PROJECT · 2025 — Present
Designing the AI layer of a creative team
Rebuilding creative production around generative tools with review gates, prompt libraries, and QA rituals so speed does not erode brand quality.
Role
Independent / Lead practitioner
Timeline
Ongoing
Team
Cross-functional with client creatives
Stack
Make, GPT, Firefly, Midjourney, Figma
Teams often paste AI into the same approval chain and expect miracles. I approached AI as an operating layer: prompts as reusable assets, review steps that catch hallucinations and off-brand outputs, and automation that removes busywork instead of adding hidden debt.
01 — THE PROBLEM
Throughput without guardrails
Creative throughput was capped by manual asset prep and repetitive variations. Early experiments with generative tools produced fast drafts but inconsistent brand voice, which made stakeholders pull the emergency brake.
The constraint was not tooling — it was trust. Without a transparent pipeline, AI became a side channel instead of a production accelerator.
02 — APPROACH
Treat prompts and QA as part of the design system
I catalogued recurring tasks (social crops, localization variants, icon explorations) and built prompt templates with explicit constraints: palette, typography, composition, and negative space rules pulled from the brand system.
I paired automation in Make/Zapier with human review gates for anything customer-facing. The principle: AI proposes, humans approve — until metrics proved where automation could safely expand.
03 — DESIGN DECISIONS
Three decisions that kept brand intact
Decision 1: Prompt library versioned like code
I rejected one-off prompts living in Slack. Versioned libraries let us roll back when a model update changed tone, and let new designers onboard without relearning secret incantations.
Decision 2: Visual QA checklist before ship
The alternative was eyeballing thumbnails. We added a lightweight checklist (logo lockups, text legibility, artifact scan) tied to ticket templates so QA lived in workflow, not heroics.
Decision 3: Measure creative cycle time honestly
We tracked hours per asset class before and after automation. The goal was not “more AI” but fewer repeated hours on commodity work — freeing time for concepting and UX depth.
04 — OUTCOME
Signals so far
+60%
Design throughput
M.Sc.
AI in progress
4+
Tools orchestrated
Gated
QA on every ship
05 — REFLECTION
What I learned
AI rewards teams who design the handoffs between machine output and human judgment. The craft moved upstream: clearer briefs, tighter rubrics, better critique — the same skills product design already values, applied to a new medium.
