What this automation does
You record or write once. The pipeline turns that single asset into a week or two of channel-ready drafts that an editor approves instead of writes.
- Ingest. A new file in a Google Drive folder, a new podcast episode in an RSS feed, or a newly published article triggers the run.
- Transcribe. Audio and video go through a speech-to-text API with speaker labels and timestamps. Articles skip this step and are cleaned to plain text.
- Extract ideas. A model reads the full transcript and returns a structured list: the five to ten strongest ideas, each with a one-line summary, supporting verbatim quotes, and timestamps.
- Verify quotes. Code checks that every quote appears word for word in the transcript. Paraphrases dressed up as quotes are rejected before anyone sees them.
- Draft per channel. For each idea, the model writes the formats you chose (a LinkedIn post, a short email section, a clip caption, a pull quote for a graphic) using your voice guide and length limits.
- Queue for review. Drafts land as cards in Notion or Airtable, grouped by source asset, with the transcript excerpt beside each one. Approved cards flow to your scheduler or CMS.
When it is worth building
This is worth it when you already produce long-form content regularly, at least two or three pieces a month, and the derivative content is not happening because nobody has the hours. If you publish one article a quarter, an editor with a chat assistant will do the job faster than any pipeline.
It is also worth it when you have a voice that matters. Generic AI-written posts are easy to spot. The pipeline’s value is in forcing every draft through the same voice guide, the same quote checks and the same review step, which keeps quality steady when volume goes up. If your goal is simply more posts with no review, an off-the-shelf repurposing app is cheaper and you should try that first.
Architecture
| Layer | What we use | Why |
|---|---|---|
| Trigger | Google Drive folder watch, RSS feed or CMS webhook | Marketing already drops finished assets in one of these places |
| Transcription | A speech-to-text API with timestamps, such as OpenAI’s transcription models | Timestamps are what make clip suggestions usable |
| Idea extraction | Claude Sonnet 5.5 with a long context window | Reads a full hour of transcript in one pass and returns structured JSON |
| Quote check | Code step with normalized string matching | Catches invented or altered quotes deterministically |
| Drafting | Claude or OpenAI’s GPT models with a voice guide and examples | Per-channel prompts keep length and tone rules separate |
| Review board | Notion or Airtable | Editors see source, draft and status in one place |
| Publishing | Buffer or your scheduler, Webflow for blog excerpts | Only approved items leave the board |
How to build it
- Write the voice guide. One page: who you are writing for, words you use and avoid, how formal, whether you use emoji, and six example posts you were proud of. This document matters more than any prompt tweak.
- Pick three formats to start. For example LinkedIn post, newsletter blurb and clip suggestion. More formats can come later; starting with eight means none of them get tuned.
- Build ingest and transcription. Store the transcript with timestamps in the review board next to the source link, so editors can check context.
- Extract ideas as JSON. Ask for ideas, quotes and timestamps in a strict schema. Reject and retry any response that does not parse.
- Add the quote verifier. Normalize whitespace and punctuation, then require an exact match. Flag near-misses for the editor rather than silently fixing them.
- Draft with one prompt per format. Each prompt gets the idea, its quotes, the voice guide and that format’s rules (character limits, link placement, hashtags or none).
- Wire approval to publishing. When an editor sets a card to Approved, push it to the scheduler with the planned date. Record the post ID back on the card.
The downloadable starter template below covers Drive ingest, transcription, idea extraction and the Notion review board.
What breaks in production
- Transcript errors on names and jargon. Speech-to-text mangles product names and guest names. Keep a custom vocabulary list and run a find-and-replace pass before extraction.
- Quotes stitched from two places. Models sometimes join two sentences said minutes apart into one quote. The exact-match check catches this, which is why it exists.
- Timestamps that drift. If the video was trimmed after transcription, every clip suggestion is off. Transcribe the final edited file, not the raw recording.
- Same idea, every asset. If your speakers repeat a favorite point, the pipeline will keep surfacing it. Store past ideas and ask the model to skip close matches.
- Platform limits. Each network has its own length and link rules, and they change. Keep limits in a config table, not hardcoded in prompts.
Timeline and cost
A pipeline with one source type, three output formats and a Notion or Airtable review board usually takes two to four weeks. Lumeor builds this in the range of $6,000 to $12,000, depending on the number of source types (video, audio, articles), the number of output formats, and whether approved items publish automatically. Model and transcription usage for a few long assets a month is typically a small monthly cost compared with editor time. If you also need the posts scheduled and tracked, pair this with AI social media scheduling, and for the email side see AI newsletter generation.
How to measure success
- Time from a long-form asset going live to its first derivative post
- Number of approved derivatives per source asset
- Share of drafts approved with light edits versus rewritten
- Quotes rejected by the verifier (should be low and stable)
- Engagement on derivative posts compared with posts written from scratch