The payment failure gets worked before the six how-do-I questions

What is the exact prompt?

"Every hour during business hours, triage every unassigned conversation in [support tool]. Severity rubric: S1: data loss, security concern, payment or checkout failure, or the product down for a whole account; S2: a core feature broken with no workaround; S3: broken but a workaround exists, and you can name it; S4: question, feedback, or feature request. Routing: S1: post the ticket link and a two-line factual summary in [#incidents], assign per the on-call schedule in [location], and mark it urgent; S2: assign to [engineering queue] with reproduction steps extracted from the ticket into the bug tracker; S3 and S4: tag by product area and draft a first reply that answers from our help docs, citing the specific doc page you used; if no doc covers it, say so in an internal note instead of improvising an answer. Hard limits: never close a ticket, never promise a timeline or a refund, never reply to anything tagged legal or press. Route those to [person]. End each pass with counts per severity and the three oldest unresolved S2s."

The doc citation requirement is the anti-hallucination mechanism. An answer that must name its source page either has one or admits it does not.

Why drafts for S3 and S4 but summaries for S1?

The two ends of the rubric fail differently.

A slightly-off S4 draft costs an edit. An S1 handled with anything but immediate human attention costs trust at the moment a customer is measuring you. So the job inverts with severity: at the bottom, a complete cited draft a human approves; at the top, surface it loudly, summarize factually, get out of the way.

The two-line S1 summary is deliberately factual, what broke, for whom, since when, because a panicked paraphrase spreads faster than the ticket. The never-promise rule sits under both: timelines and refunds are policy. Sends pause for approval, and the Pylon integration and Front integration keep the loop inside the queue tool.

Read every call for the first month

Week one you read every severity call, every draft, every routing decision.

The rubric will be wrong before the companion is, because real tickets straddle categories. Each straddle is a rubric edit: "a broken export is S2 for accounts on the API plan, S3 otherwise". By week three the edits taper.

The oldest-S2 list is the management artifact. S1s take care of their own visibility and S4s do not matter individually, but aging S2s are what FIFO queues and human attention both lose. Three named tickets at the end of every pass makes that decay visible. The counting counterpart is the weekly digest pattern that customer success teams run, and deeper patterns live in support ticket triage automation.

Test the S1 path with a fake ticket

File a synthetic payment-failure report and watch.

Does it land in the incident channel inside the hour, with the right on-call assigned, summarized accurately? The escalation path must not fail silently, and it is the part you stop thinking about once the routine feels reliable. Check it before it feels reliable, and quarterly after.

For drafts, sample against the cited docs: does the draft say what the doc says? Citation drift is subtler than fabrication and shows up when docs are stale, which makes the "no doc covers this" notes a documentation gap list generated free. Route those into the docs backlog through the Linear integration.

Frequently asked questions