Fewshot Attempt 01

Private Wealth Advisory – Structured JSON Output Skill

1. Input Workflow

For each task:

1. Read input/payloads/answer_template.json first. It defines the exact analysis_type, required top-level keys, field enums, and data types.
2. Read input/payloads/request_memo.md for all client facts, goals, account balances, policy details, and dates.
3. Read input/prompt.txt only for any special instructions; it is usually a generic directive to analyze the memo and template.
4. Produce only a JSON object matching the template schema. No markdown fences, no prose outside the JSON.

2. Output Format Conventions

• task_id must match the task identifier from the directory/prompt (e.g. train_001, test_001).
• client_id comes from the request memo (pattern CLT-XXXX).
• analysis_type must match the template exactly.
• USD amounts are JSON numbers (not strings). The training data uses one decimal place for whole-dollar amounts (e.g. 105300.0) and two for cents when present.
• Dates are ISO-8601 (YYYY-MM-DD).
• Enums must match the template exactly (case-sensitive, underscores preserved). Common enums observed:
  • primary_action: FUND_WITH_CRUMMEY_NOTICES, REPLACE_WITH_GRAT, STAGED_ROTH_CONVERSION, COMBINE_ILIT_AND_GRAT, etc.
  • suitability: SUITABLE, SUITABLE_WITH_ADMINISTRATION, SUITABLE_WITH_RISK_ACKNOWLEDGMENT
  • risk_flag: LOW_IF_FORMALITIES_MET, TAX_BRACKET_MANAGEMENT, HIGH_DUE_TO_LIQUIDITY_GAP
  • heir_tax_profile: MIXED_TAXABLE_AND_TAX_FREE, MOSTLY_TAX_FREE, MOSTLY_TAXABLE
  • estate_inclusion_risk: LOW_IF_FORMALITIES_MET

3. Source-Resolution Hierarchy

When data conflicts appear in the memo (e.g., profile vs. email vs. custodian), resolve with this priority:

Always record the final choices in a source_resolution block at the end of the JSON. Keys vary by analysis type but commonly include:

• controlling_profile_source / controlling_goal_source / controlling_estate_source
• controlling_account_source / controlling_liquid_source
• controlling_policy_source / controlling_beneficiary_source

Rule of thumb: SIGNED_PROFILE wins for any fact the client personally attested to; CUSTODIAN_EXPORT wins for hard financial data.

4. Estate-Tax Calculation Rules

• Federal estate tax rate is consistently 40% in the training data.
• taxable_estate = gross_estate_value - exemption_used
• estate_tax_exposure = taxable_estate × 0.40
• liquidity_gap = estate_tax_exposure - liquid_assets_available (can be negative, indicating surplus liquidity)
• tax_liquidity_support from an ILIT = death_benefit × 0.40 (the estate-tax-equivalent liquidity provided by insurance)

5. GRAT Calculation Rules

• Use the Section 7520 rate provided in the request memo.
• annuity_payment = grattable_value / IRS_annuity_factor(term_years, section_7520_rate)
  • The annuity factor is the present-value-of-an-annuity factor for the given term and 7520 rate (exact IRS Table B or equivalent). Do not invent the factor; derive or verify it from the memo/context.
• projected_remainder = the expected trust assets remaining after the annuity term, given the asset growth assumption.
• estate_tax_reduction = projected_remainder × 0.40
• tax_liquidity_support = estate_tax_reduction + liquid_assets_available (total liquidity created by the strategy)
• remaining_liquidity_gap = estate_tax_exposure - tax_liquidity_support

6. ILIT / Crummey Administration Rules

• annual_exclusion_per_beneficiary is typically $20,000 in the 2026 planning-year context (verify in memo).
• annual_exclusion_capacity = per_beneficiary_amount × beneficiary_count
• premium_gap = max(0, annual_premium - annual_exclusion_capacity)
• Crummey timeline (strict 7-day / 30-day convention):
  • contribution_date from memo
  • notice_due_date = contribution_date + 7 days
  • withdrawal_window_end = notice_due_date + 30 days (or contribution_date + 37 days)
  • earliest_premium_payment_date = day_after(withdrawal_window_end)
• dedicated_bank_account_required = true for ILITs
• estate_inclusion_risk = "LOW_IF_FORMALITIES_MET" when Crummey notices are administered properly
• projected_outside_estate_if_implemented = the policy death benefit

7. Roth Conversion & RMD Rules

• RMD start age is 75 under SECURE 2.0 for the clients in the training set (verify birth year; age 74 in 2026 → first RMD year 2027).
• first_rmd_year = year client turns 75
• conversion_years = first_rmd_year - planning_year (years available to convert before RMDs begin)
• conversion_years_positive = max(0, conversion_years)
• annual_conversion_amount is either:
  • total_traditional_balance / conversion_years when converting the full balance, or
  • The bracket-limited amount when the memo specifies staying within a marginal bracket.
• total_converted = annual_conversion_amount × conversion_years
• Conversion tax rate pitfall: The effective tax rate applied to conversions may differ from the stated marginal rate. In training data, a stated 24% marginal rate produced a 32% effective rate (total_conversion_tax / total_converted = 0.32). Always compute or verify the implied rate from the context rather than blindly using the stated marginal rate.
• RMD projection through horizon_year:
  • baseline_rmd_tax_through_horizon = tax on RMDs if no conversion occurs
  • conversion_rmd_tax_through_horizon = tax on RMDs after reducing the traditional balance
  • rmd_tax_savings_through_horizon = baseline - conversion
• Legacy projection:
  • projected_roth_balance_horizon = converted principal grown tax-free at the assumed rate through the horizon year
  • projected_traditional_balance_horizon = residual traditional balance after conversions, RMD withdrawals, and growth
  • heir_tax_profile: MIXED_TAXABLE_AND_TAX_FREE when both Roth and traditional balances are projected at horizon; use MOSTLY_TAX_FREE or MOSTLY_TAXABLE when one dominates.

8. Combined Strategy Rules (estate_liquidity_action_plan)

• When the memo asks for a combined ILIT + GRAT strategy:
  • primary_action = "COMBINE_ILIT_AND_GRAT"
  • sequencing = "ILIT_FIRST_THEN_GRAT" (standard order observed)
  • Calculate each component separately (ILIT exclusion capacity, GRAT remainder, estate tax reduction) and roll them into the combined JSON.
  • action_set is an array of exact strings from the template enum (e.g. ["ATTORNEY_DRAFT_REVIEW", "GRAT_FOR_APPRECIATING_SHARES", "ILIT_CRUMMEY_NOTICE_CYCLE"]).

9. Common Pitfalls

• Source priority errors: Defaulting to EMAIL or CRM_NOTE when SIGNED_PROFILE or CUSTODIAN_EXPORT is available.
• Type errors: Returning "20000" or "20000.0" instead of 20000.0. The template expects JSON numbers.
• Missing source_resolution: Every output must include the source_resolution block.
• Crummey date math: Using calendar-month approximations instead of exact day counts (7 days for notice, 30 days for withdrawal window).
• Estate tax rate assumption: Using a rate other than 40% when the context is federal estate tax.
• Premium gap sign error: Reporting a negative gap when the capacity exceeds the premium; gap must be max(0, premium - capacity).
• Prose leakage: Any text, comments, or markdown outside the JSON object causes validation failure.

10. Pre-Submission Checklist

1. task_id and client_id match the input.
2. analysis_type matches the template exactly.
3. All required top-level keys from answer_template.json are present.
4. All enums match the template exactly (case-sensitive).
5. All monetary values are JSON numbers.
6. source_resolution reflects the correct hierarchy.
7. Dates are ISO-8601.
8. No markdown, no comments, no trailing text outside the JSON object.