Skip to main content
The ralphy-assets companion repo at github.com/alecs5am/ralphy-assets is the source of truth for reusable heavy assets — gameplay loops, Italian Brainrot character references, trend music, stock b-roll, and complete example projects. The repo is organized as a pool: generic, by kind, not curated per-project. The agent’s rule (AGENTS.md invariant #12) is to consult the catalog before writing a prompt that names a specific character, footage, or track. If the asset exists in the pool, pull it. If it does not exist for a real-world entity, propose adding it rather than generating drift from text.

​
Why a pool, not curated

A curated “per-template assets” model would mean the agent picks assets after committing to a template. The pool model puts the asset selection before the prompt — the agent surveys what is already canonical, then writes the prompt around the pulled asset. Three reasons this matters. Identity drift. A “Tralalero Tralala” generated from text alone produces five different sharks across five scenes. The same Tralalero pulled from the pool — same canonical image, passed as --ref on every i2v call — locks identity. Recognizability. The Soviet Nostalgic template requires the canonical Soviet trend music bed. Generating a substitute “Soviet-sounding” track in ElevenLabs Music loses half the format’s appeal — the recognizability is the bit. The pool ships the exact tracks that work on TikTok in 2026. Cost. A pre-rendered 56-second CS:GO surf loop is ~21 MB. Generating a substitute in Veo or Seedance is $1-3 and 90 seconds; pulling from the pool is free and instant.

​
The three layers

The companion-repo manifest splits assets into three layers: required — template-bound mandatory assets. Auto-pulled when you run ralphy template use <slug>. Listed in the catalog for transparency; you rarely pull these directly. Today:
TemplateRequired assetSize
brainrot-ai-memecs-surf-loop.mp421.49 MB
soviet-nostalgictrend-soviet-bed.mp30.94 MB
pool — generic, reusable assets grouped by kind: open-ended categories like italian-brainrot-characters, gameplay-loops, trend-music, stock-broll. Pulled on demand. The two pools with the most entries today:
  • italian-brainrot-characters (33 entries) — canonical Italian Brainrot meme character reference images. Used by the italian-brainrot template as i2i references for character regeneration.
  • trend-music (27 entries) — recognizable horror / dread music beds heavily used on horror-TikTok in 2024-2026, sourced from the canonical Horror TikTok YouTube playlist. Each ships the full mp3 plus a pre-rendered 20-second hook clip under hooks/.
examples — ready-made example mp4 outputs per template, for visual reference and regression baselines. Currently empty in the manifest; future templates will ship example renders here. The catalog file at docs/assets-catalog.md is auto-generated from the manifest. Regenerate it after any manifest change with ralphy assets catalog --write.

​
The pull flow

# Survey what is in the pool.
ralphy assets list                                       # everything
ralphy assets list --kind italian-brainrot-characters    # one kind
ralphy assets list --template italian-brainrot           # required + matching pool items

# Pull one pool item to the cache.
ralphy assets pull-pool italian-brainrot-characters/tralalero-tralala

# Pull and copy into a project's assets/uploaded/ directory.
ralphy assets pull-pool italian-brainrot-characters/tralalero-tralala \
  --install coffee-with-tralalero-001

# Pull all required for a template.
ralphy assets pull brainrot-ai-meme

# Cache management.
ralphy assets cache-info                  # cache size, last-pulled
ralphy assets clean                       # wipe the cache (safe — re-pull on demand)
--install <project> is the form the agent uses most. It pulls the asset to the cache (or no-ops if already cached), then copies it into workspace/projects/<project>/assets/uploaded/, then logs a user-assets.jsonl entry. The downstream generate verbs see the asset at the conventional path and use it as a ref.

​
The cache

workspace/.ralph/asset-cache/ is the local cache of the companion repo. Layout: 
workspace/.ralph/asset-cache/
├── pool/
│   ├── italian-brainrot-characters/
│   │   ├── tralalero-tralala/
│   │   │   ├── tralalero-tralala.png
│   │   │   └── ref.json
│   │   └── ...
│   └── trend-music/
│       ├── caretaker-its-just-a-burning-memory/
│       │   ├── caretaker-its-just-a-burning-memory.mp3
│       │   └── hooks/
│       │       └── caretaker-its-just-a-burning-memory.20s-hook.mp3
│       └── ...
└── required/
    ├── brainrot-ai-meme/
    │   └── cs-surf-loop.mp4
    └── soviet-nostalgic/
        └── trend-soviet-bed.mp3
Every asset has a SHA-256 in the manifest. ralphy assets pull-pool verifies the hash after download — corrupt or tampered files fail with a clear error. The companion repo is public and unauthenticated — no API key needed. ralphy assets clean wipes the cache. It is safe to run at any time — the next pull re-downloads.

​
The rule that fires the catalog check

AGENTS.md invariant #12, paraphrased: before writing a prompt that names a specific real-world entity, run ralphy assets list --kind <kind> or read the catalog. If the asset exists in the pool, pull it via ralphy assets pull-pool instead of generating from text. If it does not exist for a real-world entity (e.g. a specific brainrot character not in the pool), propose adding it to the pool first, rather than generating a drift-prone reference from text alone. Three concrete cases. Case 1 — user asks for a Tralalero Tralala video. The agent runs ralphy assets list --kind italian-brainrot-characters. Tralalero is in the pool. The agent pulls it with --install, scaffolds the project from the italian-brainrot template, and writes prompts that pass the Tralalero PNG as --ref on every i2v call. Case 2 — user asks for a horror TikTok with the caretaker-its-just-a-burning-memory music bed. The agent runs ralphy assets list --kind trend-music. The track is in the pool. The agent pulls the 20-second hook clip from hooks/, scaffolds an analog-horror PSA, and drops the hook into the editor stage. Case 3 — user names a brainrot character that is not in the 33-character pool. The agent does not attempt to generate the character from text. The right move is to either (a) propose adding it to the pool — find the canonical reference image, register it with ralphy assets register-pool italian-brainrot-characters/<slug>, and submit upstream — or (b) treat it as a ref-required case and ask the user for a reference image to drive a one-off project.

​
License posture

The pool’s category default for Italian Brainrot characters and trend music is fair-use-meme-reference with attribution. The characters are AI-generated TikTok memes from 2024-2025; the music is from a canonical “Horror TikTok” YouTube playlist. The licensing posture is:
  • Each asset’s ref.json carries its specific license tag.
  • fair-use-meme-reference covers reference use inside generated content where the character / track is recognizably part of the meme format being referenced.
  • For commercial publication — branded ads, paid placements — the agent should warn the user that the fair-use posture is meme-context-specific and may not extend to commercial use.
Always read each asset’s ref.json before publishing. The companion repo will flag breaking-license cases in the manifest.

​
The catalog file

docs/assets-catalog.md is the human-readable export of the manifest. It is auto-generated by ralphy assets catalog --write. The agent reads this file (or runs ralphy assets list for live data) before drafting any prompt that names a specific real-world entity. Regenerating after a manifest change is part of the asset-add workflow.

​
Example projects in the companion repo

The ralphy-assets companion repo also ships complete example projects under examples/<project-id>/ — the full workspace/projects/<id>/ tree as a tarball, including the rendered mp4, the prompts, and the logs. These are the source artifacts behind the eleven shipped-project-derived templates listed in /concepts/templates. Pull one with: 
ralphy example list                            # list available example projects
ralphy example pull noski-people-001           # pull into workspace/projects/
This is the fastest way to see exactly what a finished project looks like — including the postmortem, the gen-log, and the asset versioning history.