← All use cases

中文

Finance & bookkeeping · Claude Code

Turn brokerage exports into a written investment plan to review

Idea + starter prompt, not run by us yet

For: Investors who hold accounts at several places and want one organised view

The author pointed Claude Code at brokerage CSV files, a typed note of his other balances and a long goal prompt. Claude wrote code to sort the holdings, ran a gap analysis, looked up funds and produced a ten-section plan. A CLAUDE.md file kept context between sessions over several evenings.

Steps

  1. Export CSVs from each account into one folder, remove account numbers from a copy, and add a typed note of anything else you own.
  2. Write the goal first: your constraints, accounts you will not touch, and target mix. The goal prompt mattered most.
  3. Tell Claude to ask you questions until it has enough, then to critique its own plan.
  4. Save progress notes in a CLAUDE.md file so later sessions pick up where you stopped.
  5. Treat the plan as research. Check each fund and tax point with a licensed adviser before you trade.

Starter prompt

Starter prompt. Change the [brackets] first.

Trial first: work on a redacted copy of my exports, with account numbers removed, in a folder called [redacted-copy]. Put the trial plan in [trial-plan.md]. Never open the original files until I say yes.

The redacted CSV exports from my brokerage accounts are in [redacted-copy], plus a note called [other-balances.txt]. My goal: [goal]. My limits: [accounts you must not touch, risk limit, time horizon]. Read the files and tell me what you found before anything else. Ask me questions until you are sure of my constraints.

Then write a plan with a gap analysis and a staged action list. Rate your own plan and say what is weak. It only reads and drafts for my review. Do not log in, place a trade or move money, and wait for my yes before any step that touches an account. A plan is not permission.

Original post: mattstockton.com · found via agentusecases.ai. Written up in our own words. We haven't run it, so test it on made-up data first.