# How Dheeraj Sharma turns a video transcript into a YouTube thumbnail with a 5-stage n8n pipeline

Dheeraj, a solo creator and host of GenAI Unplugged, used to spend 30-40 minutes on each thumbnail before he could publish a video. He built a 5-stage n8n pipeline. A Google Sheet queues the video, the pipeline fetches and cleans the transcript, and one AI pass writes the title, description and thumbnail concepts. A second AI pass writes the image prompt, and Gemini's Nano Banana Pro generates the thumbnail into Google Drive. He says it goes from transcript to finished thumbnail in under 5 minutes, at $0.15-0.18 per image.

| | |
|---|---|
| **Who** | Dheeraj, solo creator, host of GenAI Unplugged and self-described "AI systems builder" ([genaiunplugged.substack.com](https://genaiunplugged.substack.com)) |
| **Problem** | Thumbnails were the last task blocking publish, and each needed both creative direction and visual execution |
| **Stack** | n8n, Google Sheets, YouTube Transcriptor API on RapidAPI, Gemini 3 Pro (or Claude), Gemini Nano Banana Pro for images, Google Drive. Optional: a YouTube Update node, or a Claude Code skill instead of n8n |
| **Confidence** | Medium (his own numbers from his own newsletter, which sells the workflow) |
| **Last verified** | 2026-10-16 |

## The problem
By the time a video was edited and the description written, the thumbnail still wasn't ready. His old process was to "open ChatGPT or Claude for creative direction, then Canva for execution". In the session he said: "I would spend about 30 to 40 minutes just figuring out what thumbnail is needed. And thumbnail is not ready by the time I have to publish."

He says the launch of Google Gemini Nano Banana Pro is what changed this for him.

## The blueprint (as he describes it)
1. **Image queue.** A Google Sheet is the control panel. Each row holds a YouTube video ID and a status: scheduled, new upload, or thumbnail created. The workflow filters for "new upload" rows. A Limit node makes it process one video at a time.
2. **Transcript fetch.** He uploads the video to YouTube as unlisted (not draft, not published) and waits 30-40 minutes for auto-generated captions. Then he fetches the transcript through the YouTube Transcriptor API on RapidAPI. A Code node runs ASR cleanup. This is a find-and-replace array that fixes misspelled tool names and niche vocabulary.
3. **AI SEO Brain.** The cleaned transcript goes to Gemini 3 Pro (or Claude) with a system prompt covering brand voice, headline rules and thumbnail context format. One pass returns a video title, description, hashtags, and three ranked thumbnail concepts with CTR rationale. An optional parallel branch sends the title, description and tags to a YouTube Update node, which updates the video's metadata before it goes public.
4. **Director prompt.** A second AI stage picks a pose from a library of 12-13 options, such as shocked, pointing or thumbs up. It then writes the image prompt using a 4-layer composition: abstract background, subject at 80% right, floating 3D icons, text overlay.
5. **Image generation and upload.** The avatar reference image and any background references are converted to base64. They are sent to Gemini with the prompt at 1K resolution and 16:9. The returned image is decoded and uploaded to a Google Drive folder. If no image comes back, an if-node loops back for up to three retries.

**Identity lock.** He describes his avatar in detail at the top of the image prompt, puts the composition instructions in the middle, and repeats the avatar description at the bottom. He pairs this with a clean upper-body reference image on a transparent or solid background. "Accuracy started improving when I only provided my image without the destructive elements."

**Debugging.** "I would dump the entire output of all the nodes into Claude, put the image in Claude as well and ask: what do you see? What is wrong? What am I missing?"

**Solo shortcut (raised by Carrie Loranger).** He adds a transcript column to the Sheet and pastes the transcript in, or points the pipeline at a Google Doc. This removes the unlisted-video step, the RapidAPI account, the YouTube transcript API call and the ASR cleanup node.

**Claude Code version.** A `yt-thumbnail` skill holds the brand colors, poses and system prompt. A Python file (`generate_thumbnail.py`) makes the Gemini API call. It runs when he prompts it with a description and a transcript path. Unlike n8n, it doesn't run on a schedule.

He also runs a second version of the pipeline tuned for Substack cover images.

## Stated results
- Transcript to finished thumbnail in under 5 minutes, down from 30-40 minutes per thumbnail.
- $0.15-0.18 per thumbnail at 1K. Gemini's published rate is $0.134 per image, and he says the input tokens push the real bill past that.
- "The pipeline is 90-95% reliable." Regenerating takes two minutes and produces a new concept.
- The identity lock cuts misses "from 3-in-10 down to roughly 1-in-10."
- He says he "no longer needs a designer."
- In the live demo, the Claude Code version was "slightly better than the n8n version for that particular session."
- No click-through, view or subscriber figures are stated.

## Caveats
- The post sells things. It pitches PluggedIn, a paid tier with "the exact n8n workflow template", prompt library and pose system. It also promotes his Claude Code cohort on Maven (with a discount code) and his products SubflowAI and the Content OS Agents Toolkit.
- The article is a write-up of a live conversation with Carrie Loranger (Nine to Thrive). It switches between first person and third person ("Dheeraj describes"), so parts may be written by someone else from the recording. The quotes are from the session.
- The source only gives his first name, "Dheeraj". The surname comes from the brief and is not confirmed by the source.
- The "under 5 minutes" figure starts at the transcript. The default path first waits 30-40 minutes for YouTube captions. The solo shortcut avoids that wait.
- The monthly cost is stated two ways: "under $1.50/month" and "$1.20-$1.44" at 8 videos per month.
- The image model is named inconsistently: Nano Banana Pro, Gemini Image Pro and the Gemini Imagen API.
- X: [UNVERIFIED] not checked.

## Sources
- https://genaiunplugged.substack.com/p/n8n-youtube-thumbnail-ai-automation-workflow (primary)

---
*Want agents like these working your distribution? [See the Agent Stack Teardown](/agent-stack-teardown). Get new teardowns weekly: [subscribe](/subscribe) (double opt-in).*