AI design system

It holds what a designer would need to start a job, and what a designer would otherwise carry in their head. Each part depends on the ones before it:

• Source assets: logo files, fonts, colours, imagery and product screenshots.
• Work the team has already accepted, with the brief for each where it survives.
• The decisions behind that work: what leads, what reads second, where the logo sits.
• A working method for each format, such as a launch banner or a deck section, often packaged as an agent skill.
• A connection that lets the agent reach all of it when a brief arrives.
• Lessons from review, each checked by a person before it joins the guidance.

The name is also used for a product’s design system made readable to an AI coding tool, so the code it writes uses the team’s tokens and components. The idea is the same: the agent gets the decisions along with the files.

A brand kept as a logo folder and a PDF works for a designer who already knows it. An agent given the same folder can find the logo and has no basis for deciding where it goes, so it makes work that is the right size and wrong for the brand. Written-down decisions give the agent something to carry into a new image or a longer headline, and give the reviewer words for why a result is wrong. Whether the whole thing saves time has to be measured against the team’s current process.

We start a trial with one task your team already repeats, such as a launch banner, and write down how it gets done now: who makes it, how long it takes from brief to accepted work and how many rounds of changes it needs. We name the reviewer and agree what accepted means before any candidates exist.

Then we build the system for that one format, in the order above, run a new brief through it with the revisions a normal job gets, and compare the time to accepted work, the corrections and the setup effort against that baseline. Today we have an initial skills library, our first brand pack and a local, read-only connection. The integrated review loop, remote access and native delivery are still to be built.

Say a team needs a LinkedIn banner for a product launch, with a long headline and a screenshot full of small controls. The agent pulls the logo, the fonts and three accepted banners, then reads the banner method. Past a set length, the headline changes the composition, and a screenshot too detailed for feed size is cropped to the one control the launch is about. The reviewer says the crop hides the new button. The agent revises that exact version, and the accepted file is exported for the team to edit.