Faceless YouTube Shorts maker with an idea picker

Faceless YouTube Shorts maker with an idea picker

An n8n workflow. Pick a category or type your own idea, choose one of ten hook ideas an AI suggests, and the workflow
researches the topic, writes a scene-by-scene script and has JSON2Video render a faceless short. The finished video's
address arrives by email.

What is inside

• workflow.json: the workflow as n8n exports it, 27 nodes. It holds no credential and calls no sub-workflow. It does not
  hold the JSON2Video template either: you bring your own (below).
• README.md: these notes.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key, for the idea step and the research summary.
• A Tavily API key, for the web search.
• An Anthropic API key, for the script.
• A JSON2Video account, its API key and a template that reads the variables listed below.
• A Gmail account that n8n can send from.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In JSON2Video, make or copy the template, and note its template ID.
3. Open Paste your template id and replace YOUR_TEMPLATE_ID with that ID.
4. Add an OpenAI credential and select it in Submit Idea Prompt to OpenAI, Search (Regenerated) and
   Summarize Web Context.
5. In Search Web for Script Context, replace <YOUR_TAVILY_KEY> after Bearer with your Tavily key. The brackets go too.
6. Add an Anthropic credential and select it in Anthropic Chat Model.
7. Add an HTTP Header Auth credential with the header name x-api-key and your JSON2Video key as the value. Select it in
   Send to Video Generation API and Check Video Status.
8. Add a Gmail OAuth2 credential and select it in Email Video to User.
9. Save, send the form once from its test address, pick an idea, then press Publish at the top of the canvas (or Shift+P).

The form

| Field | Needed | What it does |
|---|---|---|
| Choose a Category | No | A starting phrase or topic, from a list of eight |
| Or Enter Your Own Idea | No | Your own topic. When filled, it replaces the category |
| Duration | Yes | 30 seconds, 1 minute, 1 minute 30 seconds or 2 minutes. It sets the number of scenes |
| Generative Style | Yes | The look of the images, from a list of nine |
| Voice Model | Yes | One of ten Azure voices |
| Audio url | Yes | The address of an audio file for your template to play |
| Email | No, but the last step needs it | Where the finished video's address is sent |

What the template receives

Build Video Creation Payload posts this to JSON2Video's movie endpoint, and your template reads the variables:

{
  "template": "<your template ID>",
  "comment": "<the script writer's note>",
  "variables": {
    "voiceModel": "azure",
    "imageModel": "flux-schnell",
    "audioURL": "<the Audio url you gave>",
    "introImagePrompt": "<the opening image's prompt>",
    "introText": "<the title and the hook, on two lines>",
    "scenes": [{ "voiceOverText": "<one sentence>", "imagePrompt": "<one image>" }]
  }
}

The code also names voiceID, but no value reaches it (see Good to know).

How a run goes

1. The form's data passes through Merge and Prepare Input Variables, which keeps the topic, the audio address, the
   duration, the style and the voice.
2. Generate Video Ideas Prompt builds a request for OpenAI's Responses API with gpt-4.1 and its web search tool, and
   Submit Idea Prompt to OpenAI sends it. A category such as What if or Did you know asks for ten one-sentence hooks
   that start with that phrase. Any other topic asks for ten hook ideas about it. Parse Idea Response JSON strips code
   fences and reads the JSON array.
3. Idea Suggestions shows the ten as a dropdown on a second page. Regenerate Idea Decision reads the second dropdown:
   Yes goes through Build Payload (Regenerated), Search (Regenerated) and Parse Results (Regenerated) and shows a new ten.
   Anything else goes to End Form, which tells the reader the video will come by email.
4. Search Web for Script Context asks Tavily for 10 results on the chosen idea, with an answer and the raw page content,
   limited to 64 sites in include_domains. Summarize Web Context has gpt-4.1 write one paragraph from the first four results.
5. Calculate Scene Count turns the duration into 6, 12, 18 or 24 scenes.
6. Script Generator (AI) is an agent on Claude 3.7 Sonnet. It writes an intro with a title and a hook of at most eight
   words, then one-sentence scenes that each carry an image prompt in the chosen style.
7. Combine Input and Script Output joins the script to the form's data. Extract Script JSON strips code fences and parses
   it, and Build Video Creation Payload shapes the request.
8. Send to Video Generation API posts it to JSON2Video. Wait for Video Rendering pauses 3 minutes, and Check Video Status
   asks for the project.
9. Handle Video Render Status reads movie.status. done and error both go to Email Video to User. running goes
   through Retry Wait Before Recheck, ten seconds, back to Check Video Status.

Change the look

• Categories, styles and voices: the dropdown options in User Input Form. The categories that ask for hook phrases are
  also named in the code of Generate Video Ideas Prompt and Build Payload (Regenerated).
• Scene counts: Calculate Scene Count.
• Research sites: the include_domains list in Search Web for Script Context.
• Tone and structure of the script: the system message in Script Generator (AI).
• The email: subject and text in Email Video to User.
• Waits: Wait for Video Rendering and Retry Wait Before Recheck.
• Models: gpt-4.1 in three nodes, and claude-3-7-sonnet-20250219 in Anthropic Chat Model. Pick a current Claude model
  there if Anthropic reports that one retired.

Good to know

• Every run spends money on OpenAI, Tavily, Anthropic and JSON2Video, and each regeneration spends the idea step again.
• The voice you pick does not reach JSON2Video as exported. Extract Script JSON reads voiceID from the script writer's
  reply, which has no such field, so Build Video Creation Payload sends none. To make the pick count, set voiceID there to
  $("Prepare Input Variables").first().json.voiceID, as the node already does for the audio.
• The person who sends the form picks an idea on the next page, so they stay until End Form appears.
• A failed render still sends the email, with no usable link. The status loop for a render that stays running has no limit.
• The Email field is optional on the form, but the Gmail node needs an address. Without one, the last step has no one to
  send to, after the video is already paid for.
• The agent has Require Specific Output Format switched on, yet no parser is attached, so it answers in plain text and
  Extract Script JSON reads it. If n8n flags the node, switch that option off.
• The original copy's code had lost its backslashes: the error messages in Parse Idea Response JSON,
  Parse Results (Regenerated) and Extract Script JSON, and the .join("\n\n") in Summarize Web Context. This copy has them back.
• Once the workflow is published, anyone with the form's address can start a run on your keys. The form trigger can ask for
  a username and password.
• The topic goes to OpenAI, Tavily and Anthropic, and the script goes to JSON2Video. The finished video is on JSON2Video's address.