A data engineering team at Halodoc shares seven production-tested patterns for building reliable AI skills (agent workflows) that automate warehouse onboarding, replication setup, and cost analysis. The patterns cover: loading only the relevant runbook per request to avoid context bleed, probing internal systems before asking users for input, explicitly mapping each workflow step to a specific tool/database, using exactly three human-in-the-loop confirmation gates, writing hard constraints as short imperatives derived from real production failures, defining explicit per-mode parameter sets to prevent hallucinated fields, and specifying output format with a concrete filled-in example. Post-deployment observations showed that replacing free-text user inputs with system-queried lists eliminated an entire class of errors, and showing intermediate reasoning steps increased engineer trust. A checklist of six pre-ship questions is provided as a practical framework.
