MoneyPrinter Turbo — Claude Skill Hub

SKILL.md

MoneyPrinterTurbo

Upstream: <https://github.com/harry0703/MoneyPrinterTurbo> (Python 3.11+, FFmpeg). It takes a subject, has an LLM write a script and search terms, pulls matching footage, synthesises a voiceover, burns subtitles, mixes BGM, and encodes an MP4.

Drive it through cli.py, not the WebUI. The WebUI (webui.sh, port 8501) is interactive and you cannot read its state. The CLI is headless, takes every setting as a flag, and prints a JSON result to stdout. Use the REST API only when the user explicitly wants a long-running service.

1. Locate or install

Look for an existing checkout before cloning anything — users usually have one:

ls config.toml cli.py webui 2>/dev/null            # already inside it?
ls ~/MoneyPrinterTurbo/cli.py 2>/dev/null

If there is none, ask the user where to put it, then:

git clone https://github.com/harry0703/MoneyPrinterTurbo.git
cd MoneyPrinterTurbo
uv python install 3.11
uv sync --frozen

Every command below runs from the project root and is prefixed uv run. If the project was set up with a plain venv instead, activate it and drop the uv run.

Docker is the other supported path — docker compose -f docker-compose.release.yml up serves the WebUI on 8501 and the API on 8080, but gives you no CLI, so prefer a local checkout for generation work.

2. Configure `config.toml`

cp config.example.toml config.toml, then fill in only what the requested run needs. See references/config-reference.md for the full key list.

Minimum for a default run:

| Setting | Why | | --- | --- | | llm_provider + <provider>_api_key | writes the script and search terms | | pexels_api_keys | supplies the footage (free key, array of strings) |

subtitle_provider = "edge" and the default Edge TTS voice both cost nothing and need no credentials — leave them alone unless asked.

Never invent an API key, and never read one out of the user's shell history or another project's config. If a required key is missing, stop and ask for it. Keys go in config.toml, which is gitignored upstream — do not commit it, and do not echo key values back in your replies.

3. Gate anything that costs money

Free by default: Edge TTS, Pexels/Pixabay/Coverr footage, local whisper.

Billable, and each needs the user's explicit go-ahead before the run — say which provider bills and roughly per what (per clip, per character):

--video-source volcengine_seedance and WaveSpeed AI — paid text-to-video,

billed per generated clip. Seedance additionally refuses to start without --confirm-seedance-charge; treat that flag as the user's consent, never as a box for you to tick on your own.

ElevenLabs, Azure, Fish Audio, SiliconFlow, Gemini and MiMo TTS.
Whichever LLM provider is configured.

--video-count 5 multiplies every one of those by five. Confirm the count.

4. Preview the script before rendering

A full render spends TTS characters and minutes of FFmpeg on a script nobody has read yet. Stop early, show the script, then continue:

uv run python cli.py --video-subject "How AI is changing everyday life" \
  --video-language en-US --paragraph-number 3 --stop-at script

--stop-at accepts script, terms, audio, subtitle, materials, video. Show the generated script, take edits, then render with the approved text passed verbatim through --video-script so the LLM does not rewrite it:

uv run python cli.py --video-script "<approved script>" \
  --video-terms "ai,technology,city" \
  --video-aspect 9:16 --video-clip-duration 5 \
  --voice-name en-US-JennyNeural --bgm-type random --bgm-volume 0.2 \
  --subtitle-position bottom --font-size 60 --stroke-width 1.5

Skip the preview only when the user hands you a finished script.

On success the CLI prints {"task_id": "<uuid>", "result": {...}} and the artifacts land in storage/tasks/<task_id>/. Report the MP4 path — do not just say it finished. Exit codes: 0 success, 1 task failure, 2 bad arguments or manifest.

Full flag list, with defaults and choices: references/cli-and-api.md.

5. Batch

For more than two or three videos, write a manifest instead of looping the CLI — one process, one JSON summary, and a per-task failure report:

uv run python cli.py --batch-file tasks.jsonl

JSON array or JSONL, up to 100 tasks and 1 MiB. Each object sets VideoParams fields in snake_case (video_subject, video_script, video_aspect, …); unknown fields are rejected, every task needs video_subject or video_script, and relative paths resolve against the manifest's directory. --batch-file and --task-id are mutually exclusive.

The summary reports succeeded/failed with a failed_stage per task. A batch that reports failures has not succeeded — surface the failing tasks and their stage rather than only the total.

6. API service mode

uv run python main.py     # 127.0.0.1:8080, interactive docs at /docs

Endpoints are under /api/v1 — POST /videos starts a task and returns a task_id, GET /tasks/{task_id} polls it, GET /download/{file_path} fetches the result. Poll with a bounded number of attempts and a real interval; never busy-loop. Endpoint table in references/cli-and-api.md.

7. When a run fails

Check references/troubleshooting.md before improvising — FFmpeg not found, Too many open files, a stalled whisper model download, and non-ASCII project paths on Windows cover most failures, and each has a known fix.

Two rules for failures: a render that produced no MP4 is a failed run even if the process exited quietly, so verify the file exists before reporting success; and if the LLM step fails, read the actual provider error rather than switching providers blindly — an unset key and a rate limit need different fixes.

When Claude uses it

SKILL.md