Descriptions:
In this AI Engineer conference talk, Michal Cichra — formerly of Microsoft and Red Hat, now building Spec 27, a new agent testing product at Safe Intelligence — presents a practical framework for capturing software decisions in ways that both human developers and AI coding agents can understand, enforce, and build on over time.
Cichra walks through three complementary documentation practices: Architecture Decision Records (ADRs) that explain why a codebase is structured the way it is and include linting rules to enforce that structure; Product Requirements Documents (PRDs) that capture problem context and user journeys at a lightweight level; and Behavior-Driven Development (BDD) via Cucumber, which makes specs executable and reviewable rather than leaving them as inert markdown. He also covers design system documentation — defining component rules, visual states, and reusable snippets — as a mechanism for enforcing consistent UI generation when AI agents are writing frontend code.
The core insight Cichra returns to throughout the talk is that LLMs have no persistent memory, and human teams forget. Without structured, machine-enforceable documentation, agents produce inconsistent or conflicting outputs across sessions. His enforcement loop — git hooks running the same checks locally and in CI, covering linting, formatting, type checking, architecture rules, and document validation — closes this gap. For engineering teams actively using AI coding agents, this talk offers a concrete, immediately applicable architecture for maintaining consistency at scale.
📺 Source: AI Engineer · Published June 03, 2026
🏷️ Format: Deep Dive
