Addy Osmani on Making AI Agents Act Senior

Addy Osmani published a set of Claude Code slash commands that force AI agents through proper engineering phases – spec, plan, build, test, review, ship – instead of letting them sprint straight to implementation. The core insight is that agents behave like junior engineers, taking the shortest path to a working diff while skipping the invisible work that separates production code from prototypes. What makes these skills interesting is the anti-rationalization tables: pre-written rebuttals for the plausible-sounding excuses agents generate to skip steps. The whole thing encodes Google-flavored engineering norms like Hyrum’s Law, 100-line PR limits, and trunk-based development into actionable workflows. Between this, Augment’s AGENTS.md research, and the growing library of CLAUDE.md templates circulating on GitHub, a real discipline is forming around constraining agent behavior rather than just prompting it.