The result is two live consumer apps: Refrain, a folk song verse app available on the web, iOS, and Android (coming soon), and Tuning Note, a reference tone app for string players on iOS. Both were built entirely in collaboration with Claude.

I made all product, design, and creative decisions. I designed both app logos independently, with no AI assistance. Over the course of both projects, I took on more and more of the smaller code edits myself—adjusting copy, changing margins, modifying text styles—while continuing to direct Claude on anything architectural.

Visual design

The initial color palette came from my description of the app: a seasons-based experience rooted in the natural imagery in old folk songs. Claude proposed a starting palette; I refined it through iteration. Getting 16 distinct palettes—each readable, each evocative of its moment—required several rounds of back-and-forth, and Figma was useful for evaluating options between rounds.

The typeface, IM Fell English, was Claude's suggestion. It's a digital revival of a 17th-century face that fits the folk song material without feeling costumed. I kept it.

The hierarchy on the page and in the home screen widget came from my written descriptions of what I wanted, rendered by Claude, then refined through multiple iterations I managed closely.

The curation workflow

The content in Refrain is the whole point. I worked through more than 3,000 folk songs across 12 collections to find the ~223 verses now in the library. All 12 collections are fully triaged.

We started with the Child Ballads, the largest collection. I described to Claude what I needed: scan the collection, identify verses with explicit references to time of day or season, and assemble candidates with full metadata. Working through the initial findings together established the nuance quickly—what counts as a morning image, what doesn't, how self-contained a verse needs to be to work in context. Out of that first pass came a documented set of curation rules that applied consistently to every subsequent collection.

From that point, I could open a new conversation, name a collection, and say "let's curate this one." Claude would scan the text, present candidates in batches, and flag borderline cases. I reviewed and approved or rejected every verse. The curatorial judgment stayed with me; the work of scanning hundreds of pages was something Claude could do quickly and consistently.

Refining the animation

The animation required the most back-and-forth of anything in either project. The original implementation moved notes horizontally, in a flat line. That wasn't what I wanted. Getting the arc geometry right—the radius, the tangent rotation, the drag direction—took several rounds of precise description and correction. The drag direction was initially inverted. The arc curvature needed recalculation. Opacity behavior on adjacent notes was refined.

The experience made clear that, unsurprisingly, describing motion in words is hard. The quality of what Claude produces in these areas tracks closely with the precision of the input. A vague description gets something technically functional but wrong in feel. A precise one—including the sketch, and specific language about what the drag should feel like at the boundary—gets something much closer.

Post-launch: expanding to all string instruments

The first version covered violin range only. A post-launch update expanded to all four orchestral string instruments, which required rethinking the pitch system from a fixed list to a master list (C1–C6) with per-instrument slices. This also prompted a structural cleanup: what had been a collection of state variables scattered through the main view was extracted into a dedicated NoteNavigator class, making the instrument-switching logic—and the codebase generally—much easier to reason about.

That cleanup came directly out of one of the periodic simplicity reviews. The restructure was Claude's, but the prompt to do it was mine.