Selfie-style UGC product videos from a Google Sheets queue

Selfie-style UGC product videos from a Google Sheets queue

An n8n workflow that makes short selfie-style product videos, the kind a customer might film of something they bought.
Once a day it takes the next product marked Ready in a Google Sheet, has an OpenAI model write the video prompt, and sends
it to Kie.ai for a Veo or Sora video. The finished video's address goes back into the row.

What is inside

• workflow.json: the workflow as n8n exports it, 31 nodes, three of them sticky notes that label the branches. It holds
  no credential and nothing from its author's sheet.
• README.md: these notes.

What we changed from the original copy

• The author's sheet is gone. The four Google Sheets nodes ask for YOUR_SHEET_ID and find the tab by its name,
  Sheet1.
• The original copy had lost the backslashes in the four request bodies that clean up the prompt. They read
  .replace(/r?n|r/g, ' '), which would have turned every letter n and r of the prompt into a space. They are back to
  .replace(/\r?\n|\r/g, ' '), which turns line breaks into spaces as intended.
• Analyze image named chatgpt-4o-latest. OpenAI's deprecations page, read on 2 November 2026, lists it as shut down
  on 17 February 2026, and its listed replacement, gpt-5.1-chat-latest, as shut down on 23 July 2026 in favour of
  gpt-5.6-sol. The node now names gpt-5.6-sol, which reads images.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A Google account and a Google Sheet laid out as below.
• An OpenAI API key.
• A Kie.ai account and its API key.
• A public address for each product photo, such as a link from your shop or an image host. Kie.ai fetches the photo
  itself.

The sheet

One tab, Sheet1, with these headers in row 1:

| Column | What goes in it |
|---|---|
| No | A number unique to the row. The update nodes find the row by it |
| Product | The product's name |
| Product Photo | The photo's public address |
| ICP | The buyer the video speaks to, such as busy parents in their thirties |
| Product Features | The points the speaker can praise |
| Video Setting | Where the clip is filmed, such as a sunny kitchen |
| Model | Veo 3.1, Nano + Veo 3.1 or Sora 2, spelled exactly so |
| Status | Ready for a row to make. The workflow writes Finished |
| Finished Video | Left empty. The workflow writes the video's address |

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Make the sheet, then copy its ID: the long part of its address between /d/ and /edit.
3. Add a Google Sheets OAuth2 credential. In Get row(s) in sheet, Update row in sheet, Update row in sheet1 and
   Update row in sheet2, select it and replace YOUR_SHEET_ID with the ID.
4. Add an OpenAI credential and select it in Analyze image and OpenAI Chat Model.
5. Add a Header Auth credential for Kie.ai: name Authorization, value Bearer followed by a space and your key. Select it
   in the eight Kie.ai nodes: Nano Image, Get Nano Image, Veo Video, Get Veo Video, Veo Video1,
   Get Veo Video1, Sora Video and Get Sora Video.
6. Open Schedule Trigger. The file leaves it on n8n's default, once a day at midnight in the workflow's time zone. Set
   the hour you want.
7. Add one row with Status set to Ready, then press Execute workflow to try it.
8. Press Publish when the test row comes back Finished.

How a run goes

1. Schedule Trigger starts the run, and Get row(s) in sheet reads the first row whose Status is Ready. Only that
   one row is made in a run.
2. Switch reads Model and picks one of three branches.
3. Veo 3.1: Veo Video Prompt1 writes a prompt for an eight-second selfie video, and Veo Video1 sends it to Kie.ai's
   veo3_fast model with the product photo, in 9:16.
4. Nano + Veo 3.1: Veo Image Prompt writes a prompt for a still of a person holding the product. With it, Nano Image
   has Kie.ai's google/nano-banana-edit make that picture from the product photo. Then Analyze image describes the
   picture, and Veo Video Prompt writes the video prompt from the row and the description. Last, Veo Video sends it to
   the same Veo model with the new picture.
5. Sora 2: Sora Video Prompt writes a prompt for a ten-second video, and Sora Video sends it to Kie.ai's
   sora-2-image-to-video with the product photo, portrait, watermark removal on.
6. Each branch asks Kie.ai for its task again and again until the result is there: every 5 seconds for the picture and
   every 10 seconds for a video.
7. The branch's update node finds the row by No, sets Status to Finished and puts the video's address in
   Finished Video.

The four prompt writers share one model, gpt-5-mini in OpenAI Chat Model.

Change the look

• The speaker, the tone and the length of the line: the system message of each prompt writer. The three video writers
  ask for a casual, handheld selfie clip in 9:16 with no phone in the shot and one or two spoken sentences. The image
  writer, Veo Image Prompt, asks for the product in someone's hand, left exactly as in the photo.
• The frame: aspectRatio in Veo Video and Veo Video1, aspect_ratio in Sora Video, image_size in Nano Image.
• Quality: model in the two Veo nodes. veo3_fast is Kie.ai's faster Veo 3.1 tier, and its full tier is veo3.
• When it runs: Schedule Trigger.
• More data in the prompts: add a column and put it in the text of the prompt writer that should read it.

Good to know

• Sora 2: OpenAI's deprecations page also lists sora-2 and its Videos API as removed on 24 September 2026, with no
  replacement. Kie.ai's documentation index, read the same day, lists no Sora model. Expect a
  Sora 2 row to fail at Sora Video, and use one of the two Veo values.
• gpt-5-mini: OpenAI's page lists it for shutdown on 11 December 2026, with gpt-5.6-terra as the replacement. Pick
  that in OpenAI Chat Model before then.
• Kie.ai's current documentation describes a newer Veo call through
  /api/v1/jobs/createTask. The /api/v1/veo/generate and /api/v1/veo/record-info calls this file uses are filed under
  its older version, so check them if the Veo branches stop answering.
• No loop has a way out for a failed task. A picture or video that fails is asked about every few seconds until you stop
  the run under Executions, and its row stays Ready.
• A Model value that none of the three spellings matches leaves the run with nothing to do. That row stays Ready and
  first in line, so the rows after it wait until you fix it.
• The status changes only at the end. If you schedule runs closer together than a video takes, a second run can pick the
  same row while the first is still rendering.
• Sora Video adds two stray closing braces after its prompt expression, so the prompt it sends ends in }}.
• OpenAI bills the prompt writers and the image description. Kie.ai takes credits for each picture and video.
• The product name, buyer, features and setting go to OpenAI, and so does the made picture in the Nano branch. The
  photo's address and the prompts go to Kie.ai.
• The person in each clip is generated and praises the product as if they used it. Check your ad platform's rules and the
  advertising law where you sell before you run one as an ad.