Read the screen

Exports do not carry the bank's own balance, the For Review count, the bank rules, whether a feed is connected, or who authored an entry. Four of those five decide what a catch-up consists of, so a run that skips them is a run doing arithmetic on an unknown.

Read this first: there are two ways to get them, and the better one is work-in-the-browser. Where the owner runs an agent with browser access and has QuickBooks open in their own signed-in window, the agent reads four of these five off the screen itself, records each with the date, and keeps the read on disk so it can be checked afterwards. A number a person recalls cannot be checked, and a number read off the screen and saved can be.

python3 bin/books.py browser confirm --company "the name the header shows"
python3 bin/books.py browser read --surface banking --from reports/browser-reads/banking.json

This skill is what runs when that is not available: a chat agent with no browser, a session where the owner would rather not have anything driving their books, or any moment where the browser read comes back in a shape that does not prove what it claims. Then ask, because the owner is already logged in and can read five numbers off a screen faster than any scraper.

The one figure browser mode still cannot get on its own is the bank's own balance, unless the owner also has their banking portal open. It does not come from the QuickBooks tile. The tile shows what the books think the account holds, and that is the figure under test.

Hard gates

• Ask, record, and stamp the date. An undated answer is indistinguishable from a guess in a month. Every fact goes in with the day it was given and who gave it.
• Never infer one of these from an export. A feed that stopped and an account that went quiet look identical in a General Ledger. Say which is which only when a person has looked.
• Where QuickBooks can export the thing, ask for the export instead of the number. The For Review list per account (For review, then Export to Excel) and the rules file (Rules, then Export rules) are both better than a count, because a count cannot be checked afterwards.
• Keep it short. Six answers across five screens, not an interview. A long list gets no answers at all.

Run it

python3 bin/books.py intake

With no arguments it prints each screen, what to read off it, why it matters, and the exact line to type back. Run it again at any point: it keeps everything already given and lists only what is still missing, with what each gap blocks.

python3 bin/books.py intake --account 101000 --bank-balance 37.14 --as-of 2025-12-31
python3 bin/books.py intake --account 101000 --for-review 241 --feed stopped --feed-last 2025-03-14
python3 bin/books.py intake --account 101000 --reconciled-through 2025-06-30
python3 bin/books.py intake --rules 15 --auto-add 15
python3 bin/books.py intake --last-human 2026-02-28 --note "the bookkeeper left"

The five screens

How to ask

One question per paragraph, each ending in a question mark, ordered by what unblocks the most. Say where the screen is before you say what to read. Never ask for something the exports already answer, and never ask them to decide something that is yours to decide.

A good round reads like this: "Open Transactions, Bank transactions in QuickBooks. On each account tile there is a line under the balance saying when it last updated. What does it say for each account?"

The same facts, read off the screen instead

Where browser mode is available, these are the commands that replace the asking. Each one validates the shape of what it was given before it records anything, so a short read of a virtualized grid is refused rather than believed.

python3 bin/books.py browser read --surface banking            # tiles: queue depth, feed state
python3 bin/books.py browser read --surface for-review         # one account's queue, in full
python3 bin/books.py browser read --surface reconcile-summary  # reconciled through, every account
python3 bin/books.py browser read --surface rules              # rules, and how many post unseen
python3 bin/books.py browser read --surface audit-log          # who last worked in the file

They write into the same answers/live-screen.json with the same dates, so every later command reads them the same way and cannot tell which route a fact arrived by.

What is done with the answers

They go in answers/live-screen.json, each with its date. Every later command reads them, and each one says plainly what it still lacks rather than assuming a zero. Three refusals turn on them directly: a merge cannot proceed without the For Review count, the queue cannot be called finished without the bank balance, and nothing produces figures for the return without the rules count.

What good looks like

Six facts recorded, each dated. Two exports on disk rather than two numbers recalled. Every command that runs afterwards naming what it used.