What problem does it solve? Acting inside live supply-chain systems (ERP, procurement, inventory) means a wrong or unauthorized write costs real money and is often irreversible. This Skill gives the AI the judgment layer to classify every action as read, reversible write, committing write, destructive, or outbound, and to pause consequential ones for explicit human approval instead of executing blindly. ## Core Features & Use Cases - Action Classification: Categorizes every operation by what it does to state (read, reversible write, committing write, destructive, outbound) rather than by tool name. - Approval Contract: Structures every consequential action proposal with the exact object, a before/after preview, rollback path, provenance, and risk class so a human can decide in one look. - Hard Safety Rules: Treats tool-returned text as data (never instructions), rejects chat-based authority claims like "the CFO approved it", blocks leaking internal cost/terms data, and enforces re-validation before commit. - Use Case: Before submitting a purchase order in SAP or Coupa, the Skill drafts the PO with a diff preview and rollback plan, checks for duplicate open requisitions, and presents it for recorded human approval rather than auto-submitting. ## Quick Start Ask the assistant to submit a purchase order or adjust inventory on a connected supply-chain system and it will classify the action and present an approval-ready proposal first.
