Descriptions:
Priscila Andre de Oliveira, senior software engineer at Sentry, shares a data-backed look at how she actually uses AI in her day-to-day work on a codebase with over 15 years of history, 100,000 dependent organizations, and roughly 100 pull requests merged every day. The finding that anchors the talk is striking: after having Claude analyze 116 of her past sessions, 67% of her AI usage was categorized as comprehension — understanding the codebase — while only 2% was code generation.
This inversion of the popular narrative around AI-accelerated coding leads to her main contribution: a custom Claude skill she calls “Catch Me Up,” a structured prompt template that organizes comprehension queries into six modes — architecture, convention, feature trace, syntax, testing, and history. She explains why deep code understanding is a prerequisite for steering AI agents correctly, and how even Claude can go off-track if the engineer doesn’t have sufficient context to recognize and correct the mistake.
Oliveira also shares Sentry’s broader internal AI ecosystem, including Warden (a code review agent), Junior (a Slack bot that triages bug reports and opens PRs), and an internal tool called Abacus that tracks company-wide AI usage. The talk is one of the more honest first-person accounts of what high-volume AI-assisted engineering looks like inside a serious production environment.
📺 Source: AI Engineer · Published May 27, 2026
🏷️ Format: Workflow Case Study
