How to auto-write social posts from YouTube videos using AI

Executive overview

Most people either avoid AI for writing or accept mediocre output. The gap is in the system: what you feed the model matters more than the model itself.

Convert a YouTube video or podcast to a transcript, pair it with a curated style guide from someone who writes well on your target platform, and use that to generate a system prompt. The AI then writes posts that match your voice and platform conventions.

The key insight: AI doesn't replace taste — it scales it.

Building the content pipeline

• Consume content daily across podcasts, YouTube, newsletters, discords, and X
• Aggregate interesting finds in Apple Notes with a link and a 3–5 word memory prompt
• Batch-write 8–10 posts on Sundays from the week's curated ideas
• ~60–65% of posts derive from audio or video, converted to text first
• A short Python script (~35 lines) pulls transcripts from YouTube links or podcast files

Finding your style reference

• Identify someone who writes well on your target platform and is known in that space
• Many will share format guides, PDFs, or interviews explaining their approach
• Convert that material to a knowledge base document the AI can reference
• This is a one-time setup step, not something you repeat each session

Creating the system prompt

• Use a strong reasoning model (o3, Claude 3.7 thinking, Gemini 2.5 Pro) to write the system prompt from your style reference
• Prompt it to search for current prompting best practices — be explicit, since uploading a PDF can disable search
• The generated prompt should include: persona, format rules, input/output templates, self-evaluation step
• Remove chain-of-thought instructions — reasoning models don't need them
• Add manual refinements: fifth-grade reading level, favour simplicity, avoid cliché phrases like "game changer"

Iterating the prompt over time

• After 5–10 post cycles, patterns emerge in what the AI does well and poorly
• Make small, targeted edits to the system prompt rather than rewriting it
• Specific additions that improve output: binary CTAs (not open-ended questions), explicit ban on filler words, instruction to keep hooks authentic to the content
• AB test Claude and GPT projects side by side to see which performs better per use case

Writing a post: live example

• Drop the transcript into the project alongside any explicit framing (e.g. "the primary idea is X")
• Add a hook instruction: "scroll stopper but authentic to the content, not overly dramatic"
• Request credit attribution in parentheses at the end, preserving post flow
• Mention notable brands or names early if relevant — they attract attention
• Expect to iterate once: the first draft is often too short or too long; one follow-up prompt is usually enough
• Paste the final output into a doc, make minor inline edits, strip any remaining AI clichés
• Optionally, use the system prompt to generate an image prompt for a complementary visual