Jessmatth

Invoice Auditor — AI skill for Claude Code

AI community

A Claude skill that extracts invoices and proves the extraction is correct.

How to install Invoice Auditor

This entry records only its repository, not the path inside it, so there is no exact command to give. Open Jessmatth/invoice-auditor and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Invoice Auditor does

A Claude skill that extracts invoices and proves the extraction is correct. Model extracts, code audits the arithmetic. 0 false positives on 76 real invoices; found 2 mislabeled records in a public training dataset.

Alternatives in AI

Design Md Figma — Figma plugin that extracts local style guidelines and generates DESIGN.md and SKILL.md files for AI tools such 147 ★
Tma1 — Local-first observability your agent reads back 116 ★
Aurora — Aurora - independent community smart home orchestrator 103 ★

README

invoice-auditor

A Claude skill that extracts invoices **and proves the extraction is correct.**

Most document-extraction tools return a confident JSON blob. You have no way to know whether the model misread a column, lost a thousands separator, or transposed two digits, and the model will report high confidence either way. Self-reported confidence is the model grading its own homework.

Invoices happen to be checkable. An invoice line carries five numbers (quantity, unit price, net, VAT rate, gross) of which only three are independent. The document over-determines itself, so the arithmetic either closes or it does not, and finding out requires no model at all.

This skill extracts with a model and then audits with code.

47 invoices   44 clean   2 flagged   1 unverifiable

FLAGGED
  inv-2291  line 3: quantity x unit price = 1,240.00 but stated net is 1,420.00
            (digit transposition; page 2, "Freight surcharge")
  inv-2310  net total + VAT does not equal gross total; delta 900.00

Results

Measured on 76 synthetic invoices with human-annotated ground truth, plus 8 hand-built fixtures covering shapes that corpus does not contain. Full method, corpus composition and reproduction in [EVALUATION.md](EVALUATION.md).


False positives on the corpus	0 / 76
Planted arithmetic errors caught	583 / 583
Smallest error reliably caught	5 cents
Real annotation errors found in the public dataset	2
Invoice shapes outside the corpus, hand-verified	13 / 13
Real photographed receipts reconciling cleanly	83 / 100

That last row is the one worth dwelling on. Running the audit over the ground truth of a public dataset used to train extraction models surfaced two mislabeled records one column shift, one stray digit, with no model