Codex Builds a Browser-Based Sampler

Deep-Efficiency9146 · reddit · 2026-07-18

The author treated GPT-5.6/Codex as a somewhat chaotic but architecture-aware product team, iteratively building a 9-pad sampler inside a browser using randomly clipped YouTube segments.

During this experiment, it successfully generated complex browser audio logic:

9 independent YouTube audio tracks/Pads
Audio capture via browser tab sharing
Post-sampling features like trimming, reverse playback, and speed adjustment
16-step sequencer with Elektron-style conditional triggers
Per-step pitch, volume, micro-timing, and effect parameters
Probability and LFO modulation, transient slicing, and keyboard playing
Local project saving and mixdown export

The author noted that the hardest parts weren't the UI, but rather browser audio permissions, keeping the sequencer stable during BPM changes, avoiding pops between buffers, and preventing old audio from leaking into new Pad states.

An interesting takeaway: when tackling complex issues within large files, the most effective way to prompt Codex isn't to "guess the API error," but to describe symptoms like a musician, such as "the next loop sounds like the tail is cut off too early."

The author shared the finished product to see if others are using these models for long-term, iterative creative tool development, rather than just "one-shot app generation."

Original post →

Codex Builds a Browser-Based Sampler

More from coding & agent