Case study ·
CS-00
·
Independent
Evidence OS: this website, run like a product
I rebuilt this portfolio by directing a team of AI agents: a governing rubric, bounded briefs, review gates, and a QA bar. You’re looking at the deliverable. This is its case file.
01
A portfolio is a claim
Every design-lead posting in 2026 asks some version of the same question: how do you work with AI? Most portfolios answer it with an AI feature the designer once shipped. That felt thin to me. If I was going to claim I could direct AI-era design work, the portfolio itself should be the evidence, not just the container for it.
The practical situation was less philosophical. My site was three years old, a career transition was weeks away, and the honest math said a solo redesign at the quality bar I wanted was a couple of months of nights and weekends I didn’t have. So the question changed shape: not “can I redesign my portfolio,” but “can I run a redesign the way a lead runs a team, where the team happens to be machines.”
02
Stop using AI. Start directing it.
The reframe that made everything else fall into place: treat AI as a team to direct, not a tool to use. Once I made that shift, the job looked familiar. Teams need contracts, so I wrote a rubric before any pixels existed, built from two outside critiques of my old site, defining what 10/10 meant for every dimension and which metrics could be claimed in which words. Teams need specs, so the design prototype became the single source of truth, with a standing rule that when documents disagreed, the prototype won. Teams need briefs, so every agent got a bounded prompt: exactly what it could create, what it must not touch, and a required self-report at the end. And teams need review, so nothing shipped without a gate. The rubric was the contract, the prompts were the briefs, the self-reports were standups, and the review gates were the design crits.
Treat AI as a team to direct, not a tool to use.
03
Three calls: contract first, six bets, gates over autonomy
04
What directing actually looked like
The gates, not the prompts, did the design work.
The week ran as a pipeline with review gates between phases: rubric, then six parallel directions, then a full design system, then a phased build in Framer with a bounded brief per phase. Illustration ran as its own track with the same discipline. I locked a hand-drawn style with a reference image and a written spec, had ChatGPT generate an 86-image library against it, and reviewed the output like an art director, down to measuring stroke weight in pixels and sending back anything that drifted toward ornate. When a generator failed the same brief three times, I stopped re-prompting and drew the replacements myself. Knowing when to stop delegating is part of the job.
The review gates were where the design actually happened. My punch lists read like any design crit: the ledger header doesn’t match the decision block, the hover has a flash before it settles, the tablet layout collapses. The difference was who fixed them, and how fast the fixes came back.
05
The receipts
~1 working week — concept to launch-ready, alongside a career transition (measured, dated files)
6→1 — complete design directions prototyped, then chosen from evidence (measured)
28 + 9 + 26+ — color styles, text styles, and components built to spec in Framer by agents (measured)
86 / 73 — illustrations generated and reviewed; keepers after an art-direction pass (measured)
25 — defects caught at human review gates across two punch lists (measured)
06
What transfers to teams
This was a team of one, so the leadership story is really an operations story, and most of it transfers directly. Briefs beat prompts: the agents that produced the best work got the same thing good designers get, which is context, constraints, and a definition of done. Contracts beat taste-checking: writing the rubric once meant I didn’t have to re-litigate quality on every artifact. Review gates are where design happens: the machines produced volume, and the gates turned volume into a product. And evidence standards matter more with AI, not less: when production gets cheap, the scarce thing is knowing what’s true, which is why every metric on this site carries its scope and every claim its evidence type.
07
Reflection
The machines were fast. The standards were mine. That’s the whole case, honestly: nothing here required a bigger team, it required treating quality as something you write down and enforce. The part I’d do sooner next time is the specimen page, the review surface that made every later defect visible in minutes. And the part I keep thinking about is how familiar it all felt. Directing agents is mostly the job I already had: clear briefs, honest evidence, and taste applied at the gate, not sprinkled at the end.
Field note
Field assistant: Elwood. Contributed morale.
