Cases
Real problems, worked all the way through.
What the method looks like under load — including what didn't work. Less polished than a textbook, more honest than a sales pitch.
Validating 50 requirements with a prototype, not a wireframe
Fifty functional requirements with branching disposition logic needed sign-off before any engineering spend — and a static wireframe can't prove branching logic works. So I built the interactive flow in 1.3 hours and let product, QA and dev click through it instead of reading it.
Read the case →From a BRD to a shipped scoring platform in 13 hours
MagnaCarta went from a business requirements doc to a live Next.js platform — paste a Deliveroo URL, scrape the menu, score it on two rubrics — in 13.1 hours across 52 commits. The BRD → 369-line prototype spec → CLAUDE.md → code chain is exactly why I argue the spec matters more once AI writes the code, not less.
Read the case →Making 'good answer' legible across four releases
FAQ answer quality across World Mobile releases was a matter of opinion until it had to be compared release over release. I used Claude with Harvey-ball scoring and narrative reasoning so 'this got better' became a defensible claim with editor-specific guidance — not a vibe.
Read the case →