The journey
Growing as an agentic designer.
Each entry is a real shift — a new skill, a process tweak, an insight, or a project that forced me to upgrade something. I keep this honest: dates from commit history, scores from actual /ux-review runs, links to the artifacts they produced.
21 milestones and counting — filter by type to trace a thread.
Showing 21 of 21 entries
- Project
Started Widget Component Library
Angular 21 standalone library, 26 dashboard primitives — stat cards, metric cards, chart widgets, feed items. Demo app beside it. Not yet consumed by anything else; that's the case study.
Read the case study → - Insight
Tokens-as-code, not tokens-as-Figma
Decided the source of truth for color/space/typography lives in src/app/tokens.ts, not Figma. Code-first means every prototype gets identical values without manual sync.
- Insight
Signals-only UI state, no RxJS in views
Committed the launchpad store fully to signals + computed. Templates never see a Subject or async pipe. Change detection is OnPush everywhere and the framework never re-renders anything we didn't ask for.
- Project
Started cl-launchpad-protoapp-new
Angular 21 SSO launchpad with three layout themes, signal-based store, drag-and-drop app management. First real test of /lp-portals-design from blank repo to running prototype.
Read the case study → - Project
Started comp-design-library
Standalone Angular site to document tokens, palette, typography, and patterns for ClassLink. The first artifact that treated the design system as a living showcase rather than a Figma file.
See the case study → - Project
Built my-files-usability research harness
React + Vite harness that wraps my-files-react in an iframe and instruments it with task flows, heatmaps, click telemetry, and post-task feedback. Same codebase ships as runs in research.
Read the case study → - Insight
The data tells the picker what it can do
Hiding incompatible options isn't a constraint — it's the system being honest. A kanban toggle on data without a status column isn't a choice, it's a trap. Stopped offering choices users can't act on.
- Insight
Research instrumentation is the strongest design critic
Building the usability harness changed how I evaluate every other prototype. 'Would I be embarrassed to wire this up to the harness?' is now the pre-ship question.
- Project
DataView picker — five views, one source
Component system in cmc-module-reactsh: same data, five view modes (table, cards, compact, kanban, detail). The picker auto-hides views the data can't support — kanban only appears if a status column exists.
Read the case study → - Skill
First end-to-end /design-flow run
Parent-portal prototype became the proving ground for /design-flow — the orchestrator that walks a project through discovery, taste, shape, build, enhancements, review, and retrospective. First time the whole pipeline shipped a working app.
Read the meta case study → - Skill
Authored /lp-portals-design
Skill that codifies the Launchpad Portals aesthetic — Apple-inspired polish, three layout themes (Dynamic / Classic / Primary), token catalog, K-12 warmth without slipping into beige.
- Skill
Built /ux-review-loop
Orchestrator that runs /ux-review, applies P0/P1 fixes, builds, commits, and repeats until composite score ≥ 95% or 5 passes. Hands-off multi-pass quality improvement.
- Process
Cut /design-flow token budget 30–35%
Selective artifact loading on resume, Phase 4B receives only brief + visual-direction (not all artifacts), enhancement passes capped at 3, Sonnet routing on no-codegen phases. ~144–216k → ~98–162k.
- Process
Started experimenting with @google/design.md
Wrote a 264-line DESIGN.md for cl-launchpad — YAML front matter for tokens, prose for rationale and contrast notes. Lint and diff are wired; the export-to-SCSS loop isn't closed yet. Honest about it.
Read the case study → - Skill
/design-flow accepts PRDs as input
Added a fast path: if a PRD already exists, import it as brief.md and run a 2–4k verification pass instead of the full /design-discovery interview. ~6–8k savings per run.
- Process
Visual Compare Tool — discovery
Sketched a Playwright-based tool to pair prototype routes with production routes and surface visual diffs. Decided against Chrome DevTools MCP — not the right substrate for a persistent CI artifact.
- Process
Designed the executed-usability-testing skill
Sketched a skill that drives the actual app via Playwright as 9 persona archetypes (screen-reader, keyboard-only, low-vision/zoom, ELL, slow connection) — an observed-evidence layer on top of synthetic-user-testing's mental simulation. Design phase; shipped weeks later as /usability-lab.
- Insight
Two agentic substrates beats one
Lovable for divergence (fast spin-up, opinionated output that's not mine), Claude Code + /design-flow for refinement (scriptable, taste-aware, surgical). The play is using both, not picking one.
- Skill
Persona pipeline as a coherent workflow
/inclusive-personas builds the cast, /portal-personas formats it for a given product surface, /synthetic-user-testing predicts where each persona hits friction. Research now has the same orchestration spine as design.
Read the case study → - Skill
Wrote up /ux-review
Five reviewer agents fanned out in parallel — WCAG, design system, usability, mobile, behavioral — merged into one composite with a Critical floor that caps WCAG failures at C+ and dark patterns at C-. Auto-fix step bisects on build failure so a single bad swap doesn't throw out the whole batch.
Read the walkthrough → - Skill
Built /usability-lab
Shipped the executed counterpart to synthetic-user-testing: drives the running prototype with Playwright as each persona, attempts real tasks, and reports observed friction — steps-vs-optimal, dead-ends, unreachable targets, focus order, what a screen reader announces — every finding traced to an a11y snapshot, not inference. With --validate it cross-checks predicted friction against real PostHog drop-off and rage-click data.
See it in the persona pipeline →
21 milestones so far — and the loop keeps running.