Thoughts on using AI

Human in the loop

AI can be incredibly capable, but it requires supervision. My working mental model is that it’s like a super-genius toddler. It can help you solve problems, but you can’t let it wander off near the stairs.

Floors vs ceilings

It can raise the floor for someone working outside their skill set for certain. It allowed me to develop this website for instance. But what gets overlooked is that it also raises the ceiling higher for someone with real expertise. My devs would do far more with the same tools than I can. It lifts everyone, and the experts most of all.

I don't design in a chat box

Chat is great for thinking and terrible for designing. When the work is visual, I want a canvas and real controls, not a text field describing what I want in sentences. So I use AI around the design work—research, structure, code—not as the place the design actually happens.

The "AI-native" tell

This label gets the same eye-roll every big buzzword earns from me. If you're leading with it, you're probably closer to the floor of your craft than the ceiling.

Talk it through first

Before I ask it to make anything, I talk the problem through with it—the angle, the goal, what good looks like. Skipping that and jumping straight to "make me X" is where bad output comes from. Get the thinking right, then build, from the specs and boundaries I created.

Work in layers

I never ask for the finished thing. For instance, when crafting an experience map, I start with the stages, then work down into tasks and goals, then back across for opportunities. Each pass is a round trip—it drafts, I react, I tell it where it's off, it adjusts—and the next layer builds on the last. The output gets better because I'm shaping it the whole way down, not grading it at the end.

You get out what you put in

The output is only as good as what I feed it. I won't get real user analysis without giving it the actual interview notes and data—real research, done with real people. And with NDAs, be careful: a "Pro" plan is still a consumer account, and can train on what you share.

Helping my team

This isn't just how I work, it's what I coach. I push the same structured, dialogue-first approach with my team—talk it through, build in layers, stay in the driver's seat. And I have everyone exploring skills, so the good patterns we find get captured and reused instead of reinvented every time.

It will confidently lie

On a research doc it invented citation IDs and stated them like fact. We didn't catch it until a client asked about one. When I told it to recheck, it found fifteen more it had made up. It’s taught me to follow up with self-audits, and verify anything that matters now.

Verify the plausible

It’s very good at providing things that potentially sound good. But none of them are mine until I've gone through them one by one and decided. The moment I let it lead like that, I've stopped using my own expertise—I'm just editing its output. Steering it incrementally, staying in the driver's seat, is how the work stays mine.

Its logic isn't always logical

We split a platform's functionality into six categories to test whether an off-the-shelf system could replace a custom app. It picked a different solution for each category and concluded nothing could replace the whole thing—but it never thought to check every system across all six, or to mix them into one federated answer. It does what you ask, not what you meant. I had to spell that out.

Built for myself

One of the most interesting use cases using AI to help you build your own tools. This been very useful for things that are beyond just creating deliverables directly. Here’s a few examples I've created for myself.

• A set of Claude Skills for our research deliverables—interview scripts, reports, experience maps—each structured to get good input and hold to past standards
• A terminal-based site crawler that screenshots every page of a site for pattern audits
• A Figma plugin that builds word clouds from weighted terms
• An app that pulls comments out of a Figma file and sorts them into resolved, unresolved, and next steps
• This portfolio site—designed in Figma, given access via MCP, built and wired to Contentful Headless CMS, running in a GitHub repo, hosted on CloudFlare.

In client work

Our clients are bringing up AI, but most of it is still talk unfortunately. Here's a few instances where it earned a prototype.


• A planning assistant that returns bite-size, formatted answers—plus the subtler version, AI working behind the scenes in smart-planning and contextual suggestions rather than a bolted-on chatbot
• Damage assessment from photos—reads field images of an incident and pulls asset IDs and visible damage as a first draft for a report
• A zoo itinerary planner that builds routes from your constraints ("two young kids, one loves giraffes, four hours"). Surprisingly accurate. It even worked in a bathroom break

Using Agents

No coworker, no autonomous agents yet, need to find a good use case for my process. Claude Code can be controlled remotely from a phone, which makes the idea of building while taking a thoughtful stroll very appealing.

Generative design tools

Chat is not a good design interface, and you can see they are realizing that. Things like Figma Make, Google Stitch and Claude's design tools are a start, but aren't there for me yet. Fine for a quick throwaway idea, but not something I'd use to actually design projects.

AI-run research

There's some push toward AI moderating interviews and standing in for users with synthetic personas. I'm very skeptical it can replace talking to a real person, but would depend on what you trained it with. I haven't tested it enough to say for sure. Worth watching.