Slop

Skill nadavc2c/SlopMachine/.claude/skills/slop

Drive the `slop` CLI — a local, GPU-accelerated generative-content toolkit (the Swiss-army-knife of generative media, built to be operated by AI agents). USE THIS SKILL whenever the user wants to create, generate, edit, stylize, or animate visual/audio content locally: make/generate an image, apply a style (anime / cyberpunk / casino), the "AI dance" use case (animate a person from a photo + a driving dance video via pose-skeleton motion transfer — the open analog of Viggle/Kling), or find, inspect, and download the current-best open models. Prefer actually running `uv run slop ...` over explaining how. Trigger even if the user doesn't say "slop" — e.g. "make me a picture", "animate this photo", "make him dance", "best image model".From its SKILL.md

Install

npx -y skills add nadavc2c/SlopMachine --skill slop

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its file declares

Copied from the file, not written here

The file declares its own license as See LICENSE / model cards (per-model; recorded in the registry).. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

5.1 KB, as published. Nobody here has run it

slop — generative content, by CLI

slop is a thin, model-agnostic CLI over HuggingFace diffusers. Each capability is a subcommand; the registry (src/slopmachine/config/models.yaml) decides which model backs each one, so commands never change when models do. You operate slop; you don't reimplement it. Run commands with uv run slop ... so everything stays inside the project's uv-managed .venv (never global Python).

How to invoke (read first)

Run from the SlopMachine repo root. Most commands are core:

uv run slop <command> [args]

The one exception is slop dance, which needs the optional pose/video deps at runtime — always run it with --extra dance (a bare uv run re-syncs to the lean core and would uninstall them):

uv run --extra dance slop dance --reference me.jpg --driving clip.mp4 -o outputs/dance.mp4

First, set up (once)

pwsh  ${CLAUDE_SKILL_DIR}/scripts/bootstrap.ps1   # Windows
bash  ${CLAUDE_SKILL_DIR}/scripts/bootstrap.sh    # macOS/Linux

Bootstrap finds the repo, checks uv, runs uv sync --extra dance, and does a GPU health check (slop info). Heavy model weights are not bundled — they download on first use into the repo-local cache. Re-run uv sync --extra dance if slop dance ever reports a missing module.

Command surface (the map)

Command	What it does	Detail
`slop info [--json]`	GPU / VRAM / capability / cache location	—
`slop capabilities [--json]`	List every capability + its default model (start here)	—
`slop styles [--json]`	List style presets (anime, cyberpunk, casino, …)	references/image.md
`slop image "<prompt>" [--style S] [--identity me.jpg] -o out.png`	Text → image (+ optional identity)	references/image.md
`slop dance --reference me.jpg --driving clip.mp4 -o out.mp4`	AI dance: animate a person with a driving video (pose-skeleton motion transfer)	references/dance.md
`slop models list/show/download [--json]`	Inspect / pre-fetch the registry's models	references/models.md
`slop assets download <kind>`	Fetch supporting assets (e.g. `pose` ONNX models)	references/dance.md

Discoverability: every command self-documents — run uv run slop <command> --help for its exact flags and examples before composing a call. --json (where shown) gives machine-readable output.

How to operate

Orient: uv run slop capabilities (what can it do? which model is default?).
Pick the command from the table; read its references/*.md for examples + gotchas, or run uv run slop <cmd> --help.
Ensure weights: the command auto-downloads on first use, or pre-fetch with uv run slop models download <capability> (and uv run slop assets download pose for dance).
Run it, always passing -o <path> so the output location is explicit, and surface the Saved: <path> line to the user.

Operating notes (set expectations, don't treat as a hang)

First run of any model downloads large weights (default image ~24GB; the dance model ~77GB) — minutes, one time, into the repo cache.
slop dance is slow on 16GB (CPU↔GPU layer streaming): expect minutes per clip. It is not hung. Keep clips short and resolution modest; raise --steps only for final quality.
On CUDA out-of-memory, retry with fewer frames (--fps lower), a shorter clip, or fewer --steps. The defaults are tuned to fit 16GB; larger settings may not.
Errors are single-line Error: ... with a non-zero exit code — read them; they're actionable.

Extending

New capability = a Stage in src/slopmachine/pipeline/ + a row in src/slopmachine/config/models.yaml + a slop <verb> command (see CLAUDE.md "Adding a new capability"). To refresh a model to the current best, use the slop-models skill.