It digs, I sequence
25 October 2026
New music finds me while I scroll Instagram. For a while those reels went into Todoist as tasks. By late August there were 37 of them.
Now every reel I like goes to WhatsApp, to the chat with myself. Share, WhatsApp, message yourself. That chat is the inbox of a research setup that finds records, checks them against what I own and argues with me about my sets. It does almost everything except the part I care about most.
From reel to track ID
A skill called /reels drains the self-chat. For each link it pulls the caption with yt-dlp, because most music reels name the track in the caption or the audio credit. If nothing is named, it downloads the audio and runs Shazam on it. If Shazam fails too, the line says “listen manually” with the link. It is not allowed to guess an ID.
The sweep on 20 September: 30 links, 28 fetched, 22 with captions, 9 with full tracklists. Ten reels showed the track names only as cards inside the video, so it pulled 393 frames at one per second and read the cards visually, since OCR was unreliable on footage filmed off a screen. That one sweep produced 17 records I didn’t own and 6 where I owned the original and the reel named a better version.
The library decides
The ownership check is what makes this useful. The setup reads my Rekordbox database directly through an MCP server called reklawdbox: the collection, play counts, BPM and key, and a score for how two tracks would blend. The access is read-only. Changes go into a staged XML file that I import myself. On 20 September the library had 5,323 tracks, and every candidate was checked against it.
It also found problems I hadn’t asked about. 497 tracks in my library are below 320 kbps. The upgrade list sorts them by play count, because a 128k file I have played 20 times does real damage on a club system. At the top: Kings of Leon, Sex on Fire, 128k, 35 plays.
Crates and one shopping list
What survives goes into crates, genre and energy playlists I pull from during a set, or onto the shopping list. It is short on purpose: 35 records, in order, each one a record without which a set I have actually planned is visibly worse. About 700 other gaps from the research notes are parked as reference. By 16 September, 25 of the 35 were bought.
Two rules decide what gets on it. For a known record I buy a version: the remix, the re-edit, the extended 12“. For a modern record I buy the extended original, because radio edits have no place in a club crate. The reel sweep found Soulsearcher’s Feelin’ Love in a Ruze & Chesster remix. I only had the 3:16 original.
The second rule came from a correction, as most of them do (I wrote about that habit in Every correction becomes a rule). The setup once built me ten canonical Detroit and Chicago records to close a gap a critic had named. My verdict: “boring - hard, yes, but not the ones I’d pick when the crowd is high energy.” Since then pedigree is never a reason to buy, and neither is popularity with other DJs. In the September sweep, Route 94’s 4Get The Girl showed up in reels from two separate accounts in one week and came to me as a priority buy. I said “not my style, we dont buy that one.” Two DJs playing a record tells me it works on their floors. Now every line says what the record does to a floor and which record I love it sits next to. The replacement list asked one question per record, does the floor lift or just nod, and brought in Alison Limerick’s Where Love Lives and Chez Damier’s I Never Knew Love (Made In Detroit Mix).
Pairs
Some reels show a transition another DJ found. Those go into a registry of transition pairs. I send a message like “pair: Daft Punk - One More Time >> Prospa - Free Your Mind (filter vocal into rave piano)”. The setup checks both tracks against the library, logs the pair with its direction and the reason, and adds it to a pairs playlist in Rekordbox. In the booth, Rekordbox’s Matching panel remembers the pair: when A is loaded, B shows up.
Direction matters, and the reason matters more. The reason is the part I can reuse in another set.
Seven DJs who disagree
The setup also critiques. It started with five archetype judges: a purist, an A&R, a digger, a booker and a technician. Their first review of my Happy Tunes crates in September was harsh. The purist gave the lane a 5: “a crate is an argument; this is a hard drive.” About 280 of roughly 500 tracks across three crates were actually distinct. I had under two hours of material between 114 and 120 BPM, so the lane could not open a night on its own. The fix was to split the crates by role: 50 openers, 257 peak records, 65 wildcards, no overlap.
Then came seven personas modeled on real DJs, built from their tracklists and interviews: MENU, Melvo Baptiste, Folamour, DJ Tennis, Solomun, Dixon and Diass. Each note lists what I already own from their sets. For Dixon, 3 of 159.
The useful part is where they disagree. Dixon fixes the first two and last two records and throws away about 75 percent live. DJ Tennis says never plan. Melvo wants one unmistakable vocal anthem about 77 percent of the way through. I read a run list against all of them and pick whose advice I take.
The rule
The setup is not allowed to sequence a set. That part is mine.
It did try. For a four-hour night on 26 September it built running orders three rounds in a row, and all three failed. BPM sorted into a slope, songs and artists repeated, and the well-known records landed wherever the sort put them. I told it it was “great at finding good songs and categorizing them” and bad at the order, and that became the rule.
So the work is split. It finds, checks, hunts versions, writes buy lists and builds crates. I do the running order: the energy story, the transitions, and one known record for every five energy records in each block of six.
Which part of your work would you never hand over?