Install the data plugin so Claude writes the SQL, runs the analysis, and validates it before you ship

A plugin in the Claude Code / Cowork sense is a bundle of skills, slash commands, and MCP connectors you install as one unit. The data plugin, published in Anthropic's knowledge-work-plugins marketplace, turns Claude into a data analyst that talks to your warehouse. It ships six slash commands — /data:analyze, /data:explore-data, /data:write-query, /data:create-viz, /data:build-dashboard, and /data:validate — backed by five background skills (sql-queries, data-exploration, data-visualization, statistical-analysis, interactive-dashboard-builder) that fire on their own when the work matches.

Why pick this as Thursday's plugin. It is the technical sibling to two plugins already featured: engineering covers the developer's day, productivity covers personal working memory, and data covers the analyst's loop — ask a question in plain English, get SQL that runs on the right dialect, get a chart or a dashboard you can hand to someone. The reason to install the bundle instead of prompting ad hoc is the same as it always is with plugins: the commands have a fixed shape and the skills encode the conventions, so your tenth query looks like your first instead of starting from a blank prompt each time.

write-query. /data:write-query generates SQL for a specific dialect, not generic ANSI that breaks on your engine. Ask for cohort retention on Snowflake and you get common table expressions plus a note on partition pruning; ask the same thing for BigQuery and the generated SQL shifts accordingly. The sql-queries skill carries the dialect differences and the performance patterns, so you describe the question and let the skill handle the syntax that varies between warehouses.

build-dashboard. /data:build-dashboard reads from a connected warehouse or a CSV you drop in and emits a single self-contained HTML file with Chart.js wired in — interactive filters, no build step, no server to stand up. It is the same instinct as the xlsx skill featured yesterday: produce the deliverable, not a description of it. The output is a file you can open in a browser or send to someone who will never run a query themselves.

validate. /data:validate is the command people skip and shouldn't. It runs QA on the methodology and the numbers before you ship: did the join fan out and double-count rows, did a filter silently drop records, does the percentage have the right denominator. An analysis that looks clean and is quietly wrong is worse than no analysis, because someone will act on it. Running validate on anything that drives a decision is the cheap insurance the rest of the plugin earns.

The plugin stays tool-agnostic through a placeholder system. Skills refer to a category like ~~data warehouse rather than naming Snowflake directly, and .mcp.json maps that category to whatever you actually run — BigQuery, Databricks, Hex, Amplitude, Definite, or a Snowflake endpoint you fill in. Swapping warehouses is a one-line config edit, not a rewrite of every skill, which is what makes the same bundle usable across teams that standardised on different stacks.

The trap to avoid is pointing it at production and trusting the first query. The commands generate good SQL, but generated SQL still runs against real tables; read the query before you execute it, and run /data:validate on anything headed for a slide or a decision. Treat the plugin as a fast first draft with a built-in checker, not an oracle. See the try-it block for the install and your first query.