Descriptions:
Margot Vanlar, an applied AI engineer at Anthropic in London, leads this breakout session from the Code with Claude event series on systematic prompt engineering for production systems. Rather than presenting a generic list of dos and don’ts, Vanlar structures the talk around two realistic scenarios practitioners encounter: maintaining a legacy prompt that has degraded after a model migration, and building a brand-new prompt for an agentic use case from scratch.
The working example throughout is a customer support bot for a fictional telecom company called Meridian Mobile. Vanlar uses a five-case evaluation suite — covering control cases, proration billing calculations, policy-edge handling, escalation logic, and information transparency — to show how evals provide the diagnostic rigor needed to distinguish model capability gaps from fixable prompting issues. The session walks through a live iteration loop: cleaning up a cluttered, multi-contributor system prompt, isolating individual failure modes, and making targeted fixes while monitoring for regressions across the full eval set.
Key lessons include recognizing when older model-specific patches have become counterproductive with newer Claude models that follow instructions more precisely (a “hotspot” case where over-cautious instructions caused unnecessary deflection), and using structured outputs as a programmatic forcing function for output consistency. Vanlar also covers when to reach for tools like Claude’s extended thinking for complex multi-step reasoning versus standard prompting, making this a useful reference for engineering teams actively maintaining Claude-powered production systems.
📺 Source: Claude · Published May 22, 2026
🏷️ Format: Tutorial Demo
