Gatefold: an AI design process where a person makes every decision

In one sentence

Gatefold runs the design of a screen with AI agents, from moodboard to native code, in a fixed order in which the model proposes and a person makes every decision.

The problem

When a design team starts using AI, the work usually turns into prompting until something looks good. Nobody can say why a variant came out the way it did, what the model was given, or whether it copied a reference. The next round starts from zero, because nothing the team learned is written down.

The idea

I split the work into stages with a gate after each one, the way a design team already works: brief, moodboard, two or three directions described in words, a handoff, variants, a verdict, a diagnosis. The model does what it is good at: searching, proposing, generating variants. A predictable program keeps the records, checks the rules and stores the results. A person gives the only signals that count: "I'll take this one" and "this is not it".

What the team learns becomes a lesson that the next handoff picks up automatically. A lesson only counts after a person accepts it, and switching one off for a project needs a name, a date and a reason.

What it has produced

• This site's course pages: 12 patterns accepted from 6 handoffs of 6 variants each became the course on the job-posting subdomains, in production since 10 October 2026.
• A native screen: a payment confirmation screen in SwiftUI, accepted as a code package, with all 30 VoiceOver checks passing.
• The next home page: a star-atlas concept with a HUD mode, 15 variants and 6 accepted patterns, now in review on the preview subdomain.

So far I am its only user.

What I learned along the way

• Passing finished work through a model is a waste. Designing 3 motion sheets with an agent cost 0.80 USD, but having the agent pass 8 ready scripts on to the animation tool cost 3.34 USD, because it rewrote each script in full. A plain program now does that step for 0 USD.
• A model should not carry data it cannot verify. When long encoded images were passed through the model, content appeared that was not in the file. That step went back to a person, with a size and checksum check.
• A rule from experience can overreach. One accepted lesson rejected 25 of 30 references in a new project. Instead of quietly bending it, a project can now switch a lesson off, with a name, a date and a reason.
• Not everything works. An animated poster with liquid drawn in vectors was closed without a result; the motion stage was narrowed to interface motion.

Why it belongs in the portfolio

Gatefold is my answer to the question of what a design team's AI workflow should look like: the model speeds up searching and producing variants, while judgement, taste and responsibility stay with a person. It is also the tool I used to design the pages you are reading.

My role

I came up with, designed and built Gatefold myself, working with AI agents.