AI workflow · How I actually work

I don't just prompt AI.
I lead it like a team.

Every AI on my desk is good at something different. My job is the same one I do with designers: know who's good at what, hand them the right context, review everything, and make the call. Days become hours, hours become minutes, and the judgment stays mine.

The teamCopilot · in-house AI · Claude · Lovable · Figma Agents
My roleDirector, editor, final call
Proof on this pageReal quotes, real timestamps

Right question, right AI.

The skill isn't typing prompts. It's knowing what each tool can see and what it's good at, then asking the question only it can answer.

Copilot & in-house AI

The insiders

SeesCompany documents, data and history, plugged into the matrix.

I ask“Pull every result, date and quote about this project.”

Claude

The build partner

SeesEverything I load into it: files, memory, project docs, custom skills.

I ask“Screen this for anything that shouldn't go public, then build it.”

Lovable & Figma Agents

The prototypers

SeesReal design-system templates and components.

I ask“Three structurally different flows. Not three color swaps.”

Figma image generation

The illustrator

SeesOnly my prompt, tuned to how each image model behaves.

I ask“Exact on-shirt text in quotes, print-safe colors, no logos.”

How a project actually moves.

The real loop, the same one that built this portfolio. AI does the heavy lifting in between. The rust waypoints are mine, every time.

Hoursnot days, for research, flows and first builds
80+data points stitched into four case studies
1 dayto rebuild this portfolio, start to live
Me: framing, judgment, anything irreversibleAI: research, options, building, checkingwhoopstabletopdoublerollerskickerbraking bumps↺ then back to 01, a little smarter01I briefShort and direct,often by voice.02I load the contextFiles, past decisions,the right skill.03AI proposesOptions, a pick, andthe risks.04I decideAccept, reshape oroverrule.05AI builds & checksNumbers vs sources,layout on a phone.06I own the irreversiblePublishing, domains,sign-ins.07It remembersDocs, memory andskills for next time.01I briefShort and direct, often byvoice.02I load the contextFiles, past decisions, theright skill.03AI proposesOptions, a pick, and therisks.04I decideAccept, reshape oroverrule.05AI builds & checksNumbers vs sources, layouton a phone.06I own the irreversiblePublishing, domains, sign-ins.07It remembersDocs, memory and skillsfor next time.↺ then back to 01, a little smarterMeAI

Proof 1 · The case studies on this site

Connecting dots I'd need a week to find.

The four case studies on this site started in Copilot. Because it's plugged into company knowledge, I didn't ask it to invent. I asked it to find and connect. Then a second AI checked the work, and I made the calls.

A/B tests + confidenceLaunch phases & datesResearch & lab signalsBusiness case vs actualTraffic & adoptionAward write-upsDesign files & flowsPartner map4case studies
  • A/B tests + confidence
  • Launch phases & dates
  • Research & lab signals
  • Business case vs actual
  • Traffic & adoption
  • Award write-ups
  • Design files & flows
  • Partner map
  • ↓ 4case studies
  • Claude flaggeda “confidential” footer on every page, an internal ID in a PDF, and headline numbers that needed a human to explain them.
  • I decidedwhat could go public: percentages yes, counts yes, dollar amounts redacted.
  • I suppliedthe facts only I knew: how many designers I led, and what “+136%” really meant.

The insider AI is brilliant at gathering. It's not the one that should decide what's safe to publish.

Proof 2 · This website

From audit to live in one working day.

The actual log from the day this site was rebuilt with Claude, including the moments I said no.

My callAI did it
8am9am10am11am12pm1pm2pm12345678Audit to live: one working day
  1. 01 · 8:22Kickoff. Picked up the redesign plan and handed over four case studies plus a leadership summary.
  2. 02 · 8:34My call. Claude proposed three new projects plus one older one. I wanted all four new ones, and a plan before any building.Provide your thoughts before any actions are taken.
  3. 03 · 8:57My call. Claude had deleted the dollar figures to play it safe. I wanted readers to see the result exists, without the number.Can you redact the dollar amount, instead of removing them?
  4. 04 · 9:35Deployed. Five pages up on a new host. Claude found the site set to private and, with my OK, made it public.
  5. 05 · 10:09I pressed the button. Domain settings are mine to change. Claude gave me the exact records, then verified them.
  6. 06 · 13:03Iterated. A sticky project bar went live, plus a fix for a mobile chart bug nobody had noticed.
  7. 07 · 13:15Caught its own mistake. While rebuilding my résumé, Claude found it had mislabeled my job titles on the site, and fixed them.
  8. 08 · 13:54My call. The first draft of this page leaned on a friend's framework and filled gaps with guesses. I sent it back.Are you just templatizing Matt's and calling it mine?

AI produced a polished draft of this page in minutes. It took a human to notice it wasn't true.

Proof 3 · Teaching the team

I write skills for my AIs.

A skill is a playbook an AI follows every time, like onboarding a new designer. When a workflow matters, I turn it into one.

My apparel side project is the clearest example. After a product line got pulled over trademark issues, I stopped trusting ad-hoc AI output and wrote a skill that forces trend research, tool-tuned prompts, SEO copy grounded in the actual image, and a compliance pass on every design.

merch-design-batch / SKILL.md

# Replaces generic output with deliberate passes.
# Skipping straight to prompts is what this exists to prevent.

1. Trend & keyword research  do not skip
2. Visual direction          one named style per design
3. Tool-tuned prompts        exact text in "quotes"
4. SEO copy from the image   not a mad-lib
5. Compliance pass           no logos, no model names

Excerpt, lightly simplified. The same approach now runs this portfolio, including the design rules behind this page.

Pushback is a feature.

Good AI collaboration looks like a good design critique: both sides are allowed to say “wait.”

AI stopped me

  • Every case study carried a “confidential” footer, not publishable as-is.
  • A PDF I'd handed over included an internal employee ID.
  • My public résumé still listed a phone number.
  • Don't cancel the old site plan yet: the domain renewal may depend on it.

I stopped the AI

  • “Percentages are fine, the dollar amounts should be redacted.”
  • “Provide your thoughts before any actions are taken.”
  • “Fictional can be tried, we can assess once you create it.”
  • “Are you just templatizing Matt's and calling it mine?”

What never crosses the wall.

In a regulated industry, the first AI decision isn't a prompt. It's where the data is allowed to go.

Inside the company walls

The insiders

  • Copilot
  • In-house AI assistant
  • Lovable on company templates
  • Figma Agents & skills

Real data and internal context stay in the tools built for it.

Out in the open

My own projects

  • Claude
  • Figma image generation
  • This portfolio
  • Side projects

Everything here is public, fictional, or already screened for release.

When I roll up my sleeves.

AI is a fantastic engine and a mediocre driver. Some days the fastest path is closing the chat and doing it myself.

Bottom line

The skill isn't prompting.
It's knowing which AI to ask, what to hand it, and when to say no.

Thanks to my friend Matt Peacos, whose AI workflow write-up got me thinking, and whose Lovable design-system templates I prototype on.

Next projectSplitPayTeaching a bill to split itself.
← All projects