# The AI Content System Behind 200M Views

Sarah Chieng, Cerebras & Cova Lee, OpenAI | AI Engineer NYC 2026 | 53:18

Source: https://www.youtube.com/watch?v=CFbnVpSsDFQ
Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks.
Page: https://aietalks.com/talks/the-ai-content-system-behind-200m-views
Published: 2026-10-21
Tags: agents, computer-use, video, workflows

## TL;DR
- Sarah Chieng and Cova Lee organize content work into six stages: ideating, scripting, recording, editing, repurposing, and publishing.
- Agents can handle repetitive production work such as research, shot lists, editing, captions, scheduling, and process reviews, while the creator keeps control of the ideas and voice.
- AI should remove tedious work so creators can spend more time developing taste, pursuing real experiences, and treating content as a craft rather than producing generic slop.

## Summary
Sarah Chieng and Cova Lee walk through the content system they use across ideation, scripting, recording, editing, repurposing, and publishing. They show how a swipe file gives agents examples of a creator's taste, while custom research tools can compare titles, thumbnails, and performance outliers. Voice input and interviews help turn spoken explanations into clearer scripts, and script skills preserve a creator's style. The production system then generates shot lists, organizes footage, applies design systems, edits through APIs or computer use, creates visual effects, cuts long videos into short clips, writes captions, and schedules posts. A weekly review finds repetitive tasks and process waste. The speakers are clear about the limit: AI can execute and iterate, but it cannot supply the creator's point of view. Their advice is to stay involved in the ideas, references, and final choices, while delegating work that consumes time without adding much creative value.

## Key ideas
### A six-stage pipeline makes content production easier to automate
[03:59](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=239s)
Chieng and Lee structure their work around ideating, scripting, recording, editing, repurposing, and publishing. Ideation covers concepts and audiences. Scripting covers hooks and testing different versions. Recording turns the script into footage. Repurposing cuts longer work into related videos, while publishing handles captions and distribution. This structure gives each agent task a clear place in the process. It also makes it easier to identify which parts should remain under the creator's control and which parts can be delegated.

### A swipe file gives an agent concrete evidence of a creator's taste
[05:31](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=331s)
Lee describes a swipe file as a single place for favorite references, questions, and sources of inspiration. Hers is a colorful website containing videos and saved Pinterest images, while a simpler version can be a Notion page containing links from Twitter. The file helps an agent understand the visual and creative style the creator wants. Lee says references communicate impact more effectively than text alone. Chieng adds that titles and title images deserve substantial attention because they determine whether people watch or read, and can affect whether a post gets 500 views or 5 million.

### Custom research tools let creators test packaging before making the video
[06:15](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=375s)
Lee built a YouTube research tool that pulls videos from favorite creators into one place, allowing her to study titles and thumbnails and remix thumbnail ideas. She then built a simulator showing how a proposed package might look on the YouTube homepage. The point is to see whether an idea stands out before spending time making the full video. The speakers also recommend comparing performance with a creator's own normal results and asking agents to find outliers, rather than copying whichever topic appears popular.

### Speaking ideas aloud produces clearer scripts than starting with polished copy
[09:29](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=569s)
Chieng prefers to explain a technical idea out loud, then give that explanation to ChatGPT or use voice mode to interview her. The agent can ask for unclear parts, expose weak connections, and suggest cuts. She summarizes the writing goal as being concise and clear. Lee recommends turning existing work into a script skill that captures how the creator wants to sound, then generating three script options. Those options can be ranked against goals such as engagement or ease of cutting.

### Shot lists and organized assets turn a script into a practical recording plan
[12:50](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=770s)
Lee creates two passes over a script. The first labels sections as A-roll, B-roll, screen recordings, or graphics. The second produces a checklist for filming, so the creator can collect each required shot without repeatedly searching through the script. Coast can search locally recorded computer activity for earlier demos, while CleanShot helps record, annotate, blur, outline, and label screenshots. After recording, the speakers recommend naming and organizing footage by type so an agent or editor can attach the right media quickly.

### A design system lets agents reproduce a creator's visual style
[17:01](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=1021s)
The speakers use design systems to give agents reusable visual rules. Chieng describes generating several branding and logo versions, then selecting the preferred direction. Lee gave Codex her 50 most recent Instagram videos and asked it to turn recurring visual patterns into a Figma document. Swipe-file references can guide this work too. The purpose is to make the creator's visual choices explicit enough that an agent can apply them consistently while the creator still chooses which version feels right.

### Editing can be automated at different levels of control
[18:43](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=1123s)
Chieng describes using video APIs such as HyperFrames to stitch angles, add text, render, review, and apply brand guidelines from Figma. HeyGen can create a short avatar replacement when a creator says one word or number incorrectly, though neither speaker recommends making an entire video an avatar. For more detailed work, computer use can operate Premiere, Final Cut, or other editors. Lee also builds small editing tools, such as a sound pad that inserts selected effects into a specific track. The choice depends on how much control the creator wants over clips, layers, and timing.

### Repurposing and scheduled publishing turn one long video into a content queue
[26:18](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=1578s)
A long interview can be analyzed for strong sound bites and cut into Shorts, TikToks, and Instagram Reels. Captions should reflect the creator's voice, platform, format, and brand rather than using one generic template. Chieng uses a list of signs of AI writing as a final check and asks the system to avoid those patterns. For distribution, Buffer can receive a folder of cutdowns, schedule them across platforms, research suitable posting times, and draft captions for approval. The result is a publishing queue that can run for several weeks.

### A weekly process review finds automation opportunities the creator misses
[29:26](https://www.youtube.com/watch?v=CFbnVpSsDFQ&t=1766s)
The speakers run a recurring skill that reviews their content process, identifies where time goes, and suggests repetitive tasks to turn into tools or automations. It found that Chieng was maintaining two places for morning triage and sending reminders to a Slack thread she never checked. It also found a Premiere check that ran every time even though it was no longer needed. Removing those steps made the process faster. The same review can examine better-performing posts and update the creator's skills over time.

## Notable quotes
- Cova Lee: "At the end of the day these tools are just to help you automate away the things that are very tedious so that you can actually treat content creation like a craft." (32:39)
- Cova Lee: "But at the end of the day this is your craft and this is your voice and you always have to be in the driver's seat otherwise nobody is going to watch your content." (33:13)
- Cova Lee: "The one goal of it is to get closer to the thing we want to say and our voice." (33:28)
- Cova Lee: "I really deeply believe the reason Sarah and I have succeeded in content creation is in part because we have had really strong POVs that we've cultivated over real experiences." (43:55)

## Tools & references mentioned
- Cerebras
- OpenAI
- Mistral
- Cognition
- ChatGPT
- Codex
- Swyx
- Notion
- Obsidian
- Wispr Flow
- Dots
- Coast
- CleanShot
- Screen Studio
- DaVinci Resolve
- Adobe Premiere
- CapCut
- Final Cut Pro
- Figma
- HyperFrames
- HeyGen
- Buffer
- After Effects
- Blender
- Sony FX3
- Canon EOS R50
- Sony ZV-1
- Matt Palmer
- Replit
- SpaceAct
- Tarik
- Jason

## Who should watch
- You publish technical or educational videos and want a repeatable system instead of rebuilding the workflow for every post.
- Your content process contains repetitive research, editing, captioning, or scheduling work that could be handed to an agent while you approve the creative choices.
- You want to use AI without losing your own voice, visual style, and point of view.

## Related talks

- [Building an Agentic Video Editor for Mass Consumer](https://aietalks.com/talks/building-an-agentic-video-editor-for-mass-consumer) (Ekaterina Deyneka, Reelful, 12:45)
- [Content Is Code](https://aietalks.com/talks/content-is-code) (Matt Palmer, Conductor, 10:53)
- [Agents for Everything Else](https://aietalks.com/talks/agents-for-everything-else) (swyx, AI Engineer, 14:10)
- [The Robots Are Coming for Your Job, and That's Okay](https://aietalks.com/talks/the-robots-are-coming-for-your-job-and-thats-okay) (Elmer Thomas & Maria Bermudez, Twilio, 08:57)
- [Lifestyles of the AI-Native](https://aietalks.com/talks/lifestyles-of-the-ai-native) (Nick Nisi & Zack Proser, WorkOS, 1:01:28)
