2026AI product deliveryTransportation, contractClaude Routine

Outcome

4hrs

a week back for a senior support engineer, once runbooks stopped needing hand upkeep.

651resolved tickets read in a 180-day proof run

14recurring problems found and reviewed with the engineer

40percent fewer senior architect escalations, via a support agent built on these runbooks

An agent keeps our runbooks current while we sleep

Context

A data support team kept its fixes in Confluence runbooks. The fixes were verified, but the runbooks drifted out of date as fast as tickets closed. Keeping them current fell to one senior engineer.

Problem

A senior engineer was losing 20% of his week to tickets and runbook upkeep.

The work never ended, and stale runbooks meant the team couldn’t self-serve, so questions kept routing back to the people who knew the answers.

The quiet tax: a stack of resolved tickets beside one senior person. Every fix that never reaches a runbook becomes tribal knowledge held by one senior person.

FIG 01: Every fix that never reaches a runbook becomes tribal knowledge.

Approach

Let resolved tickets keep the docs honest, and make the agent safe before making it autonomous.

01Read resolved Jira tickets weekly and update runbooks to match.

02Humans own the narrative pages. The agent only updates the evidence.

03Create a new runbook only when three or more tickets share a problem.

The Curator dataflow: Jira tickets flow one way into the weekly Curator agent, which writes runbook evidence to Confluence and a weekly digest to Slack.

FIG 02: Data moves one way. Jira stays the source of record.

What I did

Role: Technical Project Manager · Transportation company (contract) · May 2026–Present

Built the Curator as a Claude Routine, after two earlier attempts as Rovo agents. The prompt and config live in version control, and a lessons log corrects its judgment between runs without a redeploy.

Ran a 180-day proof run first, reviewed all 14 clusters with the engineer, then piloted the weekly routine for two months with weekly output reviews.

Outcome

The engineer got four hours a week back, and the runbooks became more trustworthy because every fact on them is backed by ticket evidence. Those runbooks then grounded a Rovo support agent, released only after it cleared 95%+ accuracy, that cut senior architect escalations by 40%.

The proof run: 651 tickets read over a 180-day window, clustered at three or more of a kind, with one-offs discarded, finding 14 clusters.

FIG 03: 180 days of resolved tickets, clustered into 14 recurring problems.

What I would do differently

I’d have tested Rovo’s limits in a day, not two full builds. A quick spike would have pointed to a Claude Routine sooner.

← All workDelivered work Next →Intranet at National Scale