Bruce Weygandt
All posts

Turning client calls into tickets, with a person in the loop

Claude Agent Skills, Automation, Customer success. 2 min read.

Scheduled Claude skills that pull bug reports and feature requests out of recorded client calls, and why the first version had too much autonomy.

Clients mention bugs and feature ideas on calls all the time, usually in passing and halfway through a conversation about something else. Our calls are recorded and transcribed, but a transcript nobody rereads doesn't fix anything. I built Claude skills to close that gap.

Bugs, in the format support already uses

The bug skill runs on a schedule, reads the calls and meetings from the past 72 hours, and pulls out anything a client described as broken. Each one is posted to a Slack channel, formatted to match the bug report template our support team already uses in Zendesk: description, location, whether it's reproducible, timestamp, environment, steps, expected versus actual behavior, and evidence.

Calls rarely cover all of that. When a field isn't stated, the skill doesn't fill it in. It turns the gaps into a short checklist for the account's CSM to confirm. A ticket with an honest "unknown" is more useful than one with a plausible guess.

There's also a draft mode: a CSM can paste rough notes or a transcript excerpt and get back a report in the same format.

The first version did too much

The first version created Zendesk tickets directly, with no one approving them. It worked, in the sense that tickets appeared. But a model reading transcripts will sometimes take a question for a bug, or file the same issue twice from two calls, and every one of those became a ticket someone had to triage and close.

Now the skill only posts to Slack. A teammate reacts with đź§  to create the ticket or đź’€ to dismiss it, and a small Google Apps Script backend handles the ticket. Review costs one click, and the person clicking knows the client.

Feature requests, counted

A second skill does the same for feature requests, with two differences. It only covers current customers, since requests from prospects and pilots go through a separate process. And it stamps each request with how many times it has come up before, so the product team sees demand rather than a list of one-off asks. It's informational and doesn't create tickets anywhere.

What I'd keep

Let the model do the reading and the formatting, and leave the decision to file with a person. The automation saves the hours spent rewatching calls, and the reaction keeps the ticket queue trustworthy.