I built Flyback, a modular synth where one patch makes both the picture and the sound, almost entirely with Claude

https://markovcd.github.io/Flyback/ I built Flyback, a free, open-source (MIT) modular synthesizer for Windows, macOS and Linux. You wire modules into a patch, and that one graph generates both a moving picture and its sound. I've been a software developer for 7 years, and outside work I'm into music and synthesizers, so this project sits where the two meet. I try to follow the coding practices I've learned, and I wanted to see whether they hold when an AI writes most of the code. They do, and they matter more. What it does • A visual patch editor: oscillators, patterns, shapes, color, math, feedback and effects modules, wired socket to socket • Picture and sound compile from the same graph, so one LFO can bend a shape and a filter at once • Per-pixel synthesis driven by coordinates and time, with feedback from the previous frame • MIDI input, live recording, and export to MP4/WebM/MP3/FLAC/WAV via a CLI • A text language to write patches as source, a standalone viewer, and a WebAssembly build that runs in the browser How Claude Code was used Claude Code writes most of the code. I direct the work, decide what ships, and write some of it myself. That's about 1,150 commits in seven weeks. Continuous integration, in the original sense. CI is the practice, not the server: integrate into main constantly, in small steps, and never leave it broken. Every change goes straight to main, with no feature branches. A Docker gate runs the full test suite (about 4,700 tests) on every push, and a change that fails it doesn't stay. That's what lets me trust code I didn't type. The agentic workflow • Parallel sessions. I usually run several Claude Code sessions at once, each on its own task in its own git worktree. When a session finishes, it rebases onto main, fast-forwards main, squashes its own churn into one commit, and builds the release artifacts without being asked • Rules every session reads. http://AGENTS.md plus rule files: one type per file, a glossary so each thing has one name, and never committing another session's uncommitted work • Skills for recurring jobs. 17 project skills cover authoring presets, retaking site screenshots, rendering the site's audio, profiling, and signing releases, so a fresh session doesn't have to rediscover how • Slash commands for whole routines. /bughunt finds bugs, proves each one with a failing test, fixes it, and files the test where it belongs. /release checks the changelog and cuts a release. /review-plugin decompiles submitted plugins and writes an accept/reject report • Memory in the repo. 167 architecture decision records that Claude checks before proposing a refactor, so settled questions stay settled • Agent drivability. Claude has to be able to run and check every feature without me looking at the window. The CLI renders frames, prints patches as JSON, and screenshots the editor, and the viewer lets Claude check a patch's picture or sound on its own. When Claude has to hack around something, the fix becomes a feature on the TODO list Biggest lesson: good old practices are what make AI-written code trustworthy. The leverage is in the repo's rules and tooling, not the prompts. Contributions welcome, by hand or with AI: the same CI gate applies to both. See CONTRIBUTING.md. Free to try Everything is free, with nothing paid or locked: • Play presets in your browser, nothing to install: https://markovcd.github.io/Flyback/presets.html • Download a build: https://github.com/markovcd/Flyback/releases/latest • Or install (Linux, macOS, Git Bash on Windows): curl -fsSL https://raw.githubusercontent.com/markovcd/Flyback/main/install.sh | bash