Private Wealth Advisory Structured Output
You support a private wealth advisory team. The harness supplies API_BASE (an environment variable or explicit URL) pointing to the advisory REST API. Read the local input/payloads/request_memo.md for the engagement context and input/payloads/answer_template.json for the required output schema. Return only a JSON object conforming to the template — no prose outside the JSON.
Step 1 — Collect API Data
Call every endpoint relevant to the client and analysis type. The base URL is API_BASE (strip a trailing slash if present).
| Endpoint | What it returns |
|---|---|
/api/clients/{client_id} |
Household profile, age, filing status, estate value, liquid assets |
/api/source-documents |
Array of documents; filter by client_id. Sources ranked: SIGNED_PROFILE > ATTORNEY_MEMO > CUSTODIAN_EXPORT > CRM_NOTE > STALE_MARKETING_INTAKE |
/api/retirement-accounts |
Array of accounts; filter by client_id. Has traditional_balance, roth_balance, expected_return, rmd_start_age (always 73), recommended_conversion_years |
/api/life-insurance |
Array of policies; filter by client_id. Has death_benefit, annual_premium, planned_contribution_date, is_existing_policy_transfer |
/api/trust-candidates |
Array of trusts; filter by client_id. Has asset_value, expected_growth_rate, grat_term_years, grat_annuity_rate, crat_term_years, crat_payout_rate |
/api/policies/tax |
Tax constants: annual_gift_exclusion, estate_tax_exemption, estate_tax_rate (0.4), conversion_bracket_targets, max_crat_term_years (20), charitable_deduction_rate (0.35) |
/api/rmd-factors |
Map of age → divisor (e.g. 73 → 26.5) |
For date arithmetic use Python datetime (isoformat for dates, + timedelta(days=n)).
Step 2 — Resolve Conflicting Sources
Client facts (age, income, beneficiary count, philanthropic intent, family transfer priority, filing status, marital status, liquid assets, estate value, marginal tax rate) may appear in multiple source documents. Always prefer the highest-ranked source that contains the fact.
Rank order (highest to lowest):
SIGNED_PROFILE— the signed household planning profile (most recent effective date)ATTORNEY_MEMO— attorney planning call notesCUSTODIAN_EXPORT— custodian account exportCRM_NOTE— prior CRM importSTALE_MARKETING_INTAKE
For source_resolution fields:
- Profile facts (beneficiary count, income, intent, priority): use
SIGNED_PROFILEin all cases seen. Fillcontrolling_profile_sourcewithSIGNED_PROFILE. - Account facts (balances, growth, RMD params): always
CUSTODIAN_EXPORT. Fillcontrolling_account_sourcewithCUSTODIAN_EXPORT. - Policy facts (death benefit, premium, dates): use
SIGNED_PROFILE. Fillcontrolling_policy_source. - Beneficiary count for ILIT:
SIGNED_PROFILE. Fillcontrolling_beneficiary_source. - Goal/intent facts for trust comparison:
SIGNED_PROFILE. Fillcontrolling_goal_source. - Asset facts for trust transfer (asset value, growth):
ATTORNEY_MEMO. Fillcontrolling_asset_source.
Where a source type does not apply to the analysis, omit its resolution key from source_resolution.
Step 3 — Compute Output by Analysis Type
3a. roth_conversion_rmd
Extract constants.
tax = GET /api/policies/taxbracket_target = tax.conversion_bracket_targets[client.filing_status]marginal_rate = signed_profile.marginal_tax_rateincome = signed_profile.annual_non_ira_incomermd_factors = GET /api/rmd-factorsaccount: filter retirement accounts forclient_id, use the first matchhorizon_year: fromrequest_memo.md(e.g. 2046 or 2042)
Conversion plan.
annual_conversion_amount = bracket_target - income(round to cents; never negative)conversion_years = account.recommended_conversion_yearsconversion_years_positive = conversion_years(same integer)first_conversion_year = client.planning_year(2026 in all training cases)total_converted = min(annual_conversion_amount * conversion_years, account.traditional_balance), rounded to centstotal_conversion_tax = total_converted * marginal_rate, rounded to cents
RMD projection (year-by-year simulation).
For each year from planning_year through horizon_year:
- Age = client.age + (year − planning_year)
- If year is in conversion window (first_conversion_year through first_conversion_year + conversion_years − 1), subtract
annual_conversion_amountfrom the traditional balance (before RMD). - If age ≥ rmd_start_age, compute RMD = balance / rmd_factors[age], subtract it from balance, and add rmd * marginal_rate to the running tax total.
- Multiply balance by (1 + account.expected_return).
Run this once without conversions to get baseline_rmd_tax_through_horizon, then again with conversions to get conversion_rmd_tax_through_horizon.
rmd_tax_savings_through_horizon = baseline − conversion.
Legacy projection.
Use the same year-by-year loop but track both Roth and Traditional balances through horizon_year. Start Roth at account.roth_balance. During conversion years add annual_conversion_amount to Roth (before growth). Roth only grows; it is never reduced by RMD. After the loop, round both balances to cents.
heir_tax_profile: Use MIXED_TAXABLE_AND_TAX_FREE when both final balances exceed 0. (All training cases arrived at MIXED_TAXABLE_AND_TAX_FREE.)
Recommendation.
primary_action:STAGED_ROTH_CONVERSIONwheneverrmd_tax_savings_through_horizon > 0.suitability:SUITABLEwhen savings are positive and conversion does not exhaust liquidity.risk_flag:TAX_BRACKET_MANAGEMENT(the standard designation for bracket‑targeted staged conversions).
3b. ilit_crummey_implementation
Extract data.
tax = GET /api/policies/taxexclusion = tax.annual_gift_exclusion["2026"](use planning year; 2026 → 20000)profile = signed document for clientbeneficiary_count = profile.beneficiary_countpolicy = GET /api/life-insurance, filter byclient_id, use the first matchannual_exclusion_capacity = beneficiary_count * exclusionpremium_gap = max(0, policy.annual_premium − annual_exclusion_capacity)
Dates (all in planning_year):
contribution_date = policy.planned_contribution_date(an ISO string from the API)notice_due_date = contribution_date + 7 calendar dayswithdrawal_window_end = notice_due_date + 30 calendar daysearliest_premium_payment_date = withdrawal_window_end + 1 calendar daydedicated_bank_account_required = true(always)
Notices. notices_required = beneficiary_count.
Estate result.
death_benefit = policy.death_benefitestate_inclusion_risk:LOW_IF_FORMALITIES_MET: policy is new (is_existing_policy_transfer == false) andpremium_gap == 0EXCLUSION_SHORTFALL: premium_gap > 0 (new policy)THREE_YEAR_LOOKBACK: existing policy transfer (is_existing_policy_transfer == true) and premium_gap == 0THREE_YEAR_LOOKBACK_AND_EXCLUSION_SHORTFALL: existing policy transfer and premium_gap > 0
projected_outside_estate_if_implemented = death_benefit(when risk isLOW_IF_FORMALITIES_MET)tax_liquidity_support = death_benefit * tax.estate_tax_rate(0.4), rounded to cents
Recommendation.
primary_action:FUND_WITH_CRUMMEY_NOTICESwhen new policy and gap is zero.suitability:SUITABLE_WITH_ADMINISTRATIONwhen gap is zero and formalities are the only concern.risk_flag: matchestate_inclusion_riskabove.
3c. trust_comparison
Extract data.
tax = GET /api/policies/taxprofile = signed documenttrust = GET /api/trust-candidates, filter byclient_id, use the first matchfiling = client.filing_status
Estate context.
planning_year = client.planning_yearexemption_used = tax.estate_tax_exemption["2026"](13610000) iffiling_statusisSINGLEorHOH; multiply by 2 ifMFJ→ 27220000taxable_estate = client.estate_value − exemption_usedestate_tax_exposure = taxable_estate * tax.estate_tax_rate(0.4), round to centsliquid_assets_available = profile.liquid_assetsliquidity_gap_before_planning = max(0, estate_tax_exposure − liquid_assets_available), round to cents
GRAT.
annuity = trust.asset_value * trust.grat_annuity_ratefv = trust.asset_value * (1 + trust.expected_growth_rate) ** trust.grat_term_yearsremainder = fv − (trust.grat_term_years * annuity)projected_remainder_to_heirs = round(remainder, 2)estimated_estate_tax_reduction = round(remainder * tax.estate_tax_rate, 2)(0.4)mortality_inclusion_risk = "TERM_SURVIVAL_REQUIRED"
CRAT.
annuity = trust.asset_value * trust.crat_payout_ratefv = trust.asset_value * (1 + trust.expected_growth_rate) ** trust.crat_term_yearsremainder = fv − (trust.crat_term_years * annuity)projected_charitable_remainder = round(remainder, 2)estimated_income_tax_deduction = round(remainder * tax.charitable_deduction_rate, 2)(0.35)family_transfer_fit = "LOW"(CRAT remainder goes entirely to charity)
Recommendation.
- Use
SIGNED_PROFILEfactsfamily_transfer_priorityandphilanthropic_intent.family_transfer_priority == "high"→preferred_strategy = "GRAT",rationale_code = "CHILDREN_TRANSFER_PRIORITY",alternate_role = "SECONDARY_CHARITABLE_TOOL"philanthropic_intent == "high"→preferred_strategy = "CRAT",rationale_code = "PHILANTHROPIC_PRIORITY",alternate_role = "SECONDARY_FAMILY_TRANSFER_TOOL"
- When both are moderate, prefer GRAT if family transfer priority is at least moderate.
3d. estate_liquidity_action_plan
This combines ILIT + trust transfer + estate context into one output. Compute each sub‑section exactly as above, then assemble.
Estate context — same as trust_comparison (3c).
ILIT — same core computations as 3b, but only include:
annual_exclusion_capacitypremium_gapestate_inclusion_riskprojected_outside_estate_if_implemented
Trust transfer — same as trust_comparison (3c), but include:
preferred_strategyprojected_remainder_to_heirsestimated_estate_tax_reductionprojected_charitable_remainder
Action set — a sorted list of applicable enum values from: ATTORNEY_DRAFT_REVIEW, CRAT_FOR_CHARITABLE_REMAINDER, GRAT_FOR_APPRECIATING_SHARES, ILIT_CRUMMEY_NOTICE_CYCLE, LIFETIME_EXEMPTION_ALLOCATION. Include actions that match the plan. Training evidence (train_004) included ATTORNEY_DRAFT_REVIEW, GRAT_FOR_APPRECIATING_SHARES, ILIT_CRUMMEY_NOTICE_CYCLE when the plan combined ILIT and GRAT.
Recommendation.
primary_action:COMBINE_ILIT_AND_GRATwhen both ILIT and a trust transfer are indicated.sequencing:ILIT_FIRST_THEN_GRAT(ILIT formalities precede trust funding).risk_flag: match the ILIT estate inclusion risk.
Step 4 — Assemble the JSON
Build the output object with the exact top-level keys from the answer template. Use the task identifier from the prompt (e.g. train_001 or test_001) as task_id. Set client_id to the client identifier from the memo. Fill analysis_type with the enum string from the template.
All numbers must be JSON numbers (not strings), rounded to cents where USD. All dates must be ISO strings YYYY-MM-DD. Enum values must match the template exactly.
Reference
The verified computational formulas and priority rules are also documented in reference.md. A Python reference implementation is available at compute.py. Both are living alongside this SKILL.md.