A design agent for company merchandise. It reads the company, writes the brief, generates four distinct designs on the actual garment, and hands print-ready files to the printer. Born out of a collaboration with a Danish t-shirt printer.
Most company t-shirts are boring. A small logo on the left chest, a colour picked by whoever ordered them, and a box of mediums nobody wears after the event. I worked with a Danish company that prints and ships exactly those shirts, and the interesting part was not the printing. It was the three weeks of email before a design was agreed: brief, draft, feedback, new draft, feedback, approval.
TexForm grew out of that. The idea is a shirt or hoodie the employee actually wants to wear on a Saturday morning walking the dog. The brand goes across the whole garment and tells a story, instead of hiding on the chest. And the three weeks of back and forth are gone, because an agent does the groundwork before any human designer touches it.
The customer fills in a five-step form: company, website, logo, the event (a run, a conference, a festival, an anniversary), and which products they want. That is all. The promise on the front page is four designs before you pay, delivered in ten days, and nobody calls you to ask clarifying questions in between.
What the agent does
- Reads the company. It fetches the website, pulls the palette, the typography character, the logo and a tone in three words, and writes down every assumption it made. Slogans and claims it is not allowed to reuse go on a blocked list.
- Picks four archetypes. Typographic, iconic, illustrative or pattern-based, chosen in code from the event type, so the four proposals are four different ideas, not four colour variants of one.
- Writes the brief. For each archetype the model fills in a concept: placements on the garment, exact wording for any text, ink colours, a one-paragraph rationale in Danish.
- Generates the artwork and the mockup. The prompt is built by code from the brief, never by the model, so the same brief gives the same kind of result every time. Eight images per project: four artworks, four garments.
- Waits for a person. An internal approval gate sits between brief and images, and between images and the customer. The customer sees four designs on their own products, picks one, and can ask for changes.
- Prepares the print files. On order approval the chosen artwork is upscaled four times, the background is knocked out, and a production sheet is written with garment, colour, item number and placement in centimetres for the printer.
Not fully AI-designed, on purpose
The generated designs are proposals, not the final print. What the agent replaces is the research and the first three rounds: who is this company, where are they from, what do their people look like, which colours and which logo, what is the event about. All of that used to be a designer reading a website and guessing. Now it is a report, and the designer starts from a brief that is already right.
The rule that made it work: the language model fills in JSON, the code builds the prompt. When a language model sat inside the prompt builder, results drifted from run to run and nothing could be tested. Moving it out made the pipeline boring and reliable, which is what a print shop needs.
Where it stands
Products today: t-shirt, hoodie, sweatshirt, polo, tank top, cap, tote bag and apron. Ordering, order data, upsell mockups of the customer's own design on a tote or cap, admin pricing and the production handoff are built. Screen-print vectorisation is deliberately left for later.
I have not seen anyone else do company merchandise this way. Plenty of shops print what you upload. None of them read your company first.
Want results
like these?
Tell me what you are building. I will tell you straight what is worth doing.
Start a conversation →