Sales CSV turned into a Google Slides deck of AI images

Sales CSV turned into a Google Slides deck of AI images

An n8n workflow with an upload form. You send a CSV of sales data and choose from one to ten slides. GPT-5.2 writes an
executive report from the rows, then a picture prompt for each slide. WaveSpeedAI draws the slides, and the workflow
puts each picture into a new Google Slides deck as a full-slide background.

What is inside

• workflow.json: the workflow as n8n exports it, 19 nodes and no sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

The copy Mindket started from had lost its backslashes in three places. Mindket put them back:

• Edit Fields1 reads the slide's number from its label with replace(/\D+/g, ''). The copy had /D+/g, which only
  removes a capital D, so "Slide 1" stayed as it was. The insertion index in Add Blank Slides then came out as NaN, the
  request body was not valid JSON, and the run stopped at the first slide.
• Create New Blank PPT names the deck after the file with replace(/\.[^/.]+$/, ''). The copy's /.[^/.]+$/ still
  cut .csv from most names, but it emptied a name that had no extension.
• The system message of Message a model asks for \n as the line break inside each slide's text, and its example
  answer uses it in all ten slide prompts. Both had lost the backslash, and four quoted lines in the example had lost
  theirs too. The example is valid JSON again.

The two WaveSpeedAI requests held Bearer <token> in their Authorization header. They now read
Bearer <YOUR_WAVESPEED_KEY>.

In the four Slides requests, the capital O in the name of the Google Slides OAuth2 credential type is written as the
JSON escape \u004f. n8n reads the same name. The escape only stops a store check from taking that long name for a
secret key.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key that can use gpt-5.2.
• A WaveSpeedAI account and its API key.
• A Google account for the Google Slides OAuth2 API credential. On your own n8n server, the Google Cloud project behind
  it needs the Google Slides API turned on.
• A CSV with a header row. Every row goes to the model in one message, so keep it to a table a person could read.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Google Slides OAuth2 API credential and sign in. Select it in Create New Blank PPT, Delete First Slide,
   Add Blank Slides and Update Image to PPT.
3. Add an OpenAI credential and select it in Message a model1 and Message a model.
4. Open Create Task and Check Status. In each Authorization header, replace <YOUR_WAVESPEED_KEY> with your key,
   brackets included. Keep the word Bearer and the space after it.
5. Save, open the form's test address, upload a small CSV and choose 2 slides.
6. Press Publish, then share the form's production address.

The form confirms at once, and the run goes on behind it. Open the run under Executions. Presentation URL holds the
link to the deck.

How a run goes

1. On form submission is the form, titled Presentation Generator AI. It takes one .csv file and a choice from
   1 slide to 10 slides.
2. Create New Blank PPT makes a deck named after the file with - Presentation added, and Delete First Slide
   removes the slide every new deck starts with.
3. Extract from File reads the CSV, with the first row as the headers, and Aggregate gathers the rows into one list,
   data.
4. Message a model1 sends that list to gpt-5.2 and asks for an executive sales report in Markdown.
5. Message a model sends the report and the slide count to gpt-5.2 with a long design brief. A JSON schema holds
   the answer to three parts: BaseStyle_prompt, Negative_prompt and Slides, each slide a label and a prompt.
6. Prepare prompts unwraps the answer if it comes back as an array, and Split Out makes one item for each slide.
7. Loop Over Items takes one slide at a time. Edit Fields joins the shared style, the slide's prompt and the list of
   things to avoid.
8. Create Task starts a flux-2-flex text-to-image task on WaveSpeedAI at 1536*864. Wait pauses 30 seconds,
   Check Status asks for the result, and Switch1 sends a task that is still processing back to Wait.
9. When the task is completed, Edit Fields1 keeps the picture's address, the slide's number and the deck's ID.
   Add Blank Slides inserts a blank slide at that number, and Update Image to PPT sets the picture as its stretched
   background.
10. After the last slide, Presentation URL builds the deck's link.

Change the look

• The report: the message in Message a model1. Rewrite it for data that is not about sales.
• Every slide's look: the system message of Message a model. It holds the style rules, twelve layouts to choose from,
  the text limits for each slide and an example answer about a film studio.
• Picture size: size in the body of Create Task. 1536*864 is 16:9, the shape of a standard slide.
• The drawing model: the address in Create Task. Another WaveSpeedAI model needs the same answer shape, a link in
  data.outputs[0].
• Waiting: Wait's 30 seconds. A shorter wait asks more often.
• The form: its title, the file field and the slide choices in On form submission. The deck's name reads the file field
  by its label, Upload_the_CSV_you_d_like_to_transform, so change Create New Blank PPT too if you rename it.

Good to know

• Each slide is one picture. Its titles, bullets and charts are drawn by the image model, so they cannot be edited in
  Slides, and a word or a figure can come out wrong. Compare every slide with the report from Message a model1.
• Switch1 sends a failed task nowhere. The loop then stops, the later slides are never made, and Presentation URL
  never runs.
• A WaveSpeedAI status other than processing, completed or failed matches none of Switch1's three outputs, and n8n
  stops the run with an error.
• The deck sits in the Google Drive of the signed-in account and is shared with no one. Share it from Slides yourself.
  The person who sent the form only sees its confirmation.
• OpenAI's deprecations page, read on 3 November 2026, gives no shutdown date for gpt-5.2.
• Billing: OpenAI bills two gpt-5.2 calls a run, the first with your whole CSV in it. WaveSpeedAI's page for
  flux-2-flex listed $0.06 an image on 3 November 2026, and each slide is one image.
• Every row of the CSV goes to OpenAI. The slide prompts, with your figures in them, go to WaveSpeedAI, and Google
  fetches each picture from WaveSpeedAI's link.
• Once the workflow is published, anyone with the form's address can start a run that spends your keys and adds a deck
  to your Drive.
• The WaveSpeedAI key sits in two header fields, not in a credential, so it travels with any export. Take it out before
  you share the workflow.