Data Enrichment Skill
gotcs108/data-enrichment-skill
Enriqueça e verifique dados de negócios locais ao vivo em CSV
Ver no GitHubSkill para Claude Code que transforma buscas fuzzy ('todos os X na área Y com contatos') em CSVs verificados ao vivo de mercados de negócios locais. Pipeline de 5 estágios com reconciliação de dados e verificação live via browser, reduzindo inconsistências entre diretórios desatualizados.
Instalação
npx convez-ai gotcs108/data-enrichment-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Gerar listas de prospectos de estacionamentos com contatos verificados
- Montar diretórios de negócios locais com ratings live do Google Maps
- Reconciliar dados conflitantes entre agregadores e fontes ao vivo
- Extrair dados de operadores, telefones, horários e acessibilidade
Veredito

Ferramenta sólida e principiada para quem precisa dados de mercado local confiáveis. A ênfase em verificação live e rastreabilidade de fontes resolve um problema real — mas exige integração com Claude Code e conhecimento de pipelines de dados para aproveitar ao máximo.
README
A tradução para o português deste README ainda está na fila.
data-enrichment-skill
Two Claude Code skills for turning a fuzzy ask — "give me all the X in area Y, with how to contact them" — into a clean, sourced, live-verified CSV of a local-business market.
The pitch in one line: directory data rots and aggregators copy each other's stale rows. The value here is reconciliation + live verification, not "I scraped a page." On a real run the live browser check corrected paid directories by >1.5 stars on multiple businesses — that delta is the whole point.
Skills
local-data-enrichment (the engine)
Generic, vertical-agnostic. A 5-stage pipeline:
- Seed —
WebSearchto discover the aggregator pages that already list many entries. - Fan out —
WebFetchthose directories into structured lists. - Enrich — one read-only research subagent per cluster fills in operator, phones, hours, rates, website/reservation, public-vs-tenant, accessibility (writes "unknown", never guesses).
- Reconcile — dedupe, resolve aliases (same place under a generic name + its operator name), flag address/phone conflicts, prefer live over directory.
- Live-verify — drive a headful browser (the
agent-browserCLI) over Google Maps to pull live star ratings. Shipsgmaps_ratings.js, a hardened extractor. (Signed-out Maps exposes ratings but gates review counts/text.)
Output: one CSV, a source per row/field, every gap marked unknown.
parking-finder (a vertical on top)
The parking instance of the engine: operator contact channels (Indigo / clicknpark / Impark / municipal agencies), the parking aggregator set (BestParking, Parkopedia, ParkMe, YellowPages), EV/accessibility columns, and an optional Kijiji P2P / classifieds supply scan — with the hard-won caveat that a " parking" keyword search is mostly condos that mention parking.
Install
Copy a skill into your Claude Code skills directory (user-level or per-project):
# user-level (all projects)
cp -r skills/local-data-enrichment skills/parking-finder ~/.claude/skills/
# or per-project
cp -r skills/local-data-enrichment skills/parking-finder <your-repo>/.claude/skills/
Then in Claude Code: "find all parking lots in with contacts and ratings into a CSV"
(triggers parking-finder), or "give me every in as a lead list"
(triggers local-data-enrichment).
Principles baked in
- Cite or say unknown — no speculation; every field traces to a source.
- Live beats directory — when they conflict, trust the browser-verified value, note both.
- Read-only / no republish — extract review themes to inform your own copy; never scrape and republish someone's copyrighted review text or photos.
- No silent truncation — if coverage is capped, say what was dropped.