Reviving abandoned side projects with Claude Code: a YouTube Music shim case study

A developer revisits a long-shelved personal project — a shim exposing YouTube Music through the OpenSubsonic API — and uses Claude Code with Opus 4.6 to finish what he never would have on his own. The setup leans heavily on giving the agent strong scaffolding upfront: a uv project with FastAPI, Pydantic, ytmusicapi and yt-dlp pinned as dependencies, the OpenSubsonic OpenAPI spec dropped in the folder, a CLAUDE.md with explicit conventions (type annotations, Pydantic v2, Google-style docstrings, pytest fixtures), and a workflow built around plan mode, follow-up verification prompts, and clearing context between tasks.

The MVP — search a song in a Subsonic client and stream it — came together in a couple of iterations once stubbed endpoints were corrected to return empty-but-structurally-valid responses. The longer tail covered roughly 80 endpoints across 15 categories: in-memory caching for ytmusicapi to dodge rate limits, SQLite-backed metadata browsing, and disk-cached streaming with cleanup on client disconnect. Authentication was deliberately skipped since the project was never meant to ship.

The author frames this as wish-fulfillment coding rather than skill-building, distinguishing between projects done to learn (where he still grinds through Rust by hand) and projects he simply wants to exist. AI assistants are well-suited to the latter — well-specified, non-novel work with a clear contract — and clearing the backlog of unfinished ideas is a legitimate use, provided stretch projects don’t get displaced.