Hybrid Fixture

Receipt Reconciler (Hybrid classifier fixture)

A skill plus Python orchestrator that helps a small business reconcile expense receipts against their bookkeeping ledger. Used by bin/tests/test_classify_project.py as a known-Hybrid target for the v1.5.3 project-type classifier: SKILL.md at fixture root AND substantial code under bin/.

The fixture is a synthetic skill, not a real one in production use. Its SKILL.md prose is shorter than its code LOC; the heuristic should classify it as Hybrid (the band where SKILL.md exists alongside dominant code).

Overview

Receipts arrive as photos, PDFs, or paper that has been scanned. The ledger is a CSV exported from the bookkeeping system. The skill's job is to match each receipt to a ledger entry, flag mismatches, and produce a reconciliation report the operator can hand to their accountant.

The skill is hybrid because the matching itself is mechanical (CSV parsing, fuzzy string matching, date arithmetic) but the judgment calls — is this $24.99 charge from the same coffee shop as that $25.04 charge in the ledger, or two separate visits — need a model to decide.

Phase 1 — Receipt ingestion

The orchestrator's bin/ingest.py walks the receipts folder, OCRs anything that needs OCR, and emits a normalized JSON record per receipt: vendor, date, amount, last-four-of-card. The JSON is the contract between the mechanical layer and the skill prose.

The skill's role in Phase 1 is to verify the OCR output: when the OCR confidence is below the configured threshold, the skill is asked to look at the original image and either correct the field or flag it as unread- able. Unreadable receipts go into a manual-review queue.

Phase 2 — Ledger ingestion

The orchestrator's bin/ledger.py parses the bookkeeping CSV, normalizes column names, and emits a JSON record per ledger entry. No skill involvement in Phase 2 — the CSV format is fixed, parsing is mechanical.

Phase 3 — Matching

The orchestrator's bin/matcher.py proposes candidate matches: for each receipt, the top-K ledger entries by combined date proximity and amount similarity. The skill is asked to confirm or reject each proposed match, with an explanation when a match is rejected.

The skill's confirmations and rejections are written back through the orchestrator's API so the orchestrator can track which receipts have been matched and which still need human review.

Phase 4 — Reconciliation report

The orchestrator's bin/report.py consolidates matched receipts, unmatched receipts, and unmatched ledger entries into a single report. The skill is asked to write the report's narrative summary: which categories had clean reconciliation, which had problems, what the operator should look at first when reviewing.

Anti-patterns

The skill explicitly does not categorize expenses against tax categories, estimate deductibility, or otherwise act as a tax advisor. Those decisions are out of scope and the skill points the operator at their accountant.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: andrewstellman
• Source: andrewstellman/quality-playbook
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.