Start here

Asking Winston

This page shows you how to write a prompt that gives Winston what it needs, how to read the answer that comes back, and how to spot-check a result before you rely on it.

How a session works

Each conversation is a session; come back to one and Winston remembers its context. New topic, New Chat. Quick lookups return in seconds; bigger reports take 2 to 4 minutes, and the "Thought for…" line shows how long it worked. Type a follow-up while it works and it queues and runs next, or type abort to stop. Sessions can be renamed, pinned, deleted, or shared with the header's copy-link button. Every prompt you send draws credits from the workspace balance (Account and access).

How to write a good prompt

What + which stores + time window + how to rank + what to do with it.

  • Name the store the way your POS does, or say "across all stores". Without a store, Winston reports across all of them and says so; comparisons and per-store reports work better named.
  • Name the time window. "Last 7 days", "this month vs last month", "since June 1". Without one, Winston picks a default and tells you what it picked.
  • Say what you want back. A table, a ranked list, a spreadsheet, a shareable page.
  • State your exclusions. "Exclude drafts, tests, and internal campaigns." Winston follows instructions like that literally.
  • Say "don't publish anything" when you only want to read the answer.
A prompt that combines these: "Rank my top 10 products by revenue for the last 30 days at Downtown, grouped by category, and flag anything with falling momentum. Give me a table."
Good first prompts
  • "List my stores."
  • "What data do you have for Metrc?"
  • "Which products went out of stock at Downtown in the last 14 days, ranked by lost sales? Don't publish anything."

Reading the answer

The answer lands at the bottom of Chat. Read the caveats first: when a data source cannot do exactly what you asked, Winston says so up front, and those warnings are the most valuable part. Raw figures come straight from your systems, pulled live. Calculated numbers, such as estimated lost sales, are worked out from those figures using formulas and assumptions that Winston states. Ask "where did that number come from" and it walks you through the trail. And trust the card, not the sentence: if Winston says it posted, published, or sent something, check Work Audits → Completed, the Library, or the channel itself.

A chat answer showing its Reconciliation and Missing fields / limitations sections
  • 1Missing fields / limitations: read this before trusting the numbers.
A real answer: the headline with its basis, then Reconciliation, a plain summary of what data was pulled, then Missing fields / limitations.

How to verify your results: the 60-second spot check

Four things to check before you act on a number:

  1. Scope line: store, date window, timezone. Winston uses completed store-local days, so today is excluded.
  2. Formula: for example lost sales = pre-stockout daily velocity × days out.
  3. Missing fields / limitations: e.g. "Estimate, not observed loss". A good answer also has a Reconciliation section, a plain summary confirming what it pulled, such as "Inventory log: 13,979/13,979 rows retrieved".
  4. Compare one SKU per store against your POS: check sellable on-hand (not total) for the same completed day. Off? Reply "Recheck [SKU] on-hand for [date]."

Naming stores

Once your POS is connected, Winston knows each location and its Metrc license. Use everyday names ("Northside", "the delivery business"); no codes needed.

Try asking
  • "Only include Downtown and Northside. One tab per store, ranked by days of supply. Give me this as an Excel file."