Essay 01 · AI workflow
Designing with agents:what changed in my process.
I build products with AI agents by writing briefs, not frames. The craft moved into the brief, the checks and the review.
Qualifi.ai, the AI career coach I have been building, runs to 13 surfaces across web and native. I didn’t build it alone, and I didn’t build it with a room of people. I wrote the strategy and the briefs, AI coding agents built in parallel lanes, separate agents audited and reviewed, and nothing merged without passing the build, the tests, the evals and my sign-off. This portfolio was built the same way.
It is the clearest example I have of how my job has changed. Not what I make, but how I spend the hours.
01
The brief became the design file
I used to put most of my care into the artefact: the frames, the components, the notes for engineering. Now a lot of that care goes into the brief. An agent will do exactly what the brief allows and fill every gap with something plausible.
Plausible is the problem. A plausible case study has a made-up metric in it. A plausible About page has a home address on it.
So a good brief now reads like a good spec used to: who it is for, what done looks like, what is out of scope, and the rules that don’t bend. Every lane gets the same hard rules and its own list of files it is allowed to touch.
brief.mdone lane · one brief
- Goal
- Build the /ai page from typed content.
- Read first
- The written spec (tokens, layouts and 50 pass/fail checks), the content model, the shared-parts API.
- Rules
- No personal contact details beyond email. No number without a source. Reduced motion. Nothing wider than a 390 px phone.
- Verify
- Capture desktop and phone, light and dark, and look at every capture. Three rounds, minimum.
- Owns
- The /ai route, its components and its content file. Nothing else.
- Report
- What was built, files, captures, open questions, and anything needed outside the lane.
02
Write the system down before you build
Before anything is built, the system gets written down: the tokens, the layouts, the motion timings, and a list of checks a reviewer can pass or fail. On Qualifi the same idea covered the AI itself. The voice rules were written for the model and for reviewers, and type sizes were locked by a test that fails if anyone adds an arbitrary size.
This is the step I would skip if I were working alone, and the one that makes everything after it work. Five agents building at once will drift five different ways unless they share a spec precise enough to be wrong against.
“Feels right” is not a check. “The headline is 58 pixels at 1440 and 37 at 390” is.
03
Parallel lanes, one owner per file
A build runs as several lanes, each in its own git worktree with its own branch and port. Shared parts (tokens, navigation, footer, the illustration kit) are built first, on their own, with a written API, so the lanes build on the same parts instead of inventing five kinds of button.
The rule that saves the most time is the most boring one. Each lane owns a list of files. If it needs a change somewhere else, it writes that down in its report instead of making it. Merges stay clean because nobody edits anybody else’s work.
It is the same rule a good design team runs on. I just had to write it down.
04
Review is where the judgement lives
Agents are good at saying they are done. So “done” has to mean something outside the agent: the build passes, the pages were captured at desktop and phone width in both themes and a person looked at them, and a reviewer that didn’t build the work went through it.
The most useful reviewers measured instead of looked. On Qualifi an audit found the long system prompt wasn’t actually being cached, because of how the SDK handles system prompts; fixing it was one of the most valuable changes in the build. Another swept every touch target on mobile up to 44 pixels. On this site, a read-only performance audit found the hero text was held invisible until JavaScript loaded. On a phone, that wait was 85 to 92 percent of the time to the largest paint. None of it shows up in a screenshot.
I ranked the findings and decided which mattered. That is the job now: not making every frame, but deciding what good is and checking we got there.
05
What I still do by hand
- Decide what is true. Every number on this site is listed with its source. Anything I haven’t confirmed is flagged for me and stays off the page until I do.
- Decide what is private. A case study for my current employer stays behind a preview flag until the company approves it.
- Decide how the AI behaves. What Compass remembers, when it says it isn’t sure, and what it refuses were my calls, written as rules before any agent touched them.
- Decide the story. Agents can draft; they can’t tell which project is the strongest proof for the team reading it.
- Decide when it is good enough to ship, and when it isn’t.
06
What actually changed
I spend less time pushing pixels and more on three things: writing the brief, building the checks, and reviewing. It is closer to leading a team than to working alone, except that this team will never push back on its own.
That is the risk. Most of my job now is making sure something in the loop still pushes back.
— ParthKitchener–Waterloo · 28 Oct 2026