About this template
The autoresponder answered the wrong question confidently
Now your customer has to argue with a robot before reaching a human. You converted a support cost into a churn cost.
Deflect the question, not the customer
The goal is that they never needed to write in — because the answer was findable. Not that they can't reach you.
So every drafted article ends with a route to a human, stated plainly. An article that dead-ends is a complaint being manufactured.
And some ticket types are marked DO-NOT-DEFLECT outright: billing disputes, cancellations, "this broke my…", anger, and any second contact about the same issue. Those are conversations. Deflecting an angry customer is how a support saving becomes a churn cost.
Volume × answerability, not volume alone
A ticket type only deserves an article if it has one correct answer that doesn't depend on the account.
"Where is my order" is high volume and account-specific — the answer is a lookup. So it tells you plainly: this needs self-serve tooling, not a help article. Drafting a page that can't answer the question is how help centres get ignored.
Each cluster is scored on volume, one-answer-serves-everyone, and answer stability — something that changes monthly will rot, and it's flagged.
It clusters by the question, not your tags
Tags reflect how staff filed a ticket. That's rarely how the customer thought about it, and the difference is where the useful clusters hide.
Their words are the keyword research
Articles are titled as the customer's own question — "When will my order ship?", not "Shipping Policy". You have hundreds of real queries in that export; that's better search data than guessing.
One question per article. Answer in the first sentence, no brand-voice throat-clearing.
What you supply
Your own Claude Code login, plus a ticket export in ~/deflect/tickets/ — subject and
body, from any helpdesk or a CSV.
Add your existing policies. Without them, policy answers come back
[POLICY NEEDED] rather than invented — which is correct, and doubles as your gap
list.
No helpdesk API key, which is also why nothing can auto-reply.
Where this stops
It analyses and drafts. It does not publish to a help centre, enable an autoresponder, close a ticket, or reply to a customer.
It won't estimate a deflection rate — it has no data on what customers do after reading — won't invent a policy answer, and won't say a ticket type is resolved.
Verified on build: the skill, the working tree, the NEVER_PUBLISH_OR_AUTOREPLY guard,
the DO-NOT-DEFLECT list, the volume×answerability rule and the article rules are
present on a fresh fork. Running it on your own tickets is yours to do.