I'm building a system where anyone on a team can generate on-brand merchandise by typing a message into WhatsApp. No designer in the loop. No feedback rounds. Brief goes in, branded mockup comes back.
The team is Starpath Robotics. They make space solar panels. The brand aesthetic is specific and deliberately weird: imagine declassified aerospace documentation from a parallel universe. Victorian patent drawings meets NASA mission patches meets contemporary streetwear. Three colors: charcoal black, warm antique white, copper red. Michroma typeface. Always uppercase headings. Every design has to reward close inspection -- micro-details, annotations, real scientific formulas in the margins.
Try getting an AI to reliably produce that.
The pipeline
Someone posts /ideator make a t-shirt with the Starlight Air panel in orbit in the WhatsApp group. If they attach a reference image, Claude's vision model reverse-engineers the style. Claude generates a concept that follows the brand system -- detailed enough to be an image prompt. Grok generates the image. The system composites it onto a t-shirt mockup and posts it back to the chat.
The trick that makes it work is the style guide. Not the pipeline, not the prompting -- the style guide. It's a markdown file that the AI reads fresh on every request. Two hundred to four hundred words of image prompt specifications, collateral formats, constraints. A companion document has few-shot examples: concrete descriptions of what "good Starlight design" looks like. The examples teach the AI the difference between generic space art and the specific Victorian-meets-physics thing this brand needs.
Update the style guide and the next design reflects the change immediately. No retraining, no fine-tuning. The document is the model's understanding of the brand.
Physical things
Beyond the digital pipeline, I've been designing physical products. Third iteration of the product packaging -- a 1255x455x90mm box with a gradient treatment that took three versions to get right. A blue energy-wave variant for the Starlight Air subbrand. A branded tape measure. Poster designs. T-shirt mockups with "Built for Space" typography and circular badge patches.
I also built Python scripts that generate brand guide PDFs from a canonical YAML-formatted style guide. Three variants: a 130-page comprehensive guide, a 5-page landscape compact, and a 6-page gallery version. All use ReportLab. All render correct colors, typography, and logo placement. The compact and minimal versions are production-ready.
What works, what doesn't
The brand guide generation works. The WhatsApp bot handles auth and commands. Claude's vision analysis and concept generation are solid. What needs work: Grok's image generation endpoint, mockup composition, and the web UI. What hasn't been attempted at all: batch generation, poster and patch pipelines, integration with actual production and fulfillment.
I keep circling back to the same conviction: the brand system should enforce consistency, not a human reviewer. If the style guide is clear enough, the AI should be able to hold the line without someone checking every output. We're not there yet. But the architecture assumes we will be, and that assumption shapes every decision about where the intelligence lives -- in the document, not in the model.