Claude Data Skills

A whole data team as one Claude Code command. Pick any subset of eight data roles, in any order; they run as one coherent pipeline over your data — raw files → a typed medallion (bronze/silver/gold) → a governed semantic layer → reports, charts, a trained model, a dashboard, and k8s deploy manifests — all in a single, auditable run directory.

This repo is a Claude Code plugin marketplace. Its first plugin is data-team (invoked /data-team:run). Everything is stdlib Python — no venv, no API key, no cloud; your local Claude is the engine.

Install

One command — npx (cross-platform) or curl (macOS/Linux). Re-run any time to pull the latest:

npx data-team-skill
# — or —
curl -fsSL https://raw.githubusercontent.com/t4tarzan/claude-data-skills/main/install.sh | bash

Or inside a Claude Code session — two commands (this is what the installers run for you):

/plugin marketplace add t4tarzan/claude-data-skills
/plugin install data-team@claude-data-skills

Then invoke it:

/data-team:run

All three do the same thing — add the git-hosted marketplace and install the plugin. npx/curl are just convenience wrappers around the native /plugin commands; no npm dependency is pulled into your project.

Keep it up to date

• Auto-update: run /plugin → Marketplaces → claude-data-skills → Enable auto-update. Claude Code then refreshes the catalog at startup and pulls new versions automatically.
• On demand: /plugin marketplace update claude-data-skills
• The one-liner above is also idempotent — re-running it refreshes to the latest.

npm? Not needed. A git-hosted marketplace is the native, one-command install path for Claude Code — zero npm involvement, versioned by git.

Quick start

/data-team:run
> where is your data?   ./sales_exports        (a folder of CSV/JSON/XML/XLSX; PDF/images optional)
> which roles?          (default: the full spine — architect → engineer → designer → analyst)
> what to answer?       "gmv by platform", "top 5 tier by aov"

You get one run directory, ./data-team-out/<name>/:

bronze/  silver/  gold/       ← typed medallion (SQLite), gold reconciled to the raw fact
semantic/                     ← governed metric DSL (+ the re-aggregation guard)
reports/  visuals/            ← answers with a receipt each + SVG charts
bi/  models/  service/  deploy/   ← dashboard, trained model + served API, k8s manifests
manifest.json                 ← the full audit trail: lineage · PII · reconciliation · conformance · SLOs
governance/report.html        ← one consolidated governance verdict

Zip it and the data, definitions, reports, model, and lineage all travel together.

Run it directly (no install)

Every stage is plain python3 — clone and go:

git clone https://github.com/t4tarzan/claude-data-skills && cd claude-data-skills
python3 skills/run/scripts/run_pipeline.py --slug demo \
    --sources ./examples/retail --metrics-dir ./examples/retail_metrics \
    --ask "gmv by platform" --ask "average order value by tier"
python3 -m unittest discover -s tests     # 51 tests

The eight roles

Pick any subset with --stages; a dependency resolver synthesizes whatever upstream a selection needs (ask for just the Analyst on a folder of CSVs and it builds the medallion first). Governance is cross-cutting — every stage records lineage, PII flags, contracts, and quality checks, aggregated into one verdict.

| # | Role (stage key) | What it does | Consumes → Produces | |---|---|---|---| | 1 | Data Architect architect | Ingest & catalog raw sources; profile types/keys/quality; land typed bronze. | raw → bronze + catalog | | 2 | Data Engineer engineer | Medallion bronze → silver → gold + lineage; every measure reconciled to the raw fact. | bronze → silver, gold | | 3 | Data Designer designer | Governed semantic layer / metric DSL with the re-aggregation guard (a ratio recomputes, a distinct-count → Theta/HLL — never summed). | gold → semantic | | 4 | Data Analyst analyst | NL questions → reports + charts, a receipt per answer (definition, owner, version, SQL, lineage). | gold → reports, visuals | | 5 | BI Engineer bi (branch) | Refreshing, RBAC-scoped dashboard off gold. | gold → dashboard | | 6 | Data Scientist scientist (branch) | Train & evaluate a predictive model (from-scratch ridge OLS) on gold. | gold → model, eval | | 7 | ML Engineer ml (branch) | Package & serve the model behind a stdlib HTTP API. | model → service | | 8 | Data SRE sre (plane) | Deploy any of the above on k8s: manifests + HPA autoscaling + Prometheus observability + SLO alerts. | service → deployment |

Topology: a spine (1→4), two branches off gold (BI; Scientist→ML), and a platform plane (SRE) that wraps any deployable.

Why it's trustworthy

• Sovereign by default — local Claude is the engine; no data leaves the box unless you opt a stage into a cloud model. Mechanical work (transforms, profiling, reconciliation) is deterministic code, so the numbers are reproducible regardless of engine.
• Governed by construction — you cannot produce a gold table without recording where it came from; a ratio cannot be silently summed up a grain; every answer carries a receipt.
• Portable — stdlib only. SQLite medallion, JSON metric defs, a from-scratch linear model, an http.server model API. Optional extras (pypdf, Pillow+pytesseract, PyYAML) enable PDF/image/YAML and degrade gracefully if absent.

Configuration

Drop a data-team.yaml in your working dir (see data-team.example.yaml): choose stages, per-stage engines (local / cloud), the missing-input policy, and bring-your-own artifacts (e.g. supply an existing gold/ and run only the Analyst).

Layout

.claude-plugin/{plugin,marketplace}.json   ← plugin + marketplace manifests
skills/run/SKILL.md                         ← the /data-team:run orchestrator
skills/run/scripts/*.py                     ← the deterministic core (stdlib)
skills/run/personas/<stage>/*.md            ← each role's judgment (the IP)
docs/  schema/  examples/  tests/           ← specs, JSON schema, sample data, 51 tests
install.sh                                  ← the one-line installer

License

MIT © t4tarzan