Online course writer with research, scripts and images

Online course writer with research, scripts and images

An n8n workflow in two parts. The first form takes a course topic, a module count and a word count. Tavily researches
the topic, OpenAI writes the modules and a voice-over script, and both become Google Docs while the web images found
along the way go to Google Drive. A second form takes the script and draws pictures for each module with gpt-image-1.

What is inside

• workflow.json: the workflow as n8n exports it, 31 nodes in two parts, each starting at its own form, and two of the
  nodes are sticky notes. It holds no credential. The two Tavily keys and the address of the Markdown-to-Google-Docs
  converter are placeholders.
• README.md: these notes.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key. Your OpenAI organization may need to be verified before gpt-image-1 answers.
• A Tavily API key.
• A Google account, connected to n8n through a Google Drive and a Google Docs credential.
• A converter that builds a Google Doc from Markdown, or the Google Docs nodes that can stand in for it (below).

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential and select it in OpenAI Chat Model, OpenAI Chat Model1, OpenAI Chat Model2,
   OpenAI Chat Model3, OpenAI Chat Model4 and Generate image1.
3. In Tavily web search and Tavily web search1, replace <YOUR_TAVILY_API_KEY> in the Authorization header with
   your key. The brackets go too, so the value reads Bearer, a space and the key.
4. Add a Google Drive OAuth2 credential and select it in Create a Folder, Upload to Folder, Create folder and
   Upload file.
5. Set up the two converter nodes, Format and Update to Doc and Format and Update Doc, as the next section explains.
6. Save, send the course form once from its test address with four modules of 300 words, then press Publish.

The converter it calls

The file does not include it. Format and Update to Doc and Format and Update Doc each post a JSON body with two
fields to YOUR_CONVERTER_URL, and they expect a service that answers by making a Google Doc.

| Field | What the node sends |
|---|---|
| output | A short JSON text holding the Markdown and the course topic |
| fileName | Course content : <topic> from the first node, Voice over script : <topic> from the second |

Both nodes authenticate with a Google Docs credential, so n8n adds its access token to the request. The original copy
sent them to a converter on its author's own server. You have two choices:

• Run a converter of your own, put its address in both nodes and select your Google Docs credential in each.
• Or delete each node and put two Google Docs nodes in its place: one set to Create, with the file name above as the
  title, then one set to Update that inserts the text. The Markdown headings then stay as plain text, but no outside
  service sees your token.

Until you do one or the other, a run stops at Format and Update to Doc, before the voice-over script is written.

How a course run goes

| Stage | Nodes | What happens |
|---|---|---|
| Instructions | On form submission, Instruction Generator | gpt-4.1-mini writes three sets of instructions from your answers: for the outline, the modules and the voice-over |
| Research | Tavily web search | One search on the topic, with a short answer and image addresses |
| Outline | Course Outline, Structured Output Parser, Split Out | gpt-4o-mini plans the modules, each with a title and a description, and each module becomes an item |
| Module research | Tavily web search1 | One search per module title, with the pages' full text and images |
| Modules | Course content, Merge the content | gpt-4o-mini writes each module with image links under the lines they fit, and the modules are joined into one text |
| Course doc | Format and Update to Doc | The converter makes Course content : <topic> |
| Voice-over | Voice over script, Format and Update Doc | gpt-4o-mini rewrites the course to be spoken aloud, and the converter makes Voice over script : <topic> |
| Web images | Create a Folder, Code, Get images, Upload to Folder | Every image address from the module searches is downloaded into a Drive folder named <topic> - Web images |

Get images and Upload to Folder carry on after an error, so a picture that will not download is skipped.

The picture run

The second part starts at On form submission2, its own form, once the voice-over script exists.

1. Open the form, enter the course title, paste the whole voice-over script and choose 5, 10 or 15 pictures per module.
2. Create folder makes a Drive folder whose name is the title followed by AI Images.
3. Image Prompts1, on gpt-4o-mini, writes that many picture prompts for each module.
4. Split prompts1 turns them into one item per prompt. Generate image1 draws each with gpt-image-1 at 1536x1024 and
   high quality, and Upload file saves it as <module>-<prompt key>.

Split prompts1 looks for a fenced block marked json that holds a list of modules. The prompt's example shows a single
module and asks for neither the fence nor a list. If the step stops with No valid JSON block, add a line to
Image Prompts1 asking for the answer as a JSON array inside such a block.

Change the result

• Course shape: the prompt in Instruction Generator. It asks for H1 module titles and H2 subheadings, and allows up to
  800 words a module when the form gives no count.
• Number of modules: the dropdown in On form submission offers 4, 5 or 6. The example in Structured Output Parser
  lists six modules, so if a run with 4 or 5 comes back with six, trim the example.
• Models: gpt-4.1-mini in OpenAI Chat Model3, and gpt-4o-mini in the other four chat model nodes.
• Research depth: the bodies of Tavily web search and Tavily web search1. The second asks for each page's raw text,
  which makes the module prompts long.
• Pictures: the size and quality in Generate image1, and the choices in On form submission2.

Good to know

• Picture costs: every AI picture is one high-quality gpt-image-1 call at 1536x1024, and six modules at 15 pictures make
  90 of them. Try 5 first.
• Web pictures: the docs link to pictures from other people's sites, and the Drive folder keeps copies. Check the rights
  to each before you publish a course with them.
• Unchecked text: the modules and the script come from a model reading web pages. Check the facts before you teach them.
• Open forms: once the workflow is published, anyone with a form's address can start a run on your OpenAI and Tavily
  keys.
• Costs of a course run: three chat model calls plus one per module, and one Tavily search plus one per module.
• Where the data goes: the topic and the module titles go to Tavily, the research and the texts go to OpenAI, the docs
  go through your converter, and the pictures land in your Google Drive.
• The original copy: two Code nodes had lost their backslashes. Merge the content joined the modules with nn where
  two line breaks were meant, and the pattern in Split prompts1 read s* and [sS] where \s* and [\s\S] were
  meant. This copy has \n\n, \s* and [\s\S] back.