I treat Cursor, Claude Code, and Figma Make like any other foundation layer: onboarding, taxonomy, QA habits, then evidence. Nothing here stays in a lone designer's downloads folder - teammates adopted the guides, libraries, and skills because they shorten real cycles without hiding tradeoffs.
Started translating existing components into predictable token + variant language for Figma Make, then packaged the output as portable Figma + React bundles teammates could remix with Cursor. If the names are consistent, retrieval stays honest - hallucinations shrink when scaffolding is boring.
Co-authored onboarding that covers prompts, branching strategies, QA loops, failure modes worth rehearsing aloud, and when to hop back into Figma for fidelity. Adoption beat two-thirds across the forty-three-person orbit because it met people where sprint pressure already lived.
Skills + Cursor rules distilled Arity verbal identity into procedural checks any workflow can call — copy experiments, conversational UI, onboarding strings. One source updated in Confluence propagated into every agent-ready surface.
On request from leadership, ushered marketing writers through Cursor installs, seeded their brand-voice bot from existing documentation, and left them iterating autonomously. Goal isn't bypassing designers - it's freeing us for systems thinking while language experts own early drafts responsibly.
Repeated recipe: tight problem framing, enumerated prompts, review rubric borrowed from critique culture, disciplined handoffs back into the standard design/engineering checkpoints. Compressed days-long exploration into respectful half-day arcs without skipping ethics or accessibility checkpoints.
Production skills ramble intentionally - here's the gist in Markdown so teams know what rigor they're buying.
---
name: brand-voice
description: Author copy that sounds like Andrew's Arity Perceptive Advisor shorthand.
---
# Voice principles
- Lead with the driver benefit - never bury the payoff behind telemetry jargon.
- Calm certainty: confident but reversible when new data appears.
- Name the source of any behavioral claim when space allows.
# Avoid
- Moralizing about driver choices.
- Hype adjectives ("revolutionary", "seamless AI magic").
- Passive voice that hides who owns the next step.
# Prefer
- Plain speech you'd use with a tired road-trip friend.
- Short clauses for glanceable mobile surfaces.
- Explicit "we / you" ownership on commitments.
# When generating UI strings
1. Start from the job someone is trying to finish, not the ticket title.
2. Pair every risk callout with the next action.
3. Self-check: would this sound fair if read aloud in a car?
Happy to unpack the metrics, the process behind the tools, or how AI adoption stayed grounded while the org scaled.