Data Extraction Pro

When SQL code, business definitions, and warehouse table structures are already scattered across analysts and product ops, the problem is less about whether SQL can be written and more about how to reproduce data requests, evidence checks, generation, and review reliably. data-extract organizes this workflow into a portable bundled skill package: the root SKILL.md acts as the controller, while skills/ contains sub-skills for routing, generation, review, and wiki maintenance. Execution requires reading the relevant files instead of relying on conversational memory.

Core Workflow

• Routing first: each task starts by reading references/exec-runbook.md, then skills/skill-router/SKILL.md produces RouterInput and RouterDecision to select the next next_skill.
• SQL main chain: extraction SQL work covers sql-generator and sql-review-controller, separating generated SQL from review JSON.
• Knowledge-base gate: steps that depend on {{WIKI_ROOT}} require an explicit knowledge-base root. If the host has not configured it and the context cannot infer it uniquely, the user must confirm the path before read/write access. For extraction SQL, once the path is ready, the flow normally fills raw/ and runs /ingest compilation before entering initial_sql.
• State loop: after a sub-skill finishes, state is updated and routing is re-evaluated until next_skill: none and the output is deliverable.

Fit and Caveats

This bundle fits data analysis and product ops workflows that already have SQL and need evidence gates plus review chains. It does not include the double-layer-memory write/retrieval chain and does not bind a warehouse brand; {{SQL_ENGINE}} is project-specific. Before external distribution, review skills/ for example business definitions, table names, and internal paths to avoid leaking environment details into a public registry.