{ "video": { "id": "CFbnVpSsDFQ", "title": "The AI Content System Behind 200M Views — Sarah Chieng, Cerebras & Cova Lee, OpenAI", "duration": 3198, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "" }, "analysis": { "video_id": "CFbnVpSsDFQ", "title": "The AI Content System Behind 200M Views — Sarah Chieng, Cerebras & Cova Lee, OpenAI", "one_liner": "Use AI to handle the tedious parts of content creation across six stages: ideate, script, record, edit, repurpose and publish. Feed agents your own references through a swipe file, and keep creative control yourself, or the output becomes slop nobody watches.", "summary": "Sarah Chieng (head of developer experience at Cerebras) and Kova Lee (OpenAI developer experience, formerly a full-time short-form creator) walk through the exact prompts, skills and tools they use at each of six stages: ideating, scripting, recording, editing, repurposing and publishing. Most of it runs on ChatGPT/Codex with the Cerebras-powered 'GPT 5.6 Soul ultra fast' model, which runs at 750 tokens per second. The core ideas are to give agents your taste through a swipe file, build custom tools and skills (a YouTube research tool, a homepage simulator, a sound pad, a design system built from your own videos), let agents drive editors and publishing APIs, and review your process and post performance regularly. They stress that the tools exist to automate tedium so you can treat content as a craft and stay in the driver's seat.", "key_points": [ "They follow six stages, ideating, scripting, recording, editing, repurposing and publishing, using ChatGPT/Codex with the Cerebras-powered 'GPT 5.6 Soul ultra fast' model at 750 tokens/sec. They say the speed matters for large scripts and files, and makes it practical to generate around 10 versions of branding or logos and cherry-pick the best.", "Swipe file: one place for every reference that inspires you (Notion, Obsidian, a website, or simply pasted Twitter links), following Swyx's swipe file strategy, so the agent can learn your taste. Kova: 'Text can very rarely get to the impact a reference has.'", "Packaging is underrated. Kova says 50% of a YouTube video's performance comes from title and thumbnail; Sarah says creators should spend 50% of their time on titles and header images, which can mean 500 views versus 5 million. Kova built a YouTube research tool that pulls thousands of videos from favourite creators, remixes thumbnails and has an outlier slider (1x–20x a channel's average views), plus a simulator of the YouTube homepage to test whether an idea stands out before making it.", "Scripting: dictate concepts (Whisper Flow) or have ChatGPT/Codex voice mode interview you, because written drafts drift toward marketing-speak. Judge engagement relative to baselines (your own past posts, or posts on the same topic). Capture your voice as a script skill, generate three script options, and optionally score them (e.g. 50% engagement). The 'two C's' are concise and clarity.", "Recording and shot planning: a two-pass shot list. The first pass annotates the script with A-roll, B-roll, screen recording and graphics; the second produces a filming checklist. They also use Coast (local, always-on screen recording searchable via the Codex CLI) to find old demos to re-record, CleanShot for fast annotated captures and blurs, and have ChatGPT rename and organise footage.", "Editing: Hyperframes by HeyGen via its API to stitch angles, add captions and text animations from raw footage plus reference videos. Sometimes the output is a one-shot final cut, but changes mean re-prompting and re-rendering. HeyGen avatars patch about 0.5-second flubs (e.g. a misspoken tokens-per-second figure). Astra computer use handles Premiere; Final Cut has a good MCP; Sarah has ChatGPT drive CapCut desktop to fix captions and add sound effects by rule (a click when an image appears, a whoosh on transitions). Kova one-shotted a nine-button sound pad, built a Figma design system from her last 50 Instagram videos, and had Astra generate Blender GLB models and composite floating objects in After Effects.", "Repurposing and publishing: a long interview takes Sarah about a month, so she clips it into shorts, writes captions with per-platform and per-voice skills, and has the AI avoid everything in Wikipedia's 'signs of AI writing' article. Buffer is set up through its API from the docs to schedule one cutdown a week across platforms at researched best times, with captions sent for approval first.", "Weekly process-review skill ('show me where I spend the most time and suggest repetitive tasks I could turn into tools or automate') plus a monthly review of post performance. It found reminders going to a Slack thread Kova never checked, and a slow, unnecessary import check in the Premiere MCP workflow." ], "takeaways": [ "Build a swipe file (even a Notion page of links) and point your agent at it for scripts, design and captions: references convey taste better than text.", "Spend far more time on titles and thumbnails. Build small custom tools to study outliers and simulate how your packaging looks on the YouTube homepage before you make the video.", "Turn your voice, platforms and editor workflows into skills: a script skill, per-platform caption skills, and a Premiere skill with sub-skills for each step. Hand agents API docs (Hyperframes, Buffer, the Twitter API) so they set up the integrations.", "Run a recurring 'review my process' skill and a monthly performance review, and update your skills from the results. When testing captions or hooks, compare cutdowns of the same long video so the comparison is apples to apples.", "Automate the tedium but write the final product yourself and keep your point of view: live in the real world, draw on hands-on engineering work for authentic content, and stay in the driver's seat." ], "topics": [ "content creation", "video editing", "agents", "skills", "computer use", "automation", "social media", "creator workflows" ], "tools": [ "ChatGPT", "Codex", "GPT 5.6 Soul ultra fast", "Cerebras", "OpenAI", "Whisper Flow", "Notion", "Obsidian", "Figma", "Coast", "CleanShot", "Screen Studio", "Adobe Premiere", "DaVinci Resolve", "Final Cut Pro", "CapCut", "Astra", "Hyperframes", "HeyGen", "Image Gen", "Photoshop generative fill", "After Effects", "Blender", "Buffer", "Dots", "ChatGPT Space", "Twitter API", "Slack", "Sony FX3", "Canon EOS R50", "Peak Design tripod", "Photodiox Gobo", "Rode", "DJI mic", "Sony ZV-1" ], "quotes": [ { "text": "when really you should be spending 50% of your time on it.", "at": "08:26", "url": "https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=506s" }, { "text": "Text can very rarely get to the impact a reference has.", "at": "09:29", "url": "https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=569s" }, { "text": "you always have to be in the driver seat otherwise nobody is going to watch your content.", "at": "33:22", "url": "https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=2002s" }, { "text": "If everything is aligned, it will go viral.", "at": "46:37", "url": "https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=2797s" } ], "words": 12157 }, "summary_url": "/#CFbnVpSsDFQ", "transcript": { "html": "/transcripts/CFbnVpSsDFQ.html", "txt": "/transcripts/CFbnVpSsDFQ.txt", "vtt": "/transcripts/CFbnVpSsDFQ.vtt" } }