The AI Content System Behind 200M Views — Sarah Chieng, Cerebras & Cova Lee, OpenAI
Loading slides…
Read the talk
How two video creators turn a script into a shoot, an edit, and six weeks of posts
Sarah Chieng makes technical ideas understandable on camera. Cova Lee builds cinematic short videos. Their workshop follows six stages of production and shows why the two creators hand quite different jobs to their agents.
From a talk by Sarah Chieng and Cova Lee
At a glance
Ideas worth remembering
Compare a video with that channel’s normal performance, then test the packaging among the thumbnails a viewer would actually see.
Annotate the script with pictures first and extract a shooting checklist second. The writer and the person holding the camera need different views of the same plan.
Use a rendered first cut when feedback can be expressed as a new instruction. Keep a layered timeline when the timing of individual clips and sounds is the creative decision.
Review the production process itself. A reminder nobody reads or a check nobody needs can be removed instead of automated again.
Two creators want different things from an editor
Give Sarah Chieng and Cova Lee the same footage and they will spend their editing time differently. Chieng, who leads developer experience at Cerebras, makes technical explainers. For her, an unclear line in the script matters more than a beautiful transition. She is happy to show an agent which caption styles and sound effects she likes and let it apply them. Lee, a former full-time short-form creator who joined developer experience at OpenAI, cares intensely about how music, light, and movement fit together. She would happily automate getting a sound onto the right track. She still wants to choose exactly when you hear it.
They organize the workshop around six stages familiar to anyone who has tried making a second video after finishing a first one: ideation, scripting, recording, editing, repurposing, and publishing. Each stage leaves something the next one can use. The idea becomes a script. The script describes the pictures someone needs to shoot. Named files let an editor find those pictures later. The workflow starts well before anyone opens an editing application.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Test the thumbnail next to the videos it will compete with
Both creators keep a swipe file of things they would like to learn from. Lee’s is a colorful website that brings videos and saved pictures together. Chieng’s can be as simple as dropping links into a Notion page. Someone trying to describe their visual taste to an assistant can spend a long time searching for an adjective. A few actual examples give both parties something to point at.
When Lee wanted to learn YouTube, she built a research tool that pulled thousands of videos from creators she followed. It gave her one place to compare titles and thumbnails and try variations on the combinations that caught her attention. Then she built a simulated YouTube homepage. An arresting thumbnail on an otherwise empty canvas might disappear beside a dozen others. The simulated feed let her change a thumbnail and see whether she would still notice it.
Raw views make a poor filter for this exercise. A famous account can draw a crowd with an ordinary post. Lee’s tool looks for outliers relative to a channel’s own average, with a slider that runs from one to twenty times that baseline. An unusually successful video from a modest channel can teach her more about an opening or a piece of packaging than a bigger number from a celebrity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Let an interviewer find the hole in the script
Chieng starts a technical script by talking through the idea and why it matters. Dictation captures how she would explain it to someone. She can then ask an assistant to interview her. If two points seem unrelated, the follow-up question forces her to supply the missing connection or cut one of them. That is useful editorial work even if she keeps none of the assistant’s wording.
Past scripts become examples for a reusable writing skill. Lee likes to generate three versions so she can compare the available openings and structures. She can ask the model to rank them against a goal, but laying the alternatives beside one another is valuable in its own right. A line that sounds plausible in isolation can sound unlike the speaker once it sits beside the version they would actually say aloud.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turn the script into a two-pass shot list
Lee’s first pass annotates the script. Beside each line it identifies what the viewer will see: the creator speaking, some B-roll, a screen recording, or a graphic. That exposes problems while they are cheap. A voiceover promising to show something needs a picture that actually demonstrates it. Her second pass collects the required shots into a checklist. She can read the script in narrative order at her desk, then use the checklist to make sure she has everything before she puts the camera away.
Software demos create a particular retrieval problem. You may have seen the program do exactly the right thing earlier and have no idea which screen recording captured it. Lee describes using a local recording tool to find those moments as references and then recording the demo again. Chieng likes being able to annotate a screenshot or blur a detail in a screen capture immediately. Neither job needs to wait until the final edit.
Once the shoot is over, naming and organizing the assets gives an agent a fighting chance of assembling them. A filename such as “screen recording 09436” tells an editor very little. Separating the main take, extra footage, and screen captures lets the next instruction refer to a shot by its job in the story. It also helps a human spot a missing demo before the editor has built a sequence around it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Make the visual style available as an artifact
Chieng describes generating several treatments for branding, logos, and the overall feel of a video, then picking the one she likes. Lee approached the same problem from the other direction. She already had a recognizable look. She asked her assistant to study her fifty most recent Instagram videos and turn their recurring visual style into a Figma document. A new project could start with the aesthetic she had spent time developing.
The useful handoff here is a document the editor can consult again. When the creators ask for a new graphic or a treatment for text, the assistant has examples of what belongs together. They can spend their time deciding whether the graphic explains this particular moment instead of specifying the look from scratch on every cut.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Choose the editing interface by the change you expect to make
Chieng describes giving an assistant the documentation for HyperFrames, three camera angles, reference clips, and a request to stitch the footage together with text. An API makes it possible to render a first cut, inspect it, and send back feedback. Lee points out a practical limit of working with a rendered file. If moving one title means prompting again and rendering the whole result, a conventional timeline with separately editable clips, titles, and music can be a better place to finish.
Lee’s nine-button sound pad is a small example of why she builds her own editing tools. Each button inserts a familiar effect, such as a click, whoosh, or shutter, onto its designated track. The shortcut saves her the hunt through a sound library. Her hand still controls which effect lands on which frame. In Premiere she groups her recurring processes for color, sound, and captions into skills. The assistant can recognize that the sound effects have been added while the captions still need a pass.
Generative editing can be similarly local. Lee shows how she masked part of a background and added set dressing while keeping herself intact in the shot. Chieng describes using an avatar for a very short correction when a word in an otherwise good take is wrong. Both examples begin with a specific change the creator already wants. They have a concrete way to tell whether the edit worked.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Six cutdowns need six openings
A long recording gives the creators material for short clips, but an excerpt loses the setup that came before it. Each cutdown needs a point someone can follow on its own. The speakers use references when choosing caption formats and distinguish their personal voice from their company’s voice. A caption that works on one account, or under a long interview, may feel out of place on an Instagram short.
Chieng illustrates the publishing stage with a folder of six cutdowns. She can ask her assistant to use Buffer to schedule one every Tuesday for six weeks, tailor the posts to several channels, and propose captions for her to approve. The calendar instruction is easy to repeat. Which moment deserves six weeks of someone’s attention, and what the first sentence needs to explain to each audience, remain editorial decisions.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Ask the workflow to show where it wasted your time
Lee describes running a weekly review of her own production process. She asks where the time went and which repeated tasks would justify a tool. One finding was almost embarrassingly mundane. Reminders were still arriving in a Slack thread she no longer checked, although she had moved her morning triage elsewhere. More reminders would only have made the unused thread longer.
Another finding came from Premiere. Her setup repeatedly checked whether footage was importing correctly each time she opened the editor. The check cost time and she no longer needed it. Removing it made the workflow faster. This gives the weekly review a useful scope: look for the actual step someone repeats, ask what it still protects, and change that step. It is as capable of eliminating an automation as it is of suggesting a new one.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What the audience asked about taste
During the Q&A, the two creators describe very different levels of organization. Chieng keeps a history of scripts and a swipe file in Notion, where an assistant can retrieve them through a connector. Lee maintains a more detailed set of editing instructions. They see value in sharing the structure of a skill, but copying the voice it encodes would miss the point. A new creator needs examples of their own decisions before a skill can reliably reproduce them.
An audience question raises the danger in benchmarking every idea against past hits. Looking only at what already performs means following someone else’s experiment. The outlier filter can identify a thumbnail worth studying, but it cannot supply a creator’s point of view. The speakers give an example from outside the usual developer feed: a conversation about fashion with a sister prompted a creator to build an AI wardrobe from photographs of clothes. That idea depended on paying attention to something away from the screen.
Asked where their own voices come through, Lee talks about the fit between music and visuals. Chieng comes back to the script of a technical tutorial. The same automation can therefore be a relief for one creator and an intrusion for another. A good first project is the repetitive task whose outcome you already know how to judge, such as fixing captions or inserting a familiar sound. The time it returns is yours to spend on the part that makes someone want to watch.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Full transcript
Loading transcript…