Step 1, Extract transactions (AI legwork · proposes → you foot)

Layer: AI drafts (reads the statement) → deterministic foot-check confirms it. The AI is fast at reading; it is not trusted until the numbers tie.

Goal: turn one redacted statement into a clean transaction table, and prove it's complete by footing it.

Inputs: one redacted monthly statement (Step 0).

Prompt (paste into Claude)

You are extracting transactions from a bank statement for bookkeeping. Statement below.
Output a table with columns: DATE | DESCRIPTION | AMOUNT  (debits negative, credits positive).
Rules:
- Preserve EVERY transaction line. Do not summarize, merge, or skip any.
- Do not categorize and do not change descriptions.
- Mark anything you cannot read clearly as [UNREADABLE].
After the table, report:
  (a) transaction count,
  (b) the sum of the AMOUNT column,
  (c) the statement's stated Beginning Balance,
  (d) Beginning Balance + your sum  → compare this to the stated Ending Balance.

STATEMENT:
[paste the redacted statement]

Verify (this is the point) ✅

• It must foot: Beginning Balance + sum of amounts = the stated Ending Balance. If it doesn't, the extraction dropped, added, or misread a line, fix it before going further.

• Count check: the transaction count matches the lines on the statement.

• Resolve any [UNREADABLE].

You move on with: a footed transaction table for the month. (Repeat for all 12, or do one to learn the loop.)