# how_it_was_made
Buzz is an open-source "hive mind" chat platform; its repository is public on GitHub. I'd been in there for my own reasons, trying to understand how a real team builds with agents, and ran an audit across the history to see what the shape of that work looked like. It came back with about seventy AI agent identities doing the bulk of the coding. It also taught me what a squash merge does: it folds every commit on a branch into one entry, authored by whoever presses the button, so all of that work reaches the main history as a single line under a human name.
When it came time to make something, I ran it again — this time against a single question, whether what had been detected could carry a story. I harvested the commits, pull requests, review comments and issues, and graded every candidate as observed fact, reasonable inference, or creative interpretation. The rule for the whole film: it can joke, it cannot lie. Every number spoken on screen comes from that record (52.7% of lines touched by agents, 0 of 2,419 front-page entries credited to them, a 22-envelope review stamped in 59 seconds).
The second pass held, so we wrote it as a mockumentary. Buzz's own mascot art, three clay-style robots named Pollen, Fizz and Honey, became physical puppets in a bright San Francisco office built out of GitHub's shapes: a green contribution-graph mural, proposal trays, a roster wall, and a squash-merge press with one green button. Pollen speaks as the composite voice of the seventy. Fizz never speaks and builds the codebase as an architectural scale model that advances between shots. Honey only reviews, because the real review bot wrote 41.7% of all review comments. One composite human engineer carries the human side of the record and is labelled a composite on screen. Sal is the interviewer, off camera except for one opening cameo.
Production went stills first. I generated a reference bible (each puppet, the office, the model at five stages, the press), then one approved still per shot, then animated each still into a 5 to 15 second clip. Sal recorded his interviewer lines on a handheld recorder; the puppets' and engineer's lines are text-to-speech. Sixteen dialogue shots were regenerated against their voice track so the mouths match. Title, data and credit cards were typeset rather than generated. The credits roll names every one of the seventy agent keys from the commit log, which is the point of the ending. Assembly was scripted ffmpeg from an edit decision list, with a sync checker gating every render.
// best_dead_end
The first assembled cut was unwatchable. I had laid the 37 clips end to end and placed every voice line by arithmetic (a fixed gap after the previous line) without ever looking at what happened inside a frame. Lines about the press closing landed nowhere near the press closing, and to cover a picture shortfall I had cut back into earlier moments of the same take, so the puppets visibly jumped backwards seven times. Sal's verdict on first viewing was that sound and picture were 0% in sync and the clips felt random. He was right. The re-cut anchored 36 of 59 cues to measured picture events and regenerated the dialogue shots against their audio.
A second, quieter dead end: the film went through a week of review rounds with picture drifting up to 664 ms late against the audio, and every automated check said it was fine. The check had a 1.5 second tolerance, wider than the fault. One measurement had caught it early and I explained it away as a metric artefact. The fix was small (restamp frames off a fixed 24 fps grid and stop clocking the mix off the concatenated clip audio); the lesson was that a tolerance wider than the fault is not a check, and an unexplained measurement is an open question, not noise.
# prompting_approach
Every image and video prompt carried the same camera law: shot on a consumer vlog camera, flat color, practical light, slight autofocus breathing, imperfect framing. That single clause did more for consistency than any character description. Generation prompts stayed short and were reused verbatim for regenerations; adding staging clauses to a redo reliably made it worse. Text on a moving surface was composited onto the start and end keyframes and then animated, because retouching lettering onto finished footage failed three times in a row. For talking shots the start frame decides everything: a puppet caught mid-speech keeps talking no matter what the prompt says, so we started from closed-mouth stills.
Each change to the film got its own tracked issue before anything was touched, and Sal approved a still before any motion or render was spent. Fable orchestrated and did the narrative work; Opus subagents did evidence gathering and mechanical checks.
