Savings.Club · How I work with AI

A design team of AI agents, working under my review

At Savings.Club I lead design across the website, the sign up, the member app and the messages that follow up. To keep that pace without lowering the bar, I built a design team of AI agents. Each agent has one role, the rules live in files they all read, and nothing they make reaches users without passing through me.

Role
Lead Product Designer. I designed the team, its roles and its rules
When
2026
Output
Copy decks, UI specs, landing pages and messages
Built with
Claude Code, GitHub

How I led it

  • I designed the roles, wrote the rules every agent reads and defined the review gate.
  • I make every final call on structure and copy, and I open every pull request.

01 · Roles · The team

One role per agent, and a lead that routes the work

A request becomes one card for the team. The Design Lead agent triages it and either hands it to one role or splits it into stages: research, UX, UI, copy, messages and front end. Like on a real team, each role knows what it owns and what it does not.

Diagram: me at the top with the brief and the rules, a Design Lead agent below, six role agents (Researcher, UX Designer, UI Designer, UX Writer, Messaging Writer, Front-end Dev), then my review as the gate and a pull request I open at the bottom.
Every piece of work passes through my review before it becomes a pull request.

02 · Rules · Shared rules

Rules written once, read by every agent

What a new teammate learns in their first week lives in files every agent reads: the official terms, the tone of voice, no em dash, numbers only from the source the business owns, and legal consent text that never changes. I change a rule in one place and the whole team follows it.

03 · Review · The gate

Nothing goes live from an agent

The gate is my review. Copy comes to me as a document in page order, and I decide the structure and the words there, not in a chat. Code comes as a branch that I open as a pull request for engineering to review. Messages are scheduled by a person.

04 · First run · What broke

A landing page, from brief to pull request, and what broke on the way

The first real workflow was a landing page in eight stages, from the brief to the pull request. It also showed me where the setup breaks: long runs hit session limits, and the lead does not always wake up when a stage ends. Rerunning the whole job wasted the stages that had worked, so I learned to restart only the stuck stage.

Learnings

  • An agent is only as good as the rules it reads. Write them once, in files, not in every prompt.
  • Review is the senior part of the job. Agents draft; I decide.
  • Plan for the pipeline to break, and know how to restart one stage without redoing the rest.

Keep reading

Savings.Club · Fintech (US)The whole member journey, rebuilt to sell, sign and keep membersWebsite, three step sign up, member app and CRM.Read the case ↗