Every design it draws, the workshop can sew.
Anyone can ask an image model for a beautiful garment. The result is a picture of clothes that do not exist, in fabrics nobody stocks, with details no pattern can produce. Our client, a clothing manufacturer, did not need pictures. They needed new designs they could put into production on Monday. So we built the agent the other way around: first the entire inventory, every fabric, every button, every collar type, every tie and trim, was cataloged as a structured asset library. Then the agent was constrained to design only inside it. It composes new garments from real materials, renders them through image models, and delivers each concept with its bill of materials. The picture is a promise the workshop can keep.
Hermes runtime Asset library Image generation models Bill of materials Supabase
New designs are the lifeblood of a clothing manufacturer and the slowest thing it produces. Each new piece meant a designer sketching, checking fabric availability, matching trims and buttons from memory of the stockroom, revising when a material turned out to be unavailable, and only then producing something a customer could react to. Days per concept, and the concepts were bottlenecked by one thing above all: keeping the entire inventory in one human head.
Generic AI image tools made the problem look solved and left it worse. The generated garments were striking and unproducible, wrong fabrics, impossible seams, buttons from nowhere. Every attractive image created a gap between what could be shown and what could be sewn, and closing that gap manually cost more than sketching from scratch.
The real brief was precision, not creativity: give us volume and variety of new designs, but only ever inside what we stock, what we can source, and what our machines and hands can actually make.
> catalog
The client's full inventory becomes a structured asset library: fabrics with their colors, weights, and behavior, buttons, collars, cuffs, ties, trims. Each asset carries its production notes, what it combines with, what it cannot do.
> brief
A design request enters in plain language: the garment type, the customer's industry, colors, formality, budget class. The brief can be one sentence, the agent asks when it is too thin to design from.
> compose
The agent designs inside the library. It selects a fabric, a cut, and compatible details, and specifies the garment as a structured design: this fabric, this collar, these buttons, this trim. Composition rules encode what the workshop knows, what drapes, what clashes, what cannot be sewn together.
> render
The structured design goes to image generation models with tightly controlled prompts built from the asset descriptions. The model illustrates the specified garment, it does not invent one. Renders that drift from the specification are rejected and regenerated.
> deliver
Each concept arrives as a pair: the visual and its bill of materials, every component traceable to a real stocked asset. A collection of variants is delivered per brief, ready to show a customer or send toward production.
> refine
The client's designer reacts in plain language, "warmer gray, hidden buttons, less formal collar", and the agent revises within the library. Human taste steers, the agent produces.
Layer I · Visual Architecture
One diagram: asset library at the foundation, brief in, composition, rendering, concept plus bill of materials out, feedback loop back to composition. The library is drawn as the ground everything stands on, because it is.
Layer II · Contracts
Two contracts govern the system. The asset contract defines how every material is described, so the library stays consistent as stock changes. The composition contract encodes the workshop's rules, valid combinations, construction constraints, what each fabric can and cannot become. The client's production knowledge, written down for the first time.
Layer III · Technical Diagrams
Library schema, brief interpretation, the composition engine, prompt construction for the image models, render validation against the specification, and the revision loop, specified before implementation. Render validation got particular attention: an image model is a talented illustrator with no respect for inventory, so its output is checked, not trusted.
Layer IV · Implementation
Hermes runtime on a dedicated isolated instance. The asset library and design history in Supabase, image generation through controlled prompt pipelines against current models, every design stored as structure plus render plus bill of materials.
runtime Hermes library structured asset catalog, fabrics to buttons composition rule-bound design engine, producibility enforced rendering image generation models, spec-controlled prompts state Supabase (assets, designs, renders, revision history) oversight designer review, human choice on every advance
The inventing illustrator
Image models add details nobody asked for, a phantom pocket, an extra seam, buttons that match nothing in stock. Every render is validated against its structured specification, and drift means regeneration, not a shrug.
Impossible combinations
Some pairings look fine as pixels and fail as garments, a trim that cannot attach to a fabric, a collar wrong for a cut. The composition contract blocks these before rendering, the agent cannot propose what the workshop rejected in writing.
Stock drift
A design composed from last month's inventory is a promise the client cannot keep. The library syncs with actual stock, and designs referencing depleted assets are flagged with suggested substitutions from what remains.
Sameness
A generator that converges on its favorite five combinations is a catalog, not a designer. Variety is enforced across each delivered set, concepts must differ in structure, not just in color.
Taste drift
The client's aesthetic is a constraint like any other. Approved and rejected designs feed back into composition preferences, so the agent drifts toward the house style instead of away from it.
300+
assets cataloged, fabrics to buttons
12
design variants per brief, standard delivery
minutes
from brief to first rendered concepts, previously days
100%
of delivered designs producible from current stock
1
designer, now reviewing instead of starting from blank
Figures from the first months in production, producibility verified by the client's own production team.
The library is where the intelligence lives. The impressive part looks like the image generation, but the value was created earlier, in the unglamorous weeks of cataloging fabrics and writing down combination rules. A generative system is only as trustworthy as the vocabulary it is confined to. Constrain the vocabulary and the creativity becomes usable.
Producibility is a feature you enforce, not hope for. The difference between AI art and a design tool is a validation step. Every mechanism in this system that rejects a render, blocks a combination, or flags depleted stock exists to protect one sentence the client says to customers: if we showed it to you, we can make it.
The designer got promoted, not replaced. Their day used to start at a blank page and end at variant three. It now starts at twelve rendered options and ends at judgment, which was always the part of the job that mattered. The agent made taste the bottleneck, and taste is a much better bottleneck than typing.
Your inventory is a design language. Nobody has ever spoken it fluently.
> ../book_a_call.sh