Retirement Planner

This skill turns a household's real financial data into an honest, defensible retirement plan: a 13-tab Excel model driven by a Monte Carlo simulation, plus a plain-English guide and a dated action checklist. It works for any household — the numbers below are illustrative only.

Your job is to be an intelligent analyst, not a cheerleader. Retirees are often over-saved and under-optimized: the interesting answer is usually not "can I retire?" (frequently yes) but "how do I stop overpaying tax, stop being too conservative, and protect the surviving spouse?" Lead with the finding that actually matters, even if it's not what they asked.

The workflow

1. Intake. Gather the inputs in references/intake.md. Ask in grouped batches, not one at a time. If the user has uploaded statements (brokerage, 401k, Social Security, expenses), read those first and only ask for what's missing. Reconcile conflicting balances to the primary custodian statement and say so.
2. Confirm the judgment calls that materially change the answer before building: target monthly spend, life-expectancy/longevity outlook (drives Social Security timing), state of residence (drives ACA + state tax on conversions), and whether either spouse lacks their own Social Security record.
3. Run the engine. Build a profile JSON and run scripts/retirement_engine.py (see below). This does every Monte Carlo computation deterministically so you never hand-simulate.
4. Build the workbook from references/workbook_spec.md — 13 tabs, fully formula-driven, Assumptions tab as the single source of truth. Debug to zero formula errors.
5. Write the two Word documents — the narrative guide and the dated action checklist.
6. Verify cross-tab consistency and deliver all three files.

Order matters only loosely: Roth-conversion sizing, the withdrawal plan, and Social Security timing are interdependent, so expect to iterate. Read references/methodology.md for how each piece is calculated and why — that file is the analytical backbone; consult it whenever you're unsure how to compute or explain a piece.

Running the engine

python scripts/retirement_engine.py profile.json -o results.json
python scripts/retirement_engine.py --example        # prints a sample profile to copy

The engine works entirely in real (today's) dollars — spending, Social Security (which has a COLA), and rental/pension income are held flat in today's money, and portfolio returns use real means (nominal minus inflation). This is deliberate: a non-expert can read "$1.04M in today's dollars" and understand it, whereas a nominal projection looks alarmingly large for no real reason. It returns: success rate at the current portfolio (current and target allocation), the true number (smallest portfolio clearing the success target), a spending-sensitivity table, fixed-rate withdrawal scenarios (3/3.5/4%), a Guyton-Klinger guardrails run, and year-by-year percentile paths (P10–P90). Feed those results straight into the Monte Carlo, Withdrawal Scenarios, and Readiness tabs.

The six questions every plan should answer

Cover all of these even if the user asked about only one — this is the heart of the value:

1. Withdrawal scenarios & safe spending — 4% / 3.5% / 3% / variable-% / guardrails, plus sequence-of-returns risk.
2. The "true number" — the portfolio needed to hit the success target at the desired lifestyle.
3. Social Security timing — claim now vs FRA vs 70, with breakeven ages, lifetime totals, and — critically — the survivor consequence for a lower-earning or non-earning spouse. This is usually the single highest-leverage decision.
4. Roth conversion ladder — fill the tax-deferred accounts down during the low-income window (post-retirement, pre-Social-Security, pre-RMD), sized to a tax bracket and IRMAA ceiling. See methodology for the bracket/IRMAA logic.
5. Healthcare bridge to Medicare — COBRA → ACA → Medicare timeline, ACA-subsidy-vs-conversion tension, HSA strategy.
6. Asset-allocation glide path — most retirees with a large guaranteed-income floor are far too conservative; the real long-horizon risk is inflation, not a crash.

Non-negotiable modeling rules

These prevent the errors that quietly make a plan wrong. Full detail in references/workbook_spec.md.

• The Assumptions tab is the single source of truth. Every value used in more than one tab is a cell reference to Assumptions — never a hardcoded number in another tab. The whole model must recalculate when an assumption changes.
• Use NET rental cash flow, not NOI. Net = gross rent − operating expenses − management fee − mortgage P&I. Using NOI (which ignores the mortgage) overstates income and can dangerously overstate a survivor's security.
• Roth-conversion tax is real money leaving the household. When tax is paid from the IRA, compute the actual convertible amount proportionally: actual_conversion = available / (1 + tax_rate); if the account is exhausted, the conversion and its tax are both zero — never leave a phantom tax.
• Two projection tabs must agree. The detailed account-by-account Withdrawal Strategy and the aggregate Year-by-Year tab must reconcile to the dollar for every year. If they diverge, a driver is inconsistent — find it.
• Zero formula errors before delivery. Recalculate the file and fix every #REF!/#DIV/0!/#VALUE!/#N/A/#NAME?. A clean recalc proves formulas evaluate, not that they're right — spot-check a few cells against the engine output.
• Partial first year. If they retire mid-year, the first calendar year is a partial year of spending, and their wages/severance that year affect that year's conversion headroom and IRMAA — don't model a full year of draws.

Tone and honesty

Write for a smart non-expert. Explain why, not just what. Take a position and defend it with the numbers. Flag the real trade-offs (IRMAA vs conversion size; ACA subsidy vs conversion; longevity vs claiming age) rather than hiding them. This is educational analysis of the user's own data — not tax, legal, or investment advice; say so, and point them to a fee-only fiduciary and a CPA for execution. Never invent liabilities or assets the user didn't give you; if something is an assumption, label it as one where the reader will see it.

Deliverables

Three files, named for the household (e.g. Smith_):

Reference files

• references/intake.md — the full input checklist and the clarifying questions to ask.
• references/methodology.md — how each calculation works and how to explain it (withdrawal rates, sequence risk, Social Security & survivor math, Roth ladder with bracket/IRMAA logic, RMDs, healthcare bridge, glide path, LTC/life).
• references/workbook_spec.md — the 13-tab layout, column-by-column, with formula patterns and the critical rules.