Projects

Open source under MIT and part of my personal and client workflow. It runs in Claude Code, Cursor, and Codex. I regularly update it after analyzing its work in real projects.

I developed this skill based on my experience and analysis of best practices for working with design systems in Figma. This skill audits a file the way a reviewer would (token bindings, variant coverage, WCAG contrast, detached instances), builds components with variable bindings and structured descriptions, and exports the whole system to code as tokens.css plus rules for the agent in the repo. It will not implement designs as production code, and it is overkill for a one-off mockup with no token layer.

The grading scale, the contradiction pass and the quote rules come from my own research-heavy client work; the skill wrote down a practice that already existed. Open source under MIT, it runs in Claude Code.

I built this skill out of my own research practice: sources pile up faster than anyone reads them. It takes a stack of documents and returns structure: claims extracted one by one, graded by evidence quality, cross-referenced, contradictions surfaced instead of averaged away. Exact quotes are preserved where the wording itself is the data, and a claim without a source is treated as a rumor. It does not gather sources for you, and a single thin document gives it nothing to cross-check.

Vanilla HTML, CSS and JavaScript: no framework and no build step, because nine screens do not need one, and zero dependencies means nothing to rot. State lives in schema-versioned localStorage, so family numbers never leave the browser. Netlify, no backend.

Six years ago I planned my mother's care on paper, because paper was the best tool available. Families still have no structured way to see who carries the care of an aging parent, or what that work is worth. The app builds a family care plan in nine short screens: who does what, hours per week, and the market value of that labor, anchored to CareScout 2025 rates. It ends with a printable plan and a private share link. A family communication tool, not a legal document; the default rates are US-based, and every one is adjustable.

React 18 with TypeScript, Vite and Tailwind CSS 4, Motion for the step transitions. A budget planner is a chain of small stateful screens, which is the shape React handles best, and typed math is math that cannot fail silently. QA with Playwright and axe-core before launch.

I built this planner around the number every kitchen renovation starts with: one confident total that survives until the first invoice. It starts from your budget instead: it splits the sum into shares (cabinets, labor, appliances, everything else) against 2025-2026 US cost ranges and flags what won't fit before the work starts. Proportions and ranges, never a single made-up total; nothing you enter leaves the page, and there is no sign-up. A planning aid, not an estimate: a contractor's quote for your actual kitchen still beats it, and it should.

React 19 with TypeScript, Vite and Tailwind CSS 4. The rules engine is plain data with cited clauses, tested with Vitest, so no screen hardcodes a standard: when a regulation changes, the data changes, not the interface.

I built this tool around a question people rarely ask of a home: will it still fit the person its resident will become? Pick a layout, correct four measurements, set a mobility trajectory (cane, then walker, then wheelchair), and the required clearances are drawn over the floor plan until they collide with the walls. A room passes or fails, and every deficit is a measurement with a cited clause next to it. UK standards drive the verdicts; US figures sit alongside, never blended. What the tool cannot cite, it does not judge. Desktop-only for now, and a planning lens, not a survey of your actual home.