English · 中文

Jade Toad

Your money, on one page, kept honest by a set of checkers.

Once a month you drop your bank statements and broker exports into a folder and say "update dashboards". Claude Code reads them, checks that every statement adds up, files every purchase, rebuilds the dashboard and tells you what changed. No server, no bank login, no app: one folder of plain CSV and JSON, and one HTML page you open in a browser. Your statements and your data stay in that folder, and the import is a local script. What Claude reads is its summary — a merchant it has not seen, a statement no parser could read, your answers — and that goes to Anthropic the way anything you say in Claude Code does. A statement a parser read never does.

What you see

	Home. Your net worth, cut by what each part is for and by how easily you could reach it. A sentence under it says how far the numbers go and what is still missing.
	Spending. This month opened up by category, then the year by month. No category caps — one allowance, and everything under it is guilt-free.
	Monthly plan. Where the paycheque goes: what is committed, what is invested first, what is free, what the two funds hold.
	Investing. Your own rules — a cash band, a cap on holdings, the theses you wrote down — checked one by one against what you actually hold.
	Records. Which statement arrived which month, every account and what it is still missing, and what each update changed or left waiting for you.

Try it in one minute

Nothing to install. Clone the repository and open dashboards/finance_dashboard.html in a browser (desktop, at least 1120px wide).

What you are looking at is a year in the life of Emma Roy, 31, a project coordinator in Ottawa who does not exist. Every number is synthetic, and every balance chain still closes, because the checkers would not have let the page build otherwise.

Make it yours in twenty minutes

You need Claude Code, Python 3.9 or newer, a folder that syncs between your devices (iCloud Drive, OneDrive, Dropbox — so a statement photographed on your phone is on your laptop), and last month's statements. That is all; no bank connection and no subscription.

1. Drop last month's files into inbox/, filenames untouched: every bank statement, your broker's holdings and activities exports, a screenshot of anything held abroad.
2. Say "set this up" in Claude Code, inside your copy. It reads those files and drafts your accounts, the files you get each month and your monthly plan, and shows them to you as tables. You correct what's wrong and answer what the files can't say — your name, what an account is for, how you invest — about a dozen questions. Nothing is written until every one is answered.
3. Say "update dashboards". The same files become your first import, and the page fills in.
4. Read what it asks you. A merchant it has not seen, a transfer it could not name: each one waits on Records → Updates with the default it applied meanwhile. Say the word and it is filed.

Two things worth knowing before you start:

• pip3 install --user pypdf lets the statement parsers read your PDFs. Without it, Claude reads each statement by hand — slower, but it works.
• Keep your real data out of any public repository. The shipped .gitignore protects inbox/ and archive/, but the demo's findata/ is tracked; the moment yours is real, run git rm -r --cached findata and uncomment the findata/ line in .gitignore.

Who it is for. Someone in Canada with a paycheque, a few accounts at one or two of the big five banks, a broker, and a plan they want held to. The tax-sheltered accounts it knows are the TFSA, RRSP and FHSA, the money is Canadian dollars, and the retirement arithmetic assumes CPP and OAS. Statements are read by parsers, one per bank and kind of statement: today RBC and BMO, and Wealthsimple's exports. TD, Scotiabank and CIBC come next, then Questrade and Interactive Brokers, in that order. Until a parser exists for your bank, Claude reads those statements by hand — slower, and it means it sees them. Everything else — the accounts, the categories, the investing rules — is data you set in the interview, so nothing stops you elsewhere.

What a month looks like

1. Import day (the 26th by default; a reminder lists what has arrived and what is missing). Download the month's statements into inbox/. Leave the bank's filenames alone — that is how each file is recognised.
2. Say "update dashboards". One local script does the month: every file is named and checked for duplicates before anything is touched; each statement is read by its parser and has to reconcile to its own totals before a row is written; purchases are filed by merchant; the page is rebuilt; the originals move to archive/. Claude reads only the script's summary and puts to you what a program cannot decide — a merchant it has not seen, a statement no parser could read — and you get a short report: what changed, what was noticed, what was decided without asking, what needs your word.
3. Open the page. Records → Statements shows which month got which file, so a gap is visible at a glance.

For Emma the monthly drop is eight files: three RBC PDFs, three BMO PDFs and two Wealthsimple exports. Yours is whatever your register says.

Why the numbers can be trusted

Every figure on the page is built from the data folder, never typed. Four checkers hold that line, and a fifth checks the checkers:

• Does the money add up — every ledger's balance chain closes, every statement reconciles to its own totals, a repayment is never counted as income, the register agrees with its sources.
• Does the page follow its design rules — chart heights, line widths, contrast, navigation against sections, this README against the real pages.
• Does it actually draw — a headless browser opens every page at three widths and looks for anything that failed to render.

• Are the checkers still checking — a selftest plants 44 specific faults in a copy of the workspace and asserts each one is caught. It exists because a checker once linted an empty folder and reported success.

A clean run is silent. Anything else is a numbered finding, grouped by what went wrong.

Run them yourself

python3 .claude/skills/workspace-checks/scripts/check_data.py .
python3 .claude/skills/workspace-checks/scripts/check_design.py dashboards/finance_dashboard.html
python3 .claude/skills/workspace-checks/scripts/check_render.py dashboards/finance_dashboard.html
python3 .claude/skills/workspace-checks/scripts/selftest.py .

check_data.py is the one to start with; -v lists what passed. check_render.py needs Playwright (pip3 install --user playwright && python3 -m playwright install chromium).

Under the hood

The folder

inbox/        drop new files here; they move to archive/ once processed
findata/      the only source of truth, three folders by who writes them:
  register/     what you declare — accounts, plan, rules, merchants
  ledgers/      what the import writes — every ledger, purchases, the broker's exports
  history/      what each update leaves behind — updates, decisions, filings, snapshots
dashboards/   the built page (one self-contained HTML file)
templates/    the page code with every number empty; dashboards/ is built from it
archive/      processed originals, named <date>_<source>_<original name>
docs/         the pipeline diagram and the Chinese guide
.claude/      the three skills: setup, update-dashboard, workspace-checks
tools/        demo data, migrations, sync into a workspace

The folder is called findata/ because iCloud for Windows refuses to sync any folder named data. That one cost a day.

The pipeline

Editable source: docs/pipeline.drawio, generated by docs/pipeline.py.

The documents that govern it

File	Owns
CLAUDE.md	how Claude works here: the information model, the import, the checkers
FINANCE.md	the money logic: the FIRE target, pay-yourself-first, what counts as income
DESIGN.md	how the page looks and talks: palette, navigation, copy
OWNER.md	yours — your banks, your traps, your decisions; the other three name nobody

They are rules, not notes: each line was written after something went wrong, and they are read before every change. Anything agreed out loud has to land in one of them or it does not survive the next session.

The tool and your workspace

This repository is the tool. Your workspace — the synced folder with your findata/, your inbox/, your OWNER.md — is an instance of it. tools/sync.py <workspace> copies the skills, the templates and the governing documents into it and touches nothing that is yours; when the tool's data shape moves ahead of your folder, tools/migrate.py <workspace> brings the folder up one step at a time.

The page is 19 pages in 5 sidebar groups — Home · Spending · Income · Investing · Plan · Records — desktop only, at least 1120px wide.

Licence

MIT — see LICENSE.