All work

04 · Systems

The AI design-to-production pipeline

Design and engineering as one pipeline. AI agents do the typing; I own the decisions.

Role
Product and design lead
Context
A creative studio's website
When
2026
Focus
Design ops, AI agents, Next.js, CI
59 → 7 MBhero animation, fixed by measuring first
24 → 57automated tests on the money paths
6bugs caught by agent review
26logged build sessions

The problem

A creative studio needed its website to work as a sales tool, not a brochure: lead capture, proposals, SEO pages, a CMS editors could use, and a signature scroll animation. There was no dedicated engineering team, and the work had to stay consistent across months of stop-start sessions.

My role

I set the brand rules, the structure and the priorities, and directed two AI tools. One designs and plans in a desktop workspace. The other builds and tests in the code repository. I review, accept or reverse every change.

Constraints

  • Two AI sessions that cannot talk to each other live.
  • Brand voice rules the copy must obey everywhere.
  • Real money paths: contact, proposals and downloads must never break.
  • The production host changed mid-project, so the release path had to move with it.

Decisions

Files as shared memory

The two agents sync through two files in the repository. One holds the rules, brand facts and current state. The other is a dated log entry after every meaningful change. Git is the source of truth for code; the log is the source of truth for intent. Any session, human or AI, picks up where the last one stopped.

Brand rules as constraints, not reminders

Brand facts live in the file every agent reads first. When the pillar framework changed from four to three, the rule changed once and every later session followed it.

Tests on the money paths before polish

Automated tests guard the lead, proposal and email routes. They grew from 24 to 57 as features shipped, and every session ends with type-check, lint, tests and build all green.

Measure, then fix

When the homepage felt slow, the fix came from measuring, not guessing. The server was fine. The real cost was a decode path sending most devices the heavy asset. A 59 MB frame sequence became a scrubbed 2 to 7 MB video, with frames kept only as a fallback.

Adversarial review by agents

A read-only, multi-agent review across six dimensions found defects normal testing missed, including a navigation dead zone that left tablets with no menu. Six confirmed bugs were fixed and verified.

A release you can trust

Every push to the release branch deploys to staging on its own: tests, a versioned build, a health check, and automatic rollback on failure.

Outcome

  • A production-grade site: an editor-friendly CMS, 15 industry and service SEO pages, structured data, a four-step project brief that emails the team a proposal draft, and lead capture with follow-up emails.
  • 57 automated tests on the money paths, and staging that deploys and rolls back on its own.
  • 34 portfolio projects and 178 media files published through one repeatable script.
  • The same pipeline built the site you are reading.

What I’d do next

  • Measure it properly: hours per shipped feature against a conventional designer-to-developer handoff.
  • Package the pattern as a template for product teams: a rules file, a decision log, a test gate.
  • Add visual regression tests, so design changes get checked as strictly as code.