·Question
Our support draft assistant told two customers this week that we have a feature we do not have, and the wording was perfect both times.
The setup: a ticket arrives in our inbox, we pass it to Claude along with our tone guide, it writes a draft reply, and an agent reads it and sends. On Tuesday it told a customer they could schedule a report to send automatically every Monday. We have no scheduling of any kind. On Thursday it told somebody our API supports webhooks. It does not. Both replies read exactly like the fifty correct ones that went out around them, which is the dangerous part: there is no tell, no hedging, nothing that looks like a guess. An agent caught the second one. The first one went out. How are people stopping this, short of checking every line against the documentation?
This answered it
Give it the documentation, and tell it the documentation is the only thing it may answer from. Ours gets the relevant help centre pages pasted in with the ticket, and one line at the end: if the answer is not in the text above, do not guess, write that you will check with the team, and flag it for a person. It still slips occasionally, but it went from weekly to about once a month.
That is exactly the gap. We gave it the tone guide and never the actual product documentation, so it had nothing to be right from. We were asking it to sound like us, not to be correct.
And nothing sends unread while it is drafting for customers. Draft is fine, auto send is not, not until you have months of evidence behind you.
Same thing here with policy wording, and the same fix. If it cannot find the clause in the document in front of it, it has to say so rather than produce something that reads exactly like a clause.