Inspiration
I used to run with YouTube playlists that layered music over a 175-cadence metronome. They helped my foot strikes stay on pace, but the songs rarely matched my taste and became repetitive.
I eventually built my own one-hour mixes in Cubase. I placed an hour of cadence clicks on one track, imported the songs I wanted, stretched and aligned them one by one, arranged the transitions, exported the mix, and transferred it to my Galaxy Watch. It worked well enough that I made three mixes. It was also tedious enough that I kept asking myself: could this be automated?
That question became Tadak. The name comes from the Korean onomatopoeia tadak. It evokes quick, light taps such as running footfalls, as well as an off-beat song and metronome clicking into alignment.
What it does
Tadak is a local Windows desktop app for preparing a running mix before the run. Add local music or load four included CC BY 4.0 sample tracks, choose a target cadence, and let Tadak analyze and transform each track. Arrange the Ready tracks, choose Even or Accented cadence cues, adjust Music and Click, use Quick check, and export one 192 kbps MP3.
Source music remains unchanged. Processing stays on the computer. Tadak is designed for music with a steady underlying beat rather than free-time performances or tracks with large tempo changes inside the song.
How I built it
Codex was more than a code generator on this project. One persistent GPT-5.6 Sol session carried most of the core work from product definition through audio research, implementation, debugging, packaging, and verification. It coordinated specialized agents, built the Electron Windows interface and local Python audio worker, connected Beat This, Signalsmith, FFmpeg, and LAME, and produced a portable Windows build with licensed samples and notices.
I supplied the running problem, product direction, and the listening judgment software could not provide. Codex turned that evidence into technical investigation and repairs.
Challenges and learning
The most important failures appeared through listening. One real track was initially compressed to about twenty percent of its original length. After that was repaired, the cadence clicks still drifted away from the music. I rejected both results. The same Codex session traced the first problem to a diagnostic overriding the natural musical choice, and the second to missing and extra beat events shifting phase across the track.
Five steady-tempo tracks later sounded natural and correctly aligned in full-length listening. The process reinforced a simple lesson: automated checks can prove mechanics, but they cannot replace listening when the product promise is musical.
Log in or sign up for Devpost to join the conversation.