vinyl-dig

Vinyl Dig

Overview

Turns "go see what's in stock at my local shops" into a parallel, read-only crawl with a scored, deduped buy list at the end — never a hallucinated listing. One subagent per shop, dispatched concurrently, each returning structured candidates; a synthesis pass benchmarks against Discogs and tiers the result.

Originated in Johnny's shelf repo (music/.claude/commands/vinyl-dig.md + music/.claude/agents/shop-crawler.md, Chicago shop list). Promoted here 2026-08-21 so it isn't locked to one repo. The crawl mechanics and scoring rubric below are the portable part; the shop list, seed artists, and collection data are not — they live in the calling project and get named at invocation time.

What the calling project must provide

This skill reads project-local files; it has no data of its own. Before running it, confirm the project has:

• A shop directive — one file (name it anything, e.g. chicago-shops.md) listing shops, neighborhoods, tiers, and standing dig rules (formats/condition/budget discipline). Johnny's version also carries a "route clusters" section for planning a physical trip, which is optional.
• A catalog access map — working URLs/API endpoints per shop (Shopify /products.json, a storefront's own catalog JSON, Discogs seller storefronts) plus known gotchas (403s, stale info-only catalogs). Crawling blind against a shop site that blocks scrapers wastes the round.
• A collection ledger (collection.csv or equivalent) — to dedupe against what's already owned.
• A wantlist (wantlist.csv or equivalent) — to flag high-conviction matches.
• A taste/verdict store (taste-db.csv or equivalent) — to never re-pitch a passed record, and to score fit for the rest.

If any of these don't exist yet, say so and stop rather than crawling without a dedupe or scoring basis — a buy list that recommends an owned or already-passed record is worse than no buy list.

Running the dig

1. Read the shop directive and the catalog access map first. Confirm the shop list and per-shop crawl method (native catalog vs. Discogs storefront) are current — a shop's site changes shape often enough that a stale access map produces zero candidates silently.
2. Stay in plan mode and propose the crawl before running it: which shops, which URLs/endpoints per shop, genre/label filters, subagent layout. Wait for approval. This catches a stale or wrong shop list before burning a full crawl on it.
3. Dispatch one subagent per shop, in parallel, each read-only (never writes files), each returning: artist, title, label/cat#, format, year/pressing, price, media + sleeve condition, listing URL, and which taste target it fills. Each subagent's prompt should carry: the shop's catalog/storefront URL, the hard filters (format rules, condition floor, "already owned" check against the collection ledger), and the benchmark instruction (price vs. Discogs stats → steal/fair/toppy).
4. Never fabricate a listing. Every candidate needs a URL actually fetched this session. A shop with a broken or blocked catalog is a reported gap, not a silently empty result folded into "nothing found."
5. Synthesize: dedupe against the collection ledger and taste-db passed rows, flag wantlist matches, score confidence/fit (see rubric below — adapt to whatever scoring the project's own taste-db uses if it differs), and build a tiered buy list ("best cart under $X" if the project sets a budget ceiling, otherwise ranked by desire band).
6. Write the report to a file the calling project expects (Johnny's convention: vinyl-recommendations.md), or ask where it should land if unclear.