n8n is a tool that connects services in a visual workflow: a row in a spreadsheet goes in, and a script, images, a voiceover and a rendered video come out. Automate the mechanical steps (file handling, calling APIs, timing, rendering) and keep the decisions (the topic, the script, the final check) with a person. A workflow that publishes the same template many times a day produces exactly what YouTube refuses to pay for.
This lesson is technical. You can run a successful channel without it; come back when a manual process works and you want to speed it up.
A note on age: Adrian has since written under his video that this build is outdated and that he now runs cheaper workflows. We keep his clips for the design (narrow agents, measuring the audio before generating video), not for the exact nodes or prices.
Plays from YouTube when you press play. Watch on YouTube
What to automate, and what to keep human
Automate what is the same every time: fetching a topic, sending prompts, saving files under fixed names, joining audio and pictures, logging the result.
Keep human what makes a video yours: choosing the topic, approving the script, checking the facts and watching the result before it goes out.
Nick Winter's cost breakdown shows the temptation. His videos cost under 30 cents each because six images loop under a long voiceover. That saving is also what makes a video look mass-produced. Nick Winter 1:30
A note on the sources: the channel Nick Winter demonstrates reworks stories from other channels, and Productive Dude's demo is a generic facts Short made from one line of text. We use their clips for the engineering only. With your own stories the same workflows produce your own videos; with other people's they produce reused content.
Anatomy of an n8n video workflow
Draw the steps before you build a single node. A typical chain for a short video has nine:
- Fetch a pending topic from a spreadsheet.
- Write the script and captions.
- Write image prompts.
- Generate the images.
- Turn the images into clips.
- Generate the voiceover.
- Combine everything.
- Render the final video from a template.
- Publish and archive.
Productive Dude builds a similar chain click by click and starts it from one typed line: script, voiceover with timing, timed chunks, one image per chunk, clips, render. Productive Dude 0:00
A second design splits the thinking over three AI agents, each with one narrow job: a strategist, a scriptwriter and a video director. The voiceover is generated first and drives the edit. Adrian / AI Agents 1:25
A spreadsheet as control panel, one topic per run
Give each topic a row with the fields the prompts need (topic, tone, niche, language) and a status column. The trigger reads the rows where the status is "pending" and returns only the first one, so each run makes one video. Pin the fetched row while you test, so the workflow reuses the same data instead of fetching again. GenAI Unplugged 3:00
This row is also where your approval lives: a topic only becomes "pending" when you set it.
Script and captions with an LLM node and a fixed output
Fill the prompt from the spreadsheet fields and put the role and style in the system prompt. Then force the answer into a fixed structure with an output parser and a JSON example: hook, captions, call to action, full script. A fixed structure is what lets the next nodes rely on the output. GenAI Unplugged 6:20
The same idea in its simplest form: describe the exact JSON in the system prompt (intro, middle, call to action, title, description), switch on JSON output, and join the parts into one script field with a set node. Productive Dude 5:25
Split that single result into separate items so each caption gets its own image prompt, and send the long script down another branch to the voiceover. Pin the outputs before any node that costs money. GenAI Unplugged 12:45
For long videos, write in steps inside the workflow: a three-act outline first, then one prompt per chapter with the whole outline attached, so each chapter knows what comes later. Models repeat themselves, so a code node can vary names. Nick Winter 25:40
Specialised agents and a word limit per scene
Put a description of your channel's style in the system prompt of every agent. Then match the script to the limits of the video model. If a clip lasts at most eight seconds, the scriptwriter gets a fixed number of words per scene (16 in this example), and the number of scenes sets the length of the video. A limit node keeps test runs to two scenes. Adrian / AI Agents 3:40
The video director agent writes one visual prompt per scene and avoids words that the video model's safety filter refuses. Keep that list and extend it after each failure. Adrian / AI Agents 9:25
A prompt-writing node works best with context. Productive Dude gives it the whole script, the current chunk and its position, a list of what the video model animates badly, and the keyword that the image style needs. Productive Dude 20:00
Images and video through an API
Image and video models are called with an HTTP request node:
- Store the API key as a credential, never inside the node or in a shared workflow file.
- Clean the prompt of quotes and special characters before it goes into JSON.
- Wait. A request creates a task that finishes minutes later. Add a wait step, or poll the job status every twenty seconds until it is done.
- Save each finished clip at once, so a later failure does not lose clips you paid for.
GenAI Unplugged 21:10 · Adrian / AI Agents 9:25
One look for the whole video. Productive Dude uses one style add-on for the image model, with its trigger keyword in every prompt. When the first images missed the look he wanted, he fixed it in the prompt-writing node with example prompts and a list of style preferences, not by hand per image. Productive Dude 24:00
Fetch each result by its ID in a separate request, and give video generation a much longer wait than images. Productive Dude 33:15
For stills, a sub-workflow can write the prompts, save each image under a numbered scene name and let the renderer add a slow zoom, while the voiceover is generated in parallel. Nick Winter 34:40
Long voiceovers in chunks, with checks
Check the length before you generate video. Each scene's audio is measured. If it is longer than the clip, the workflow retries once at a higher speed, and if it is still too long it stops with a clear error so you can lower the word count. Adrian / AI Agents 6:25
Ask for timestamps. Request the voiceover through a plain HTTP node and use the variant that returns timing per character. A short piece of code can then cut the script into chunks that match the clip length of your video model, and a split-out node turns them into one item each. Productive Dude 9:20 · Productive Dude 13:55
Split long scripts. A long text sent to a voice service in one go produces glitches. A code node splits the script into numbered chunks, a small table tracks which workers are free, and the numbered chunks are joined into one file when all are done. Nick Winter 37:25
Status, queue and restart; costs and self-hosting
One video at a time. An organiser workflow reads the table, stops if any row is "processing", and otherwise takes the first "queued" row. All values that differ per channel sit in one settings node at the start. Nick Winter 19:30
Assembly. Code pairs each audio file with its clip and calculates the timings from the audio durations. The finished video goes to the project folder, and the title, description and link are logged in one sheet. Adrian / AI Agents 12:45
The renderer needs links, not files: upload the voiceover to a media host, collect the clips into one list and merge both into a single item. A code step then writes the description of the video that the rendering service expects. Post it, wait, and fetch the result. Productive Dude 38:20 · Productive Dude 42:10
Captions and restart. Timed captions come from a transcription of the finished voiceover; check the offset against the audio. The run ends by setting the status to done and starting the next queued row. Nick Winter 45:30
Costs. With generated video clips, the clips are the main cost: about two dollars per short video in Adrian's build, which he now calls outdated and more expensive than his later ones. With still images, the voiceover is. Adrian / AI Agents 1:25
Self-hosting. An automated channel needs storage, a tracking table and a renderer. Hosted services run into request limits or get expensive at volume, and a rented server of $10 to $20 a month can run open-source versions of all three plus n8n. That takes server knowledge; start on hosted services. Nick Winter 6:40
A useful side workflow. YouTube's free channel feed lists a channel's latest fifteen uploads. A workflow can read it for a list of competitors every twelve hours and keep the videos with many views for their age. That is an outlier monitor for your research. Nick Winter 10:10
Checklist
Your ticks are saved in this browser only.
Income and cost figures mentioned in these videos (for example "under 30 cents per video" or "$6,150 last month") are the creators' own claims and are not verified.

