Reorder before you run out, not after

An operations agent that reads sales velocity and stock on hand from Shopify, Square or a CSV, projects the stockout date per SKU against supplier lead time, sizes a 60 day reorder, drafts one purchase order and vendor email per supplier, and stages the coded cost for the bookkeeper, with three separate approvals before anything is sent or recorded

The problem

A small shop usually learns an item is out when a customer asks for it, and learns it over ordered when cash sits in stock that does not move. Reordering runs on the owner's memory of what sells, even though the rhythm shifts every month and nobody counts what is already on its way from the supplier.

How the agent works

1. It starts

   The restock skill in Anthropic's small-business plugin for Claude Cowork starts when the owner asks what to reorder or what is about to run out.

2. The agent works

   It reads sales history and stock on hand from Shopify, Square or NetSuite, or from a sales CSV plus a stock count for shops that count by hand, and states the date of that count because an old count skews every stockout estimate. Its inventory-planner sub skill computes units per day over the last 7 days (the current trend) and 28 days (the stable baseline), removes distortions such as a single wholesale order, treats out of stock days as suppressed demand rather than zero, divides stock by the higher rate for a cautious stockout date, compares it with lead time and sorts every item into past the point of no return, order now, order soon or fine. The default reorder is 60 days of cover plus lead time, adjusted for pack size, minimum order quantity and inbound stock, with a cost on every line; items with no sales in 28 days are never reorder candidates, more than 120 days of cover is flagged as overstock and 90 days without movement as dead stock, with the cash tied up in each. The owner approves, trims or changes the buy list (gate one), then the agent drafts one purchase order per vendor and a matching email in the owner's voice, and each send needs its own yes naming vendor, lines and total (gate two). The ap-processor sub skill then files a coded memo (PO number, total, expected date, account) so the later three way match against the bill and the receiving ticket has something to check, and the coding is approved before anything is written to the ledger (gate three).

3. The result

   No money moves here; paying the bill belongs to a separate skill.

Worked example

The default output is an HTML page that leads with the total spend this round would commit and the count of items about to stock out, then one row per SKU with velocity, stockout date, suggested quantity and an inbound status tag; the purchase orders and import files are attachments, and without an accounting connector the coded memo becomes an import file. Seasonality is applied only with at least a year of history to compare against the same period last year, otherwise the skill asks the owner about their season; when history is thin it reports which SKUs lack data instead of inventing velocity, lead times or stockout dates. After a useful run it offers once to add a recurring Google Calendar block for the reorder review and drops the offer if declined. The source gives no results figures, so measure your own: stockouts per month and the cash sitting in dead stock.

Sources

• https://raw.githubusercontent.com/anthropics/knowledge-work-plugins/main/small-business/skills/restock/SKILL.md
• https://raw.githubusercontent.com/anthropics/knowledge-work-plugins/main/small-business/skills/inventory-planner/SKILL.md
• https://raw.githubusercontent.com/anthropics/knowledge-work-plugins/main/small-business/README.md