AI-assisted delivery, and what it changed
The handoff is where design intent goes to die, so that's where the pipeline got rebuilt: two MCPs in one context window, and a spec generated from the design source instead of written alongside it. The first study is what changed. The second is what the pipeline revealed once it was running — a documented process becomes a training set, and the efficiency gain turns into a question about what the design function is actually for. The third is that model turned back on this site, including the page you're reading.
-
The delivery model in full: two MCPs in one context window, a spec that gets generated rather than written, and the stakeholder group that is still only partially convinced.
Arcos, IncHow restructuring the design-to-dev pipeline at Arcos cut kickoff time from two weeks to three days — one shared design language for UX, engineering, and AI.
At Arcos, design specs were thorough but not executable — developers made hundreds of small judgment calls per feature, and AI entering the pipeline automated the ambiguity instead of solving it. I built Harmony as a context-rich npm package paired with Storybook and Zeplin MCPs, giving UX, engineering, and AI a shared design language. Kickoff compressed from two weeks to three days; a testable coded concept now takes 12 hours instead of ~40.
-
What the pipeline revealed once it was running: a documented process becomes a training set, and the efficiency gain turns into a question about what the design function is for.
Arcos, Inc — UX Leadership & AI DeliveryWhat an AI-assisted delivery model revealed about where design creates its highest value — and why speed without clarity of purpose just accelerates the wrong thing.
When the delivery pipeline started working, the recovered capacity had to go somewhere. This is the story of what the efficiency revealed: that the upstream problem — purpose, data shape, user intent — is harder and more valuable than the downstream one. And what it means when your team’s documented ways of working become the training set for an AI agent someone else built.
-
The same model turned on this site: three systems that looked reusable from a distance, and the reasoning about which one actually was. You are currently reading its output.
jason.sonderman.infoThis redesign shipped three systems that looked like one reusable pattern. They weren’t. Here’s how I figured out which one actually was, and extracted only that.
This site’s own redesign produced three candidate patterns worth reusing elsewhere: a static, agent-facing content layer; an intent-based lens entry system; and an offline fit-brief generator. Investigating them side by side showed they were architecturally independent, not one coupled idea. I extracted only the agent layer — the part with the highest reuse value and the clearest thesis — into a standalone template, and left the other two out on purpose.
- Artifact
agent-layer-template
qaContext, llms.txt generation, and the eval harness this site's own agent layer runs — extracted and published as a GitHub template repo, so it's a thing you can fork rather than a paragraph you have to take on faith.
(opens in a new tab)