How I Used ChatGPT Dots to Build a Music-Pattern Catalogue

I wanted to build a site where someone could look up a musical pattern, hear a short example, and check what the source supports. I used ChatGPT Dots to research sources, prepare changes, run checks, and verify what reached the public site. I set the selection criteria, tested playback, and decided what to publish.
The original plan was smaller and slower
On November 14, 2026, Dots proposed a small site with eight searchable examples and an initial target of three playable clips. The plan called for deciding whether to publish on November 27. I asked Dots to start the later steps immediately, then expanded the scope.
The first Japanese listening site went live on November 14. We published v1.0.0 on November 15, and I considered the project complete on November 16 after further refinements. By then, Notable Musical Patterns was a Japanese and English catalogue with 16 source collections, 6,729 searchable records, and 2,983 catalogue entries with on-site audio. The record count is not a count of distinct compositions. All dates are in Japan time.
The English homepage during development, showing a listening example and the Alberti bass pattern.
What changed was both the pace and the deliverable. November 27 was the proposed publication decision date for the small site, not a deadline for the larger catalogue or the result of a separate run without my involvement.
What I could delegate
Early on, Dots spent too much time comparing individual scores and MIDI files while I was still looking for a broad range of candidates. I changed the order: define the musical feature, find public datasets, check the terms for commercial use, modification, and redistribution, then link likely examples to the feature. I reserved detailed checks of specific editions, recordings, and passages for decisions about what to present as a supported example or host as audio.
That rule produced different outcomes. For Mutopia, the site hosts synthesized excerpts from passages checked against the relevant scores and MIDI files. These are not recordings of performances. For O’Neill’s 1903 collection, the site provides bibliographic and original-source references rather than hosting synthesized audio. The collection and selection methodology describes those decisions.
The site uses static HTML, CSS, and JavaScript on GitHub Pages. The browser-side search matches catalogue entries to clip metadata and feature records by stable ID. The Python release checker checks record counts, local links, and whether audio files and credits match the catalogue. It does not test playback or establish that the music supports a label or that the audio is cleared for reuse. Checking the site in a browser was a separate step. It was how I noticed that the playback controls behaved differently across pages.
Where I still had to steer
Dots could carry out individual rounds of source research, but I repeatedly had to tell it to move on to the next round. I had already asked it to continue routine work without waiting for me. Early updates about brief pauses and restarts also made it hard to tell whether work was running or merely scheduled. I asked for updates only when there was a meaningful result, a blocker, or a decision I needed to make.
From the point I handed this work to Dots on November 13 through the completion decision on November 16, the project chat contained 898 messages: 325 from me and 573 from the assistant. That includes progress updates but excludes automated runs, tool activity, and other chats. The count documents the conversation; it does not measure time saved.
Try the catalogue · Browse the source code
AI was used to draft and revise this article.