← ContentsThe Consilience Student Knowledge PackAbout

Every episode of the Chronicle needs a face. Every bookcast needs a cover. Every Spark article needs an image that does more than illustrate — an image that invites. The image generation system produces these at scale, rendering from structured prompts that know more about the content than any human art director could hold in memory at once.

The primary engine is Gemini 2.5 Flash Image, with DALL-E 3 as fallback. The preference for Gemini is not ideological. It is practical — lower latency, lower cost, and output that aligns more consistently with the Danish Modern palette without requiring extensive post-generation correction. When Gemini is unavailable, DALL-E steps in with a compatible prompt format and produces results that, while stylistically different, remain within the visual language of the platform.

Authentication runs through three methods, tried in order. First, Workload Identity — the cleanest path, available when the calling service runs on Google's infrastructure with a service account already bound to the runtime. Second, Service Account — a JSON key file, the standard fallback when Workload Identity is not available. Third, Application Default Credentials — the ambient auth that the Google Cloud SDK discovers from the environment. Three methods means the system never fails because of an auth configuration someone forgot to update. It tries each door in sequence and walks through the first one that opens.

The prompts are structured, not freeform. A Chronicle episode prompt carries character DNA — physical description, emotional state, lighting preference, the palette constraint that keeps every image within the Danish Modern family. A bookcast cover prompt carries the book's title, its primary themes, its emotional register, and the requirement that the image feel like an invitation to listen rather than a demand to look. A Spark cover prompt carries the article's thesis, its tone, and the constraint that the image must read clearly at thumbnail size. Telegram backgrounds are generated from the same system but optimized for the constraints of a chat interface — darker, lower contrast, nothing that competes with message text.

The character DNA is the hardest part to get right, and the most important. A character described inconsistently across episodes fractures the reader's trust. The prompt construction system maintains a canonical description of every Chronicle character — updated as the story progresses, never contradicting what came before — and injects it into every image generation call that involves that character. The Art Director persona, a lightweight meta-intelligence that reviews every generated image before it ships, checks for consistency against that DNA. If a character's hair color drifted, if the lighting doesn't match the scene's time of day, if the palette slipped toward something that isn't Danish Modern — the image is regenerated. The cost of a bad image is not the generation credit. It is the broken spell.

Open Graph images — the cards that appear when someone shares a link — are generated by the same pipeline with an additional constraint: they must be legible at 200 pixels wide on a social feed. The title must be large enough to read. The visual must be simple enough to recognize. This is not a creative decision. It is a distribution decision. A Chronicle episode that cannot be shared is a Chronicle episode that will not be read.

23. Image GenerationListening