The Orchestrator controls the entire build process. It runs through 7 phases in a fixed order. Each phase produces artifacts that subsequent phases consume. An error in Phase N prevents Phase N+1 from starting.

The system is built as a Pipeline — not a DAG, not an event loop. The linearity is intentional: it simplifies debugging and guarantees data availability.

The system decides per topic and audience how deep the content should go. To do this, it computes a Helpfulness Score (HS). A topic like “CSS Variables” receives a higher HS for developers than for executives — same topic, different scores.

Each audience gets its own build pipeline. The pipelines run in parallel and independently of each other. Within a pipeline, levels are generated sequentially (each level needs data from the previous one), but topics and languages can be generated in parallel.

The architecture distinguishes two agent types: Pipeline Agents (one per audience) and File Agents (one per output file).

Every generated HTML file is fully self-contained — no external stylesheet, no CDN dependency. The entire design is embedded as CSS variables in each file. Three brand colors and three font families form the visual foundation.

The system generates four navigation components that together provide complete orientation. In addition, overview pages include an Audience Switch and every page features Language Flags.

Every generated page follows a fixed structure. There are no “bare” sections — everything is wrapped in .module containers. Within each module, a nested hierarchy ensures consistent layouts across all pages and levels.

Hero sections on deep-dive pages additionally contain a page-overview component with anchor links to the sections.

The skill validates every output against a 28-point checklist. Each item is a hard gate — if a check fails, the affected file is reworked before it enters the final output. The checklist covers structural, content, and visual criteria.