Latest edition: 21 August 2026London — published continuously since 2026Free forever
The Founder Gazette
Startup news, held to newspaper standards
Tools & tactics

Solo founder runs CAD software she has never learned, via Codex

Yana Welinder builds garments, vendor outreach and a Stripe-connected storefront on her own, using a written spec and agents that drive the software for her. Here is the workflow she demonstrated.

By The Gazette desk21 August 2026109

Yana Welinder is the sole founder of Yana Bana, a fashion label she describes as AI-native. She has no engineers.

She set out the workflow on the How I AI podcast, hosted by Claire Vo, in an episode published on 17 August 2026.

The first step is a written specification she calls a fashion prompt. It covers silhouette, volume, how the fabric behaves, how the garment moves, and the sound it makes.

The point is that the prompt is treated as a technical spec rather than a wish. Welinder is a former product leader, and the document reads like one.

She generates the designs in ChatGPT using Images 2.0, which she says outperforms other models for fashion work. That is her assessment, not a benchmark.

In one session she takes a hand-drawn sketch through to a product photo, a runway photo and an influencer shot.

The more unusual step is the next one. Welinder uses Codex with computer use to operate CLO 3D, the pattern-making software the industry runs on, which she has never personally learned.

The agent drives the interface. She specifies the outcome. The output is CAD files intended for 3D printing.

That is the part a founder in any hardware-adjacent business should note: the skill barrier being crossed is not coding, it is a specialist desktop tool with a steep learning curve and no API.

Vendor sourcing went the same way. She ran research and outreach emails end to end using deep research and browser use, working through Superhuman.

The storefront is her own build too: a full e-commerce site with garment voting, databases and a Stripe integration, deployed on Vercel with code in GitHub. Pre-orders run through it.

She is not betting everything on the agents. Welinder says she is testing human patternmakers and Codex side by side on the same garments, which is the cheap way to find out where the machine still falls short.

The honest reading for a founder copying this: the leverage comes from the spec, not the model. Welinder's prompt is detailed enough that the output is repeatable, and the same discipline is what lets an agent operate software on her behalf without her supervising every click.

The episode does not put a figure on cost, time saved or units sold. What it does show is a category of work — specialist software you would previously have hired for — moving inside the reach of one person.

More: Yana Bana · Welinder's account on X