Descriptions:
Philipp Schmid of Google DeepMind opens his AI Engineer conference talk with a quick audience poll: nearly everyone uses skills (structured instruction sets) with their coding agents, but almost no one has evals for them. The rest of the session is a methodical case for why that gap is dangerous and how to close it.
Schmid draws on SkillBench — a benchmark covering 50,000+ skills across open and closed models — to establish that well-crafted skills improve agent performance by roughly 15% on average, while AI-generated skills consistently underperform human-written ones. He distinguishes capability skills (temporary, teaching models things they can’t yet do consistently) from preference skills (durable, encoding company-specific workflows and style), and argues evals serve different purposes for each: for capability skills, evals tell you when the underlying model has improved enough to retire the skill; for preference skills, they protect hard-won specificity from model updates.
The bulk of the talk delivers seven concrete skill-writing tips: use progressive disclosure so models only load context they need, define goals and constraints rather than rigid step-by-step workflows, include negative cases to prevent over-triggering, and eliminate no-op instructions that inflate token usage without changing behavior. Schmid credits AI educator Matt for surfacing the no-op problem through a published skill and tweet. The session closes with a framework for building test suites of 10–20 prompts (five happy-path, five negative) before shipping any skill.
📺 Source: AI Engineer · Published July 14, 2026
🏷️ Format: Keynote Launch
