The Problem With AI and Personal Finance
Ask a generic AI assistant about your finances and you get generic answers. "Build an emergency fund covering three to six months of expenses." "Diversify your portfolio." "Maximize your 401(k) contributions." All technically correct. All useless without context.
The context problem is not the AI's fault. The model does not know your actual income, your current savings balance, your mortgage rate, or the financial goals you set eighteen months ago and have since partially revised. Every conversation starts from zero. The result is advice calibrated to a fictional average person who shares none of your specific numbers, timelines, or constraints.
Generic answers feel unsatisfying because they are unsatisfying. They do not engage with your actual situation. They engage with a template that loosely resembles your situation.
The fix is a structured memory layer. Not a third-party app. Not a paid subscription service. A single file that you maintain, that Claude reads at the start of every conversation, and that makes every subsequent answer actually relevant to your numbers rather than the population average.
The Core Setup: What Goes in CLAUDE.md
The setup uses a Claude Project with a file called CLAUDE.md stored in the project context. Claude reads this file before every conversation in the project. Everything you put there is available as context for every question you ask, without needing to re-paste it each session.
The file has five sections. Net Worth Summary: updated monthly, covering total assets, total liabilities, and net worth figure. Use round numbers rather than exact balances if you have concerns about storing precise financial data in Anthropic's cloud. A figure accurate to the nearest thousand dollars is close enough for planning purposes and meaningfully reduces the sensitivity of what you are storing.
Monthly Cash Flow: your take-home income, fixed expenses (rent or mortgage, insurance, subscriptions), variable expense baselines (groceries, dining, transportation), and your typical monthly surplus or deficit. This section is what makes spending review conversations actually work. Without it, Claude has no baseline to compare against.
Investment Positions: current allocation across your major accounts, not individual ticker-level detail, but enough to understand the shape of your portfolio. Retirement accounts, taxable brokerage, cash equivalents. Approximate percentages by asset class is sufficient for most planning conversations.
Goals With Timelines: what you are working toward, by when, and what progress looks like currently. Emergency fund target. Retirement date. Down payment goal. Each goal gets one to two sentences: the target, the current status, the monthly contribution toward it. Enough for Claude to understand priorities and trade-offs.
Decision Log: a running record of significant financial choices and the reasoning behind them. Changed jobs and accepted lower base salary for equity. Refinanced mortgage at a specific rate in a specific year. Chose to pay off a car loan early rather than invest the difference. This section gives Claude the history to understand why your current situation looks the way it does, rather than treating it as a starting point without context.
The Maintenance Workflow
The setup described requires about fifteen minutes on the first Sunday of each month. Pull up your accounts, update the balances in the Net Worth Summary, check whether your cash flow baselines have shifted, and log any decisions made in the prior month.
This is not onerous, but it does require consistency. A CLAUDE.md that is three months stale gives Claude a misleading picture, which may generate worse answers than no context at all. The discipline of the monthly update is part of what makes the system work. Without it, the memory decays into irrelevance.
The Goals section does not need updating every month, only when something has materially changed. Investment Positions need updating when you rebalance or when allocation has drifted enough to matter for planning discussions. The Decision Log only gets new entries when you make a decision worth remembering.
The practical overhead across a full year settles out to around two hours of deliberate attention to your finances. That is less time than most people spend searching for a financial calculator that handles the specific scenario they are trying to model.
The monthly update also functions as its own forcing function. You have to look at your actual numbers, not just receive bank notifications and mentally file them away. That regular confrontation with your financial state has value independent of anything Claude says in response.
The Prompts That Actually Work
The quality gap between a specific prompt and a generic prompt is enormous with this setup. The CLAUDE.md provides the context. The prompt provides the question. Generic questions produce hedged, general answers even with a complete and current context file. Specific questions with your actual numbers produce analysis that is genuinely useful.
For weekly spending review: paste your transaction list from the week and ask Claude to flag anything unusual against your stated baselines. "Here are my transactions for the week of [date]. Flag anything that looks inconsistent with my typical spending patterns from my financial context." Claude will cross-reference against your Monthly Cash Flow section and identify what is out of range, which categories ran high, and whether the variance looks like a one-time event or a pattern shift.
For scenario modeling: give Claude the exact change and ask for downstream effects on your specific goals. "If I increase my mortgage principal payments by $200 a month starting in July, how does that interact with my savings goals given my current cash flow?" This is the kind of multi-variable question where having your full goal set and cash flow in context makes the difference between a meaningful answer and a placeholder calculation using assumed numbers.
For year-end tax planning: "Given the investment positions and cash flow in my financial context, and assuming I have realized approximately $X in capital gains this year, what decisions should I be thinking through before December 31?" Claude will not give you specific tax advice. But it will walk through the relevant framework in the context of your actual situation, including the tensions between different objectives you have stated in your goals section.
The pattern: specific numbers, specific timeframes, direct questions. Vague prompts get vague answers regardless of how much context is available.
What This Setup Cannot Do
The limits matter as much as the capabilities, and stating them clearly prevents the setup from being used in ways it was not designed for.
This does not replace a human financial adviser for complex situations. Estate planning, tax optimization across multiple entities, business ownership structures, or unusual asset types all involve judgment calls, legal knowledge, and coordination across professionals that a language model cannot reliably provide. Claude will tell you this itself in most of these situations, which is part of what makes it useful as a thinking partner rather than a liability.
Claude cannot execute trades, transfers, or any account actions. Everything it suggests, you still have to do manually through your actual accounts. The system is analytical only. It thinks with you. It does not act for you.
The setup does not provide legally regulated financial advice. What it provides is an informed framework for thinking through decisions, grounded in your actual context rather than generic principles. That is genuinely more useful than most free financial content, but it is not the same as professional advice with fiduciary accountability.
On privacy: storing financial data, even rounded figures, in Anthropic's cloud is a real trade-off. If that is not acceptable for your situation, this setup is not the right approach. The round-number mitigation reduces exposure meaningfully but does not eliminate it.
What It Is Actually Good For
The setup earns its overhead on three things most people currently do poorly or not at all.
Scenario modeling before decisions. Before you change a contribution rate, pay off a debt early, or shift an asset allocation, you can think it through with a system that knows your full picture. The friction of that thinking is lower when your context is already loaded and a capable model is ready to work through implications with you.
Consistent documentation through the Decision Log. Most people cannot accurately recall why they made a financial choice two years ago. The Decision Log creates a record that exists independent of memory. When you revisit a decision, the reasoning you had when you made it is available for comparison. When you are evaluating whether a goal still makes sense, you can see what you were thinking when you set it.
A regular rhythm of deliberate attention. The monthly update is not just housekeeping. It is the practice of treating your finances as something you manage actively rather than something that happens to you. The CLAUDE.md becomes a forcing function for that practice.
This is a thinking tool. Not a replacement for professional advice. Not a trading system. Not a magic answer machine.
A thinking tool that knows your numbers.
That turns out to be rarer than it should be.