Extract data from a stack of PDF invoices into one clean CSV

← Use cases · Accounting

Extract data from a stack of PDF invoices into one clean CSV

Suppliers send invoices in every format imaginable. You re-type the fields into Excel for hours. Drop the PDFs into a folder, ask Claude for a CSV with supplier, date, amount, VAT, and invoice number — twenty minutes later you import the CSV into your accounting tool. Audience: beginner · ~25 min · 7 steps

You'll end up with

A `invoices.csv` file with one row per invoice — supplier name, invoice number, issue date, net amount, VAT amount, gross amount, due date — ready to import into your accounting tool.

Before you start, you need

• Nothing — Claude will invent 6 realistic synthetic invoices in different layouts so you can practise the extraction loop. (Once you trust it, drop real PDFs from your inbox into the same folder — the prompt is identical.)
• Claude Code installed and a terminal open (see Day 1)

1. 01 Build the workspace
2. 02 Ask Claude to invent 6 invoices in different layouts
3. 03 Pick one and let Claude show you what it sees
4. 04 Ask Claude to process all of them
5. 05 Spot-check the CSV against the originals
6. 06 That CSV is your handoff point
7. 07 Level up

Re-typing invoices into a spreadsheet is the dictionary definition of work that shouldn’t exist. Suppliers send PDFs with totals in seven different places, dates in three formats, and supplier names that don’t quite match what’s in your accounting system. The accounts-payable clerk does it because nobody else will.

This use case ends that — once you’ve done one batch by hand to set the format, every future batch is a one-prompt job.

Compliance check before you start. Invoices contain supplier names, bank details, amounts, and sometimes payment terms — internal-confidential. Same posture as policy documents: fine in a Claude-readable folder on your workstation, not for sharing on web-based AI tools. If your suppliers include large counterparties (e.g. infrastructure providers, technology vendors), confirm with your line manager before the first batch.

---

Steps

1. Build the workspace
2. Ask Claude to invent 6 invoices in different layouts
3. Pick one and let Claude show you what it sees
4. Ask Claude to process all of them
5. Spot-check the CSV against the originals
6. That CSV is your handoff point
7. Level up

← Back to gallery

Step 01

Build the workspace

First — pick your OS

The commands below switch to match. Your choice carries across every use case on this site.

macOS Windows

Open the Terminal app.

Press ⌘+Space, type “Terminal”, and press Enter.

Open the Start menu (press the Windows key), type “Ubuntu”, and press Enter. If you don’t see Ubuntu listed, install WSL first.

In Ubuntu, ~/Desktop is a folder inside WSL’s Linux home (/home/<your-Linux-username>/Desktop) — not the Windows desktop you see in File Explorer at C:\Users\...\Desktop. That’s fine: the files are real and Claude can read and write them. Anywhere this use case says “open in Finder / File Explorer”, run explorer.exe . from your Ubuntu terminal — Windows opens that exact WSL folder in Explorer.

Type each line:

mkdir -p ~/Desktop/claude-lab/invoices-may
cd ~/Desktop/claude-lab/invoices-may
mkdir invoices
claude --dangerously-skip-permissions

• mkdir -p ~/Desktop/claude-lab/invoices-may — make a folder called invoices-may inside claude-lab on your Desktop (-p creates claude-lab too if it’s not there yet — it’s the one home for all hub use cases).
• cd ~/Desktop/claude-lab/invoices-may — move into it.
• mkdir invoices — make a sub-folder where the invoice files will live.
• claude --dangerously-skip-permissions — start Claude Code. The flag stops Claude prompting you for permission on every file write — safe in a fresh, dedicated folder like this one. (If you’d rather see every prompt for your first run, just type claude — same thing, more interruptions.)

Step 02

Ask Claude to invent 6 invoices in different layouts

You don’t have real supplier PDFs and you don’t need them. Tell Claude:

Create 6 synthetic invoices inside the invoices/ subfolder, each as a separate file. The point is to simulate the real-world mess: every supplier lays their invoice out differently.

Use these filenames and layouts:

1. invoices/01-acme-cleaning.md — a simple top-aligned layout: supplier name + address top-left, invoice number / date top-right, then a single-line item table, then totals.
2. invoices/02-piraeus-stationery.html — a more chaotic HTML layout: a logo placeholder div top-centre, then a small detail block on the right with VAT number + IBAN, then a multi-line items table.
3. invoices/03-attica-electric.md — supplier on the right, “Bill To” block on the left, invoice number embedded in a sentence (“Invoice ref: ATTI-2026-0094 dated 2026-05-08”), totals at the bottom.
4. invoices/04-thessaloniki-logistics.md — line items spread across two pages (use a <!-- page break --> marker), totals on page 2, VAT noted as a separate column (24%, the Greek standard rate).
5. invoices/05-london-software-ltd.md — a UK supplier billing in GBP, VAT shown at 20%, due-date phrased as “Net 30 days from invoice date”.
6. invoices/06-nbg-coffee-machines.md — a tiny supplier, no formal invoice number — just “Receipt #4421” — to test that Claude flags this rather than guessing.

Each invoice must contain (somewhere in its layout): supplier name and address; an invoice number (or receipt number for #6); issue date in any plausible format; one or more line items with description + qty + unit price; net total; VAT amount; gross total; due date (if applicable); currency.

Use realistic Greek supplier names and addresses where appropriate. Vary the amounts — anywhere from €120 to €18,000 (or £-equivalent for the London one). Use realistic Greek VAT numbers (format: 9 digits) and IBANs (format: GR16 ... for Greek ones).

Make the layouts genuinely different — don’t just rename the same template six times. That’s the whole point.

Claude writes all six files straight away.

You’ve just done two things real bookkeepers don’t realise are possible: generated synthetic test data, and watched Claude vary layouts on purpose. When you do this on real PDFs next week, the prompts that follow are identical — only the file extensions change (.pdf instead of .md/.html).

Step 03

Pick one and let Claude show you what it sees

Different suppliers lay out invoices differently. Before you ask Claude to do all of them, have it look at one so you can confirm it’s pulling the right fields.

Tell Claude:

Pick any one file from invoices/. Extract these fields and show them to me as a table:

• Supplier name (the company you’re paying)
• Invoice number
• Issue date (in YYYY-MM-DD format)
• Net amount (the amount excluding VAT)
• VAT amount
• Gross amount (net + VAT)
• Due date (in YYYY-MM-DD format)
• Currency (EUR, USD, etc.)

If any field genuinely isn’t on the invoice, mark it as MISSING. Don’t guess.

Press Enter. Claude reads the PDF and shows you the parsed fields.

This is the most important step. If “supplier name” is coming out as “Ltd Trading Co” (the right-hand side of a name) instead of the full company name, fix the prompt now — for example: “the supplier name is whatever appears next to ‘Bill from:’ or in the top-left letterhead, full legal name including ‘Ltd’ / ‘A.E.’ etc.”

Iterate on one invoice until the fields are right.

Step 04

Ask Claude to process all of them

Once one invoice parses correctly, tell Claude:

Now apply the same extraction to every file in invoices/. Write the results to invoices.csv with these columns in this order:

supplier_name,invoice_number,issue_date,net_amount,vat_amount,gross_amount,due_date,currency,source_file

One row per file. The last column source_file is just the filename — so we can trace any row back to the original document if something looks wrong.

Where a field is genuinely missing, write MISSING (uppercase, no quotes).

Dates always YYYY-MM-DD. Amounts always with . as decimal separator (no thousands separator).

Press Enter. A batch of 30 invoices takes 2–5 minutes. Claude works through them and writes the CSV.

Step 05

Spot-check the CSV against the originals

Ask Claude:

Show me the first 5 rows of invoices.csv.

Pick two rows. For each one:

1. Ask Claude “show me invoices/03-attica-electric.md” (or whichever the source_file is).
2. Compare the row’s net amount, VAT, and gross against what the source actually says.
3. Confirm net + VAT = gross to the cent.

If a row is wrong, tell Claude which source_file mis-parsed and what the right value should be:

Row for 04-thessaloniki-logistics.md has VAT as 24.00 — the actual invoice shows VAT as 240.00. Re-extract this one and update the CSV row.

Pay special attention to:

• Decimal vs thousands separators. European invoices often use . for thousands and , for decimal (1.250,00). Claude may flip them — verify on any line where the amount looks oddly large or small.
• Multi-currency invoices. A GBP invoice in a batch of EUR ones is easy to miss if the currency column isn’t filled.
• MISSING fields. Are they genuinely missing, or did Claude not find them? Look at one or two MISSING entries to be sure — the 06-nbg-coffee-machines.md “receipt” should legitimately be MISSING in the invoice_number column.

Iterate. Once the spot-checks pass, you trust the file.

Then make Claude audit every row, not just your two

Arithmetic doesn’t need your eyes. Paste:

Audit invoices.csv row by row, with commands, and show me the results:

1. For every row: net_amount + vat_amount must equal gross_amount to the cent. Show the arithmetic.
2. For every row: compute the VAT ratio. Greek invoices should be 24%, UK 20% — flag any row that’s neither.
3. The CSV’s row count must equal the number of files in invoices/.
4. For any flagged row, re-open the source file, find the discrepancy, and fix the CSV.

Re-run the audit until it’s clean.

The VAT-ratio check is the sneaky-good one: it catches the European decimal-separator swap (1.250,00 read as 1.25) automatically, because a mangled amount almost never lands on exactly 24% or 20%.

Step 06

That CSV is your handoff point

In real life this is the moment you’d import invoices.csv into your accounting tool — SAP, Oracle, Xero, QuickBooks all take a CSV import; you’d map our supplier_name to their Vendor once and save the mapping for next month. We’re pretending here — the CSV on your Desktop is the deliverable.

The pattern for next month: drop fresh PDFs into the invoices/ folder, run the same prompt — Claude reads PDFs natively too. Twenty minutes for 30 invoices instead of two hours.

The deeper win: the CSV is also a perfect starting point for spend analysis. “Group by supplier_name, sum gross_amount, sort descending” gives you your top vendors of the quarter — a question that used to require a separate request to finance.

Make every future batch a one-liner with CLAUDE.md

The fixes you made in Step 3 (“supplier name is the full legal name, including ‘Ltd’ / ‘A.E.’”, date format YYYY-MM-DD, . as decimal separator, MISSING for genuinely absent fields) shouldn’t be re-typed every month. Tell Claude:

Create a CLAUDE.md in ~/Desktop/claude-lab/invoices-may/. Put in it the field-extraction rules I established today:

• Supplier name = the company being paid, full legal name including suffixes like Ltd / A.E.
• Invoice number = whatever appears next to “Invoice #” / “Invoice No.” / “Invoice ref”; for informal receipts use the receipt number; MISSING if there’s no identifier at all
• Issue date and due date in YYYY-MM-DD
• Amounts always with . as decimal separator, no thousands separator
• Watch European thousand/decimal swap (1.250,00 is 1250.00, not 1.25)
• Currency column always filled (EUR / GBP / USD), one PDF can be in a different currency than the rest
• Genuinely missing fields → MISSING (uppercase, no quotes), never guess

Claude reads CLAUDE.md automatically when you start it in any folder. Next month:

1. New folder, drop the PDFs in
2. cp ~/Desktop/claude-lab/invoices-may/CLAUDE.md .
3. claude --dangerously-skip-permissions and just say: “extract all invoices to invoices.csv”

Claude already knows every rule. The same CSV format lands every month, which means your accounting-tool import mapping keeps working without re-touching it.

Step 07

Level up — a review screen instead of a raw CSV

Optional: the CSV is for the accounting tool; this is for the human who signs off on it. One prompt:

Read invoices.csv and the files in invoices/. Build a single self-contained file review.html that I can open by double-clicking — no server, no internet, no external libraries. A table of every extracted row; clicking a row shows the original invoice’s text beside it so I can eyeball the extraction against the source. Highlight rows in amber when any field is MISSING and in red when net + VAT ≠ gross or the VAT ratio is neither 24% nor 20%. A counter at the top: N invoices, M need attention. Accent colour #007a8a.

Double-click it, work the amber and red rows, and the monthly sign-off becomes minutes of looking at exactly the rows that deserve attention — instead of scrolling a spreadsheet hoping wrong numbers jump out.

The pattern to remember — any output Claude can produce as text, it can also produce as a small interactive page. The upgrade prompt is always the same shape: “turn this into a single self-contained interactive HTML file I can open by double-clicking.” It works on every use case in this hub.

Next use case →

Generate a personalised first-week checklist for a new joiner

A new joiner starts Monday. HR sends them a generic 40-item checklist that scares everyone. Claude reads the job description, the team's standard onboarding doc, and the new joiner's CV — then produces a tailored 12-item week-one plan they can actually finish.

Open it → Back to all 15 use cases