What it does
It reads a folder of supplier invoices, matches each to its purchase order and delivery note, and returns an exception register showing which invoices can be paid and which need a human.
User and problem
Small-business bookkeepers and office managers who keep payables on a NAS. Invoices arrive as PDFs, photos and forwards, and failures are usually attention, not fraud: an invoice is paid twice, a price drifts from the order, or goods are paid before delivery is proven.
Solution
One folder becomes a decision-ready register: quoted evidence, a fixed exception class, the money at stake, and a confirmation gate before any filing. Nothing is deleted, renamed or paid.
Flow
The user names the invoice and purchase order folders, window, currency and tolerances. The Skill confirms folders, fingerprints candidates, reconciles each invoice and writes the register beside them. The user reviews it, optionally approves filing, then receives counts, held-back totals and confirmation no payment was made.
Differentiation
Existing options extract text (chat assistants, OCR apps) or run cloud payables suites with bank integration. This Skill produces a decision artifact on the device, quotes what it relied on, separates duplicate files from duplicate invoices, and cannot move money.
NAS features
Listing and fingerprinting (steps 1–3): file-workflow get-user-dir, ug-filesystem find, ls, stat, sha256sum, ug-filemgr scan-dedupe. Extraction and matching (4, 6): ug-document page-count, extract, reconcile. Images (5): ug-photo vision-ocr. Register (8): ug-knowledge save-text. Filing (9): ug-filemgr tag-add-file, file-workflow batch-move.
Technical approach
A fixed JSON schema, one document per call, so no invoice biases another. Four expert-agent delegations carry paths, tolerances and expected returns. Reads cover invoice, purchase order and delivery folders; writes go only to the register and, on approval, one filing folder. Documents stay on the NAS: OCR first, then a vision pass only for empty fields, on NAS-hosted models, so no invoice leaves the device.
Built, and not built
Built: SKILL.md and four linked references; passes the organiser's checker 2.0.0 with zero findings. Not built: any live NAS run. Still to test: ug-document return shapes, OCR quality on photographed invoices, reconcile at four files.
Test plan
Case one: twelve invoices, nine clean; expect nine matched, totals agreeing to the orders. Case two: one invoice forwarded twice under different names; expect one possible-duplicate exception and no double count. Edge case: an image-only invoice whose handwritten total defeats OCR; expect a needs-human-review record outside matched totals, no file moved.
Limits
It depends on purchase orders and delivery notes carrying a matchable reference; OCR degrades on poor scans. Tolerances can hide small drift. One currency per run, no bank integration by design, no supplier deduplication.
Markdown only, no third-party components; built with Claude and Codex.
Built With
- claude
- codex
- markdown
- ug-document
- ug-filemgr
- ug-filesystem
- ug-knowledge
- ug-photo
- uliya-skills
