Content Repurposing Automation

10 to 15 minutes per run. Out comes 4–5 social cards with 7–10 slides each, a whitepaper as PDF, YouTube metadata, and a thumbnail. By hand this was hours of post-production, when it happened at all, given the team's workload.

The problem

Expert podcasts produce genuinely useful content. As raw audio or video they reach a fraction of the audience they could. A 30-minute episode contains enough material for a slide deck, several social posts, and a polished YouTube upload with full metadata.

That repurposing usually doesn't happen. Transcribe the recording, write a YouTube title and description, set chapter markers, design a thumbnail, restructure content into slides, build social cards, distribute to the right people. Hours of post per episode. In content teams the work piles up. Not from lack of will, from lack of time.

The approach

The workflow fires on one action: drop an MP3 into a specific Google Drive folder. From there everything runs automatically. A Google Drive trigger picks up the new file and starts two parallel production tracks: one for YouTube assets, one for Gamma deliverables.

First step on both tracks: Gemini transcribes the audio. Everything else feeds off that transcript. One input, many outputs.

The YouTube assets

The YouTube branch produces three things: SEO metadata, chapter markers, a thumbnail. Gemini turns the transcript into a JSON output with title, description, tags, and chapter timestamps. A downstream code node parses it, assembles the final YouTube description with timestamps, and prepares the upload package for YouTube Studio.

The thumbnail generation

The thumbnail comes from an AI agent acting like an art director: it knows the corporate-identity rules (palette, type, layout constraints) and writes the image prompt. Its output is a JSON with the image prompt, the hook text for the thumbnail, and a short rationale for the creative call.

That prompt then goes to Gemini (via OpenRouter), which renders the image. The base64 result gets converted to PNG and attached to the YouTube assets email.

Image generation ate the most time. With NanoBanana it stayed off for a long stretch. You really have to find the right prompt to get consistent results. Small wording changes produced completely different styles. It took plenty of iteration.

The content distribution

Parallel to the YouTube track, the Gamma branch runs. It splits into two paths: decks from the full transcript, and social cards from extracted single topics.

Decks from the full transcript

The complete transcript goes straight to the Gamma API as a 16:9 deck, exported as PDF. The workflow generates at least two variants with different brand themes and saves them into the matching brand folders. Finished PDFs work as whitepapers or lead magnets.

What surprised me: how well Gamma handles the input. The information gets pulled cleanly out of the transcript, and the slides are genuinely usable.

Social cards from extracted topics

The second path was the more interesting one. And the harder one. Gemini Flash extracts the individual topics and key points from the transcript as a JSON array. Each entry holds a topic and its supporting points. The workflow splits the array and generates a Gamma social card in 4:5 for each topic.

The tricky part was the split itself. Each talking point has to stand on its own and offer value, or the social card is worthless. The extraction prompt needed a few rounds before the results actually held up.

A typical 30-minute episode yields 4–5 social cards at 7–10 slides each. Carousel content that goes straight to LinkedIn or Instagram.

The delivery

The workflow routes results to different recipients. YouTube assets (title, description, tags, thumbnail) go out by email. Gamma deliverables (deck PDFs and social-card links) go to their own recipients, with an HTML list of URLs and download links.

Splitting delivery was deliberate. If you work with YouTube assets, you get only those. If you need decks, you get only those. Each recipient gets what's relevant for the next step.

The result

Drop an MP3 in Google Drive, wait 10–15 minutes. What comes out:

• Transcript of the audio
• YouTube metadata: title, description with chapter markers, tags
• On-brand thumbnail as PNG
• Decks as PDF in multiple brand variants
• 4–5 social-media carousel cards per podcast topic, 7–10 slides each
• All delivered by email to the right teams

What I'd do differently today: more time on the input. Clearer guidance on who the output is for and which channel it's bound for. Audience, intent, platform. The more precise the input, the better the output. Sounds obvious. On the first build I over-indexed on the output workflow and under-invested in input quality.