How I work with AI

Using agents to explore ideas, develop designs, and build on what I’ve learned.

Practice
2025–present

Good work builds on what came before. I wanted my work with AI to do the same.

When I returned to a project in ChatGPT or Claude, I’d dig through old chats and explain decisions we’d already made. Some of the documents I’d uploaded were out of date, so I had to upload newer versions. Switching tools meant rebuilding that context.

Learning to use Claude Code meant an agent and I could edit the same project files and keep them up to date. I now use those files with Codex and Claude Code, choosing the context each task needs.

I build workflows with AI agents to compare new feedback with earlier research and decisions, explore product directions, and make working prototypes.

Shaping the design

I use AI as a thinking partner to explore how a product could work and feel. With sketches, diagrams, and working prototypes, I map the steps someone takes and how the screens connect. I check the prototype against the person’s task: does the next step make sense, and is the information they need there when they need it?

I develop the visual design in Figma and give agents clear direction to explore ideas alongside it. They work in a separate canvas or copy of the prototype. I might take just one interaction or detail from an agent’s prototype and work it into my design. As we build, I refine the spacing, how it scrolls, and how things load. Once the layout, typography, and colors are in place, the agent can handle small changes within that design. For larger changes, we go through the same process again.

From one ad to the whole campaign

An AI prototype we built could explain a single ad’s performance. A marketer reviewing it described pulling reports from several platforms, checking creative, and explaining results to other teams. She asked for maps, audience breakdowns, and reports she could share.

I asked the agent to compare her feedback with our prototype, keeping her words separate from our interpretation. It suggested that explaining one ad might not help her report on a whole campaign.

To test that interpretation, I’d sketch a few ways to review campaign results and ask her which would help her explain what happened. The adapted files below show how I got to that question.

campaign-review feedback.md
# Feedback from the prototype review
 
## Her current work
She described pulling separate platform reports, matching creative to results, checking how things were named across sources, and preparing a report for other teams.
 
## What she asked for
Maps to show where results differed, audience breakdowns to show who responded, and reports she could share.
 
I worked through these requests in the working brief.
 
## Questions to take back
What does she need to explain to other teams? What would make that easier? Which information would she need to trust the explanation?
Adapted working records

Financial preparation

While preparing my 2025 taxes, I started building workflows with agents to review expenses, check invoices, and project income. I add context the records can’t provide, question the agent’s assumptions, and keep confirmed information separate from estimates and open questions. That helps me prepare for conversations with my accountant and financial advisor. Afterward, I use agents to understand unfamiliar advice and prepare follow-up questions.

Keep what we learn

When I choose a direction or correct the agent, I save why. For tasks I repeat, I keep instructions with examples and corrections. I check them as I work and update what doesn’t fit.