brainrot

A serialized short-form (TikTok-shaped) content mill: ideate → score → plan → make.

• Breadth (ideate) runs a studio development funnel over a niche — a writers' room fans out series concepts (a recurring cast plus an overarching arc), an adversarial producers' table argues them down, an audience team projects metrics, and a greenlight committee scores them. The best are kept.
• Depth (make) writes one episode of a kept series as a multi-character scene (a two-pass writers' room: draft → punch-up) and renders it into a vertical 1080×1920 AI video that stands alone and advances the larger story.

Built as a set of Go-DSL pipelines on top of vibe: vibe supervises the model backends (LLM, ComfyUI image/video, Kokoro TTS) under capability profiles, vamp orchestrates the DAGs, and contentkit provides the shared studio primitives (tournaments, critique/revise, scorecards, review loops).

Prerequisites

brainrot is the pipeline, not the inference stack. It needs a running vibe stack:

• vibe installed and running. brainrot depends on the published github.com/gallowaysoftware/vibe module (v0.7.1+) for its vamp/contentkit packages, and drives a running vibe daemon at runtime:

  go install github.com/gallowaysoftware/vibe/cmd/vibe@latest
  vibe daemon &
  

  Configure a long_form text profile (suggested: a 27B-class model such as Qwen3.6-27B-MTP at 128k+ context).
• ComfyUI on :8188 for Qwen-Image stills + Wan2.2 image-to-video (vibe start comfyui). The two workflow graphs are bundled under internal/pipeline/workflows/.
• Kokoro-FastAPI TTS on :8880 for character narration (vibe start tts_kokoro).
• ffmpeg on $PATH for assembly (stills → video → voice → captioned MP4).
• ~30GB VRAM during generation. The LLM and the image model don't co-reside; make unloads the LLM (vibe stop) before phase 2 loads ComfyUI.

Run brainrot doctor to see what's up and what's missing; brainrot activate brings up the declared profile + services via vibe.

Install

go install github.com/gallowaysoftware/brainrot/cmd/brainrot@latest

Or from a checkout: go build ./... then go install ./cmd/brainrot.

Workflow

brainrot activate # bring up long_form + comfyui + kokoro
brainrot ideate --niche "liminal office horror"# develop + score series, keep the best
brainrot list # kept series with scores + episode progress
brainrot make <series-id># write + render the next episode

ideate prints the kept series IDs and the full studio trace (concept pool, producer debate, projected metrics) so the run is inspectable. make with no --episode writes the next unrendered one; the finished final.mp4 path is printed at the end.

Optionally: mine a canon-backed series for episodes

A series with a lore.md canon pack (and optional source.txt seed) can be mined for a deep, fresh, ranked slate of episodes:

brainrot plan <series-id># stage a ranked slate (no changes)
brainrot plan <series-id> --review # curate the slate accept/reject, one at a time
brainrot plan <series-id> --apply # append the whole slate to the bible

Mining stays fresh — it won't repeat episodes already in the bible (an anti-repeat guard drops title collisions mechanically, not just by prompt).

Commands

Command	What it does
ideate --niche <s>	Develop + score series for a niche, keep the top ones. Flags: --count (shortlist size, default 6), --keep (default 4).
list	List kept series with scores and done/total episode progress.
plan <id>	Mine a series' canon (lore.md + source.txt) into a ranked slate of episode beats. Flags: --count (default 8), --apply, --review.
make <id>	Write + render one episode as a vertical video. See flags below.
activate	Bring up the vibe profile + services brainrot needs.
doctor	Read-only: what's running, what's missing.

make flags:

Flag	Effect
--episode N	Which episode (1-based); default = next unrendered.
--shots N	Shots in the episode (default 7; a series can set its own).
--candidates N	Write N independent candidate scripts; an editor/producer/money panel picks one to render.
--finalists K	Two-stage funnel: render the top K panel finalists and let a vision judge pick the best actual render (requires --candidates > K).
--narrator <voice>	Default Kokoro narrator voice.
--publish-to <dir>	Copy the finished final.mp4 into <dir> as <id>-<NNN>.mp4.
--preview	Stills only: shot list + per-shot images, skip animation/voice/assembly (fast iteration).
--script-only	Phase 1 only: write the script, leave the LLM loaded.
--render-only	Phase 2 only: render the existing shots.json (use after --script-only).

How make renders

Two phases so the LLM is unloaded before the image model loads (they can't co-reside on 32GB):

1. Phase 1 (LLM) — draft the shot list, then punch it up. With --candidates > 1, write several and let a greenlight panel pick. Narration is mechanically cleaned (stage directions / shouty labels stripped so they don't reach the TTS or the burned captions).
2. Phase 2 (ComfyUI + Kokoro + ffmpeg) — Qwen-Image stills → Wan2.2 image-to-video → Kokoro voiceover → captioned, assembled MP4.

Series library

Kept series live under $XDG_STATE_HOME/brainrot/ (default ~/.local/state/brainrot/):

series/
<series-id>/
series.json the bible: cast, arc, episode beats, scores
lore.md optional canon pack (enables `plan` mining)
source.txt optional raw seed mined alongside lore
anchors/<slug>.png optional fixed character portrait (used verbatim for monologue narrators)
episodes/<NNN>/ per-episode run dir (shots.json, images/, final.mp4)
_planned/<stamp>/ staged `plan` slates for this series
_dev/<stamp>/ full studio trace from an `ideate` run
_rankings/<stamp>.json

The bible is yours to hand-edit; brainrot appends to it (via plan) but doesn't clobber your edits.

License

MIT — see LICENSE. The pipeline is Galloway Software's; the content you generate with it is yours.