Coolhunting
THE LABEL THAT DISCOVERED ITSELF.
Seven years of archives revealed a version of the creative company nobody had seen.
SÃO PAULO, AUGUST 2026
.TBF. is one year old. Its archive is seven. Before the company existed there were already years of research and references.
Thirteen thousand files: photographs, videos, screen grabs, recorded meetings.
One night in August we set a machine to listen to all of it.
Nobody had ever heard the whole archive: forty-three hours of audio, thirty-eight of them speech.
The tool was Whisper, an open transcription model, running on our own Mac. No cloud.
It started at eleven at night and finished at six in the morning.
It read the 3,132 files that had audio, and failed on none.
We wanted to search inside what was said.
What came back was somebody else's voice.
370,700
words said out loud across seven years
14,135
of them said into our own camera
Thirty-four of the thirty-eight spoken hours came from people who do not work here.
Saved reels, other people's courses, podcasts, tutorials.
Our own camera accounts for two and a half. The last hour and a half was never classified.
What it films is a party, a shoot, a trip. It films gesture.
We built an archive of our own voice and found somebody else talking in it.
Inside those hours, one in every seven words was said in English.
Nobody here knew that number existed.
And that English is almost all borrowed: of 55,260 words spoken in English, 50,582 come from saved material.
Art direction on TikTok, an After Effects tutorial, a sales reel, a prayer in Hebrew.
In the few English clips our own camera shot, the English is music at a party.
None of it was addressed to anyone. It is a record of what this label watches.
Chapter 1
The files had always been there, thirteen thousand of them, across seven years and four business units.
That night we gained no files. We gained a way to ask.
Before, you opened folders and tried to remember filenames.
Now you type "show me the starred photos from the client" and the archive answers.
The search looks inside what was said and inside what is written on screen.
An archive is not a repertoire.
An archive is what you kept. A repertoire is what you can connect.
Chapter 2
Everyone says AI made the tools the same. It made the inputs the same too.
EVERYONE HAS
- The same public corpus.
- The same references, three clicks deep.
- The same trend report, read by everyone.
- The same question, phrased the same way.
ONLY WE HAVE
- Seven years of recordings nobody else has.
- What we refused, and the day we refused it.
- The first client, who is still a client.
- Our mistakes, dated and kept.
Everyone feeds the machine what everyone has, and the machine hands back the average.
What only we have fits in 14,135 words. Our own tape counted them.
We spend our days in the same corpus as everybody else. The difference is that ours got written down.
Nobody escapes the average by writing a prettier prompt.
Chapter 3
Feeding the machine our own archive was half the job.
The other half was teaching it what this label refuses:
a manifesto is memory, and memory fails at some rate, and never out loud.
So here the rules stopped being remembered and started being run.
A GUARD is a small script that sits between the work and the world.
One refuses the dash this label does not use. One blocks the road to production.
One counts how many of our rules are code, and how many are still hope.
Culture that does not become code is decoration.
One counterweight, learned the hard way: a guard that guesses is worse than no guard at all.
One of those guards once rang on a file that had only been read, not written.
An alarm that rings on the wrong file teaches everyone to ignore alarms.
So a new rule only enters if it is born with a test, or retires another one.
The oldest rule here is the cheapest: a signal that already exists beats a new signal.
Cameras already write the date, the lens, the coordinates.
The field that told us whose voice each file held had been sitting in the catalogue the whole time.
Chapter 4
The alarm that rings on the wrong file has a twin: the one that never rings at all.
The same night the archive learned to answer, it also lied.
A file name with an accent can be stored two different ways.
Our database used one. The script feeding it used the other.
On screen they are identical. Byte by byte they are not.
44
files that got their text and stayed invisible to the search
We caught it before the batch was applied, and fixed it.
Had we not, a search for those files would have returned nothing, and an empty result looks exactly like an empty archive.
We would have concluded the archive was empty in the place where it was fullest.
Chapter 5
The machine did seven hours of work that night.
What thirty-four hours of borrowed speech mean for this label is still open.
Producing got cheap. Choosing got expensive.
Because the machine does not choose.
It has no provenance, and provenance cannot be generated.
It widens the field. Somebody still has to frame the shot.
Inside those 370,700 words, recorded on 27 March 2025:
"People keep asking how I do my work. I just answer: by paying attention."
Ninety-eight seconds of tape, said in Portuguese, filed under a portrait shoot, unheard for seventeen months.
It took a machine reading forty-three hours to hand back one sentence we already owned.
The archive answered: 14,135 of the 370,700 words are ours. What to do with the rest is not its question.
