# Sources One required backbone, everything else deepens it. The skill works on day one with nothing connected at all. | Source | Status | What it gives | |---|---|---| | A CRM (HubSpot, Monday.com, Salesforce, Zoho CRM), a payments connector (PayPal, Square, Stripe), **or** a storefront (Shopify, Square) | Required — one of any of the three categories | Tickets and feedback; disputes and refunds; or order history (who bought what and when) plus fulfillment and refund patterns | | Support desk (Zoho Desk) | Optional | Tickets by customer, priority, and status — complaints in the customer's own words | | Gmail | Optional | Complaint and praise language in threads | | Web research | Always available | Publicly visible reviews on Google, Yelp, Facebook | | Pasted or exported reviews | Always available | The best source there is, with dates and ratings | --- ## Public reviews **Preferred: an export.** Google Business Profile and Yelp both let the owner download their reviews. An export carries the rating, date, reviewer name, and full text, and it includes reviews the public page paginates away. Ask for one the first time this skill runs; it takes the owner two minutes and improves every run afterward. **Fallback: what is visible on the web.** Fetch the business's public review pages and read what renders. This is genuinely useful and genuinely partial — pages truncate, paginate, and reorder. Say so in the Sources section rather than presenting a partial pull as a complete one. **Also fine: pasted text.** The owner drops reviews into the chat. Tag them `[Review]` like any other source and use them. When counting, never total a rating average across sources. Google's average and Yelp's average are different numbers about different populations, and an owner who repeats a blended figure to a customer will be wrong. --- ## Disputes and tickets **PayPal disputes.** Fetch disputes opened in the window. PayPal throttles aggressively on wide date windows. On the first rate-limit error, stop, record `PayPal: rate-limited — not included` in Sources, and continue. Do not retry silently; suggest a narrower window instead. **Stripe disputes and refunds.** `GET /v1/disputes` created in the window (via `stripe_api_read`), with `reason` and `status`; `GET /v1/refunds` for the same window, since a run of refunds against one product is a theme before it is a review. A dispute's `evidence_details.due_by` is a deadline — surface any open one in the "do these three things" list. **Square disputes.** The disputes endpoint through `make_api_request` for the window. Same treatment as Stripe. **HubSpot tickets and feedback.** Fetch open and recently closed tickets, plus any feedback submissions. Zero tickets is a legitimate state for a new or small portal — record `HubSpot tickets: 0` and move on. Only flag a connector problem when authentication itself fails. **Zoho Desk tickets.** `getTickets` for the window (`receivedInDays` 30 or 90 to match it; sort by `recentThread`), then `getThreads` on the ones whose subject reads as a complaint, for the verbatim. `getTicketsMetrics` gives the open, overdue, and on-hold counts for the header. Open tickets with priority High belong in the Disputes and tickets count, not only in the theme list. `getTickets` rejects an `include` parameter — ask for extra fields with `fields`. --- ## Email Search the window for threads containing complaint and praise language. Seed list, deliberately wide: ``` refund cancel unhappy issue problem disappointed frustrated broken late slow wrong missing never again took forever thank you lifesaver recommend great job ``` Customers rarely use the word "disappointed." They write "took forever" and "never again." Over-inclusion is cheap here because theme extraction filters the noise; a narrow keyword list silently loses whole themes. Pull the subject line and one or two sentences of the customer's own words — enough to quote verbatim later. --- ## Shopify Covered in `churn-signals.md`. In short: orders, fulfillment times, and refunds, read as behavior rather than opinion. --- ## Reporting what you got Every run ends with an honest Sources block. Counts, and named failures: ``` Sources pulled — Apr 28 to Jul 27 Google reviews (export) 14 Yelp (web, partial) 6 — page shows most recent only Stripe disputes 3 Stripe refunds 7 HubSpot tickets 0 Zoho Desk tickets 12 — 2 open, priority High Gmail threads 11 Shopify orders 182 ``` Name the connector that was actually read on each line (the example above is a Stripe and Zoho Desk shop; a PayPal and HubSpot shop lists those). The dash matters as much as the numbers. A source that is missing looks exactly like a source with nothing bad in it unless it is named.