Agentic email migration pipeline
A major rideshare brand was moving its whole marketing program onto a new brand system. That meant rebuilding a deep library of email templates, and nobody wanted to put a team of designers on find-and-replace for a month.
The problem
The new brand system touched everything: colors, type, icons, logos, dark-mode rules. The email templates were hand-built HTML, some of them years old, with patterns layered on top of patterns. Every one of them had to move to the new standard.
By hand, one template took anywhere from 20 minutes to two hours when things went smoothly. In practice it ran four to six hours, once you count the trial and error, the design-system lookups, and the render tests. The time wasn’t even the worst part. The work was spread across four or five tools. There was no real QA step. The brand guidelines changed halfway through a batch. And the know-how for doing it right sat in a few people’s heads, not in anything you could hand to someone else.
The approach
We built the pipeline around one rule:
Humans inspect. Humans accept.
You hand it a batch of old templates and a task sheet. It hands back updated templates that have already been through an independent QA pass, ready for a person to look at. The reviewer’s notes go back to the agent, and approved work ships. Four layers make that work.
What’s automated, and what isn’t
The whole design rests on one split. Most of the work is mechanical: swapping fonts, recoloring, rewriting CSS, mapping logo URLs, applying dark-mode classes. That runs through plain deterministic tools, which are fast and never improvise. What’s left needs judgment, like a hero image that has to be rethought or an icon with no clean swap. That part goes to a reasoning skill, and the skill is built to stop and ask when it’s unsure rather than guess.
Then a second reasoning pass reviews the whole template from scratch. You’d never let the person who wrote the code be its only reviewer, and the same logic applies here. Every correction it makes gets written back into the guides, so the know-how stops sitting in a few people’s heads and starts building up where anyone can reach it.
The outcome
The team stopped hand-editing HTML. Now they review finished, QA’d work instead. The know-how that used to bottleneck on a few experienced people sits in a system anyone can run, and what comes out is ready to ship, not a rough draft to clean up.
What it demonstrates
- Deterministic tools and reasoning skills, kept separate. Tools do the mechanical work. Skills handle the judgment. A second skill QAs the result.
- A person stays in the loop. The agent produces. A person inspects and accepts. Nothing ships without that.
- The know-how lives in the system. Brand rules and patterns sit in versioned, self-updating guides instead of in someone’s memory.
- It compounds. Because the memory improves itself, the pipeline is worth more after every batch than it was before.
- It runs on your stack. Built around the client’s own design system, assets, and tools, not a product you have to move into.