hybrid_fixture

Name: hybrid_fixture
Author: andrewstellman

69stars7forks

How to install hybrid_fixture?

andrewstellman/quality-playbook/hybrid_fixture

$npx -y skills add andrewstellman/quality-playbook --skill hybrid_fixture

Installs into the current project.

Summary

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 un

Files · 1

View on GitHub

SKILL.md

1	# Receipt Reconciler (Hybrid classifier fixture)
2
3	A skill plus Python orchestrator that helps a small business reconcile
4	expense receipts against their bookkeeping ledger. Used by
5	`bin/tests/test_classify_project.py` as a known-Hybrid target for the
6	v1.5.3 project-type classifier: SKILL.md at fixture root AND substantial
7	code under `bin/`.
8
9	The fixture is a synthetic skill, not a real one in production use. Its
10	SKILL.md prose is shorter than its code LOC; the heuristic should classify
11	it as Hybrid (the band where SKILL.md exists alongside dominant code).
12
13	## Overview
14
15	Receipts arrive as photos, PDFs, or paper that has been scanned. The
16	ledger is a CSV exported from the bookkeeping system. The skill's job is
17	to match each receipt to a ledger entry, flag mismatches, and produce a
18	reconciliation report the operator can hand to their accountant.
19
20	The skill is hybrid because the matching itself is mechanical (CSV parsing,
21	fuzzy string matching, date arithmetic) but the judgment calls — is this
22	$24.99 charge from the same coffee shop as that $25.04 charge in the
23	ledger, or two separate visits — need a model to decide.
24
25	## Phase 1 — Receipt ingestion
26
27	The orchestrator's `bin/ingest.py` walks the receipts folder, OCRs anything
28	that needs OCR, and emits a normalized JSON record per receipt: vendor,
29	date, amount, last-four-of-card. The JSON is the contract between the
30	mechanical layer and the skill prose.
31
32	The skill's role in Phase 1 is to verify the OCR output: when the OCR
33	confidence is below the configured threshold, the skill is asked to look
34	at the original image and either correct the field or flag it as unread-
35	able. Unreadable receipts go into a manual-review queue.
36
37	## Phase 2 — Ledger ingestion
38
39	The orchestrator's `bin/ledger.py` parses the bookkeeping CSV, normalizes
40	column names, and emits a JSON record per ledger entry. No skill involvement
41	in Phase 2 — the CSV format is fixed, parsing is mechanical.
42
43	## Phase 3 — Matching
44
45	The orchestrator's `bin/matcher.py` proposes candidate matches: for each
46	receipt, the top-K ledger entries by combined date proximity and amount
47	similarity. The skill is asked to confirm or reject each proposed match,
48	with an explanation when a match is rejected.
49
50	The skill's confirmations and rejections are written back through the
51	orchestrator's API so the orchestrator can track which receipts have been
52	matched and which still need human review.
53
54	## Phase 4 — Reconciliation report
55
56	The orchestrator's `bin/report.py` consolidates matched receipts, unmatched
57	receipts, and unmatched ledger entries into a single report. The skill is
58	asked to write the report's narrative summary: which categories had clean
59	reconciliation, which had problems, what the operator should look at first
60	when reviewing.
61
62	## Anti-patterns
63
64	The skill explicitly does not categorize expenses against tax categories,
65	estimate deductibility, or otherwise act as a tax advisor. Those decisions
66	are out of scope and the skill points the operator at their accountant.

People who liked this also viewed

6 picks

Type

hybrid_fixture

69stars7forks

View on GitHub

How to install hybrid_fixture?

andrewstellman/quality-playbook/hybrid_fixture

$npx -y skills add andrewstellman/quality-playbook --skill hybrid_fixture

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Receipt Reconciler (Hybrid classifier fixture)
2
3	A skill plus Python orchestrator that helps a small business reconcile
4	expense receipts against their bookkeeping ledger. Used by
5	`bin/tests/test_classify_project.py` as a known-Hybrid target for the
6	v1.5.3 project-type classifier: SKILL.md at fixture root AND substantial
7	code under `bin/`.
8
9	The fixture is a synthetic skill, not a real one in production use. Its
10	SKILL.md prose is shorter than its code LOC; the heuristic should classify
11	it as Hybrid (the band where SKILL.md exists alongside dominant code).
12
13	## Overview
14
15	Receipts arrive as photos, PDFs, or paper that has been scanned. The
16	ledger is a CSV exported from the bookkeeping system. The skill's job is
17	to match each receipt to a ledger entry, flag mismatches, and produce a
18	reconciliation report the operator can hand to their accountant.
19
20	The skill is hybrid because the matching itself is mechanical (CSV parsing,
21	fuzzy string matching, date arithmetic) but the judgment calls — is this
22	$24.99 charge from the same coffee shop as that $25.04 charge in the
23	ledger, or two separate visits — need a model to decide.
24
25	## Phase 1 — Receipt ingestion
26
27	The orchestrator's `bin/ingest.py` walks the receipts folder, OCRs anything
28	that needs OCR, and emits a normalized JSON record per receipt: vendor,
29	date, amount, last-four-of-card. The JSON is the contract between the
30	mechanical layer and the skill prose.
31
32	The skill's role in Phase 1 is to verify the OCR output: when the OCR
33	confidence is below the configured threshold, the skill is asked to look
34	at the original image and either correct the field or flag it as unread-
35	able. Unreadable receipts go into a manual-review queue.
36
37	## Phase 2 — Ledger ingestion
38
39	The orchestrator's `bin/ledger.py` parses the bookkeeping CSV, normalizes
40	column names, and emits a JSON record per ledger entry. No skill involvement
41	in Phase 2 — the CSV format is fixed, parsing is mechanical.
42
43	## Phase 3 — Matching
44
45	The orchestrator's `bin/matcher.py` proposes candidate matches: for each
46	receipt, the top-K ledger entries by combined date proximity and amount
47	similarity. The skill is asked to confirm or reject each proposed match,
48	with an explanation when a match is rejected.
49
50	The skill's confirmations and rejections are written back through the
51	orchestrator's API so the orchestrator can track which receipts have been
52	matched and which still need human review.
53
54	## Phase 4 — Reconciliation report
55
56	The orchestrator's `bin/report.py` consolidates matched receipts, unmatched
57	receipts, and unmatched ledger entries into a single report. The skill is
58	asked to write the report's narrative summary: which categories had clean
59	reconciliation, which had problems, what the operator should look at first
60	when reviewing.
61
62	## Anti-patterns
63
64	The skill explicitly does not categorize expenses against tax categories,
65	estimate deductibility, or otherwise act as a tax advisor. Those decisions
66	are out of scope and the skill points the operator at their accountant.

People who liked this also viewed

6 picks

Type