Original video in Chinese.
Key Takeaways
- Record & Replay lets you demonstrate a computer workflow once. Codex interprets the actions and turns them into a reusable skill, reducing the need to describe a complex workflow in words.
- I tested copying an article from Ulysses into Substack and adding a subscription button. The first replay missed the button; after feedback, Codex updated the skill and reproduced that step successfully.
- Video understanding and computer use work together to turn tacit knowledge into explicit instructions. For me, this is one of the year’s most significant upgrades to agent capabilities.
Yesterday, Codex introduced Record & Replay. It is the most impressive feature I have seen this year.
Once enabled, it records the computer actions that follow. Codex interprets each step and turns the workflow into a skill. Next time, you can ask Codex to invoke that skill.
I rarely use extravagant language, but this feels like a major breakthrough: you can teach AI by demonstrating an action instead of describing it.
Previously, creating a skill required putting your workflow into words and giving it structure. Many people find that difficult. With Record & Replay, you perform the workflow in front of AI, and it learns from the demonstration.
In essence, it makes tacit knowledge explicit.
Hello everyone, and welcome back. I aim to explain both the why and the how of AI, with more than a basic tutorial. Follow the channel, or connect with me in newtype. At the time of this episode, the community had operated for 800 days and more than 2,400 people had joined as paying members.
Users in China could join through Zsxq; overseas users could join through Substack. The community included my two courses, daily newsletters, and exclusive videos.
Record the workflow
I had shared this demonstration in the community the day before. The successful test impressed me enough to post the recording immediately.
Several times a week, I copy a completed article out of my writing app, put it into Substack, format it, and prepare it for publication. Automating that repeated work would save time.
I loaded the Record & Replay plugin in the chat and told Codex I wanted a new skill. It requested permission to record the screen. After I granted it, the recording indicator appeared at the bottom.
I opened the writing app, selected the article, and copied it. In the Substack dashboard, I created a new post, pasted the text, added a subscription button at the end, returned to the top, entered the title without its numbering, and left the editor.
Once I stopped recording, Codex began creating the skill. It reviewed each recorded step, interpreted the intent and logic, and packaged the workflow into instructions.
Replay it and inspect the result
I had another member-only article ready, so I asked Codex to save it in Substack.
Following the new skill, it found Ulysses and the specified article, opened Chrome, created a Substack post, pasted the content, and moved the title from the body into the title field.
The draft was there. Its title and body were correct. The only problem was the missing subscription button.
I reported the omission. Codex reviewed the original recording, recognized an error in its interpretation, and added the subscription-button step to the skill.
I deleted the draft and asked it to run again. This time, it scrolled to the end of the body, placed the cursor after the final paragraph, and inserted the subscription button.
Inspecting the draft confirmed that it matched my demonstration.
Why this matters
Record & Replay combines multimodal understanding, needed to interpret the video, with computer use, needed to reproduce the actions.
Neither capability is unusual by itself, and skills have existed for some time. OpenAI has combined them so that workflows that are difficult to describe can be communicated visually.
That makes tacit knowledge explicit and opens the way for AI to learn many more human skills. It is why I consider this such an important Codex update.
That is all for this episode. If you want to understand AI, become a more capable independent creator, and meet people with similar interests, join the newtype community. See you next time.