BJERKE / DSGN
Senior Design Lead · IBM
Challenge
The Problem
What we were solving
Every portfolio is a static artifact. You build it, then it starts going stale — work gets added but context gets lost, voice gets diluted through editing, and the whole thing requires a manual rebuild every few years. The deeper problem: a portfolio page can't have a conversation. It can show work, but it can't explain the thinking behind decisions, answer follow-up questions, or surface the right project for the specific person reading it.
Design goal
Build a portfolio where you can talk to it. Not a chatbot bolted on as a gimmick — a system where the AI acts as Troy, reasons from a structured source of truth, selects and populates UI components to surface projects visually, and sounds like him because it was trained on his actual words through a structured interview.
Constraints
- ·
No backend database — all content lives in JSON and Markdown, deployable to Vercel as a static-first app
- ·
20-component catalog must be composable from a single JSON response (DynamicRenderer pattern)
- ·
The AI must sound like Troy — direct, plainspoken, no corporate hedging — calibrated through a structured interview, not prompt engineering alone
- ·
Projects must be surfaceable both through chat and through static project detail pages — two separate rendering paths from the same data
- ·
Component selection must feel intelligent — the system needs enough voice guidance to know when to show a MetricsStrip vs. a LeadershipStoryCard vs. a PeerQuoteCard
How we learned
Research
Findings
The hardest design problem wasn't the component system — it was voice fidelity. An AI that sounds like a portfolio assistant is useless. One that sounds like Troy is a tool. The calibration method was a structured 7-phase interview where the AI adopted a demanding hiring-manager persona and interrogated Troy's work, thinking, and leadership through a real conversation. The interview surfaced under-claimed work (a three-year research foundation, a design team that shipped production code, design-led revenue attribution), caught a leadership story being told as a tidy win when the real story was a months-long campaign to change minds, and produced a source of truth that now generates any portfolio artifact — case study copy, cover letters, talk abstracts — in the right voice on demand.
Methods
- 1.
Structured AI interview — 7 phases from seeding with real artifacts to distilling a reusable skill
- 2.
Adversarial persona framing — AI played a demanding design leader evaluating a senior hire, not a sympathetic collaborator
- 3.
Single-question-per-turn protocol — depth through follow-ups, not a questionnaire
- 4.
In-the-moment coaching — AI flagged gaps and pushed on weak answers before they became canonical
- 5.
Multi-artifact synthesis — interview progressively enriched with older case studies, Figma deck exports, and award nominations
- 6.
Source-of-truth distillation — conversation compiled into a structured document covering positioning, voice, projects, and leadership philosophy
Seeding the interview with a real artifact — the watsonx Agent Builder case study HTML — produced better questions than any self-description would have. The AI could read the actual work and interrogate the gaps, rather than reflecting back whatever Troy told it to say.
— Phase 1 — Seed with real material
The adversarial framing was what made it useful. A sympathetic interviewer confirms what you already believe about yourself. A demanding one finds the leadership story you're underselling — the one you've gotten used to telling as a win, when the real story is a months-long campaign to change minds.
— Phase 2 — Adopt a demanding persona
Component selection by the AI is not prompt engineering — it's intent design. The system prompt tells Claude not just which 20 components exist, but what kind of moment each one is built for. MetricsStrip for quantified outcomes. LeadershipStoryCard for influence-without-authority stories. SuggestedQuestions to keep the conversation moving. The AI selects correctly when it understands the intent, not just the API.
— Component Catalog Design
The output isn't a document, it's a tool — feed it back in and it produces aligned artifacts on demand.
Phase 6 principle
How we worked
Process
Phase 1 — Seed with real material · Day 1
Troy gave the AI the HTML of an existing portfolio project page — the watsonx Orchestrate Agent Builder case study. Not a prompt describing himself — the actual artifact. The AI read it and used it as the factual ground for everything that followed. Principle: start from real evidence, not a self-description.
Phase 2 — Adopt a demanding persona · Day 1
Troy asked the AI to interview him as a design leader at a mid-sized software company deciding whether to hire him for a design lead role. The adversarial-but-fair framing surfaced gaps a sympathetic interviewer would have skipped. Principle: the quality of the output is set by the quality of the pressure.
Phase 3 — One question at a time · Days 1–2
The interview ran as a real conversation — a single question per turn, each one building on the last answer. Topics moved deliberately: product-decision reasoning → handling tradeoffs → leadership and influence → a harder version of the leadership question → craft → fit and motivation. Principle: depth comes from follow-ups, not from a questionnaire.
Phase 4 — Get coached in the moment · Days 1–2
After each answer, the AI didn't just record it — it played back what was strong, flagged what was missing, and pushed on weak spots. It caught Troy telling his best leadership story as a tidy win when the more impressive truth was a months-long campaign to change minds across a team, and told him to lead with that. Principle: the interview doubled as coaching — the transcript records the gaps as much as the strengths.
Phase 5 — Feed in the rest of the evidence · Day 2
Once the spine of the story was clear, Troy added more source material: older project pages (Cloud Pak Experiences, the Kano Roadmap Study) and a working Figma portfolio deck export. The AI folded these in and caught things Troy had under-claimed — a three-year research foundation, a design team that shipped production code, current-year design awards. Principle: the system sharpens as you give it more real artifacts to reason over.
Phase 6 — Distill into a reusable skill · Day 3
The conversation was compiled into a single structured document — a source of truth covering positioning, projects, leadership philosophy, and awards, plus explicit instructions for how to generate future artifacts (decks, site copy, cover letters, talk abstracts) in Troy's voice. Principle: the output isn't a document, it's a tool — feed it back in and it produces aligned artifacts on demand.
Phase 7 — Use it (this page is the proof) · Ongoing
This case study was generated through that system. The page you're reading is itself an output of the method it describes — written in Troy's voice, structured by the same source of truth, surfaced by the same component system. Principle: a portfolio system you can run beats a portfolio you have to rewrite by hand.
Impact
Outcomes
Summary
The system generates any portfolio artifact — case study copy, cover letters, talk abstracts, interview prep — in the right voice, on demand. Adding a new project means updating a JSON file and a Markdown document; the AI handles presentation. The 20-component catalog gives Claude the vocabulary to respond differently to every conversation — surfacing work the way a person would, not the way a page template would.
Components in the A2UI catalog
Time to add a new project
Portfolio artifacts generatable from source of truth