advisory Planning
Produce structured JSON planning outputs using the task-group advisory API for high-net-worth clients. Fetch client data from API endpoints, resolve conflicting source documents, compute projections with the bundled scripts, and return a JSON object conforming to the task's answer template.
Workflow
1. Identify the engagement type
The task request memo names the engagement. Map it to analysis_type:
| Engagement keyword | analysis_type |
|---|---|
| Roth conversion, RMD | roth_conversion_rmd |
| ILIT, Crummey | ilit_crummey_implementation |
| GRAT, CRAT, trust comparison | trust_comparison |
| estate liquidity, action plan | estate_liquidity_action_plan |
2. Fetch data from the advisory API
The API base URL is provided in the task environment (usually as API_BASE).
Fetch these endpoints relevant to the client. Use exact client_id from the
request memo:
GET /api/clients/{client_id}
GET /api/source-documents
GET /api/retirement-accounts
GET /api/life-insurance
GET /api/trust-candidates
GET /api/policies/tax
GET /api/rmd-factors
Full endpoint reference: references/api_reference.md.
3. Resolve conflicting sources
Clients may have multiple source documents of different types. Resolve by selecting the most recent authoritative source. Precedence:
- SIGNED_PROFILE — signed by client, most recent
- ATTORNEY_MEMO — attorney notes
- CUSTODIAN_EXPORT — authoritative for account balances and returns
- CRM_NOTE — older import, may be stale
- STALE_MARKETING_INTAKE — oldest, least reliable
For retirement accounts, always prefer CUSTODIAN_EXPORT for balances and expected_return. For client attributes (income, marginal rate, beneficiaries, intent), prefer SIGNED_PROFILE.
4. Compute numerical projections
Use the bundled scripts for deterministic results. Do not hand-roll the computation loops; the scripts produce the exact values the evaluator expects.
Roth conversion / RMD projection
python3 skill/scripts/roth_projection.py \
--traditional <traditional_balance> \
--roth <roth_balance> \
--return <expected_return> \
--rmd-start-age <rmd_start_age> \
--client-age <age> \
--horizon <horizon_year> \
--planning-year <planning_year> \
--marginal-rate <marginal_tax_rate> \
--annual-income <annual_non_ira_income> \
--bracket-target <conversion_bracket_targets[filing_status]> \
--conversion-years <recommended_conversion_years>
The script outputs JSON with: annual_conversion, total_converted,
total_conversion_tax, baseline_rmd_tax, conversion_rmd_tax,
rmd_tax_savings, projected_roth, projected_traditional.
See references/roth_conversion.md for the computation model, recommendation logic, and field enum values.
GRAT / CRAT trust projection
python3 skill/scripts/grat_crat.py \
--asset-value <asset_value> \
--expected-growth <expected_growth_rate> \
--grat-term <grat_term_years> \
--grat-annuity-rate <grat_annuity_rate> \
--crat-term <crat_term_years> \
--crat-payout-rate <crat_payout_rate> \
--estate-tax-rate <estate_tax_rate> \
--charitable-deduction-rate <charitable_deduction_rate>
The script outputs JSON with: grat_remainder, grat_estate_tax_reduction,
crat_remainder, crat_income_tax_deduction.
See references/trust_comparison.md for the full computation model, recommendation logic, and field enum values.
Crummey timeline
python3 skill/scripts/crummey_timeline.py \
--contribution-date <planned_contribution_date> \
--beneficiaries <beneficiary_count>
The script outputs JSON with: contribution_date, notice_due_date,
withdrawal_window_end, earliest_premium_payment_date, notices_required.
See references/ilit_crummey.md for the full computation model, recommendation logic, and field enum values.
5. Assemble the output JSON
Use the task's payload answer_template.json as the schema. Fill every
required key. The template defines key names and field types (strings, numbers,
enums). Enums must match exactly as listed.
When the template lists fields with dot notation (e.g.
recommendation.primary_action), nest them as objects:
{
"recommendation": {
"primary_action": "STAGED_ROTH_CONVERSION"
}
}
Specific field notes
task_id: the task identifier string (e.g."train_001"or"test_001")client_id: stable client identifier from the request memo- Integers must be JSON integers, not strings
- Numbers must be JSON numbers, not strings
- Dollar amounts: round to cents (2 decimal places)
- Dates: ISO 8601 YYYY-MM-DD strings
action_set: when included, sort alphabeticallyheir_tax_profile: choose based on projected roth vs traditional ratio at horizon
6. Return only the JSON object
Do not include prose, commentary, or markdown fences outside the JSON. The output must be parseable as a single JSON object.
Analysis-specific details
For detailed computation models, field enum tables, and recommendation logic for each analysis type, read the corresponding reference file:
- Roth conversion / RMD: references/roth_conversion.md
- ILIT Crummey: references/ilit_crummey.md
- GRAT vs CRAT: references/trust_comparison.md
- Estate liquidity: references/estate_liquidity.md
Each reference file contains the exact formulas verified against the advisory task environment, enum lookup tables, source-resolution precedence, and recommendation decision trees.