
Tech • IA • Crypto
A new capability allows users to teach Codex workflows by demonstration, enabling it to automate repetitive digital tasks with learned preferences.
Users can now train Codex by performing a task once while it observes, allowing the system to learn workflows directly from real actions rather than written instructions. This approach captures nuanced preferences, such as file organization and platform-specific steps, that are often difficult to describe in prompts.
A typical use case involves publishing a YouTube video. The process includes pulling metadata from a spreadsheet, attaching thumbnails, adding captions, and configuring privacy settings. Codex records each step, including how assets are matched and verified, then converts the sequence into a reusable automated skill.
After observing a task, Codex generates a structured skill that encapsulates the workflow. This includes remembering where metadata is stored, how files are organized, and the exact order of operations required. These skills can be reused without re-explaining instructions.
Once trained, Codex can independently execute the same workflow on new inputs. For example, when provided with a new video package, it can match it to the correct spreadsheet entry, populate metadata, upload assets, apply settings, and confirm completion without manual intervention.
Codex can carry out tasks using a combination of computer control, browser actions, and connected plugins. This allows it to operate across multiple systems, replicating complex workflows that span different platforms and tools.
The system minimizes reliance on step-by-step prompts. Instead of repeatedly specifying instructions, users can rely on previously learned skills, making automation faster and more consistent across repeated tasks.
The same capability can be applied to other workflows, such as formatting and sharing pull requests, managing documents, or setting up calendar events. Any repeatable digital process can potentially be captured and automated.
By learning directly from user demonstrations, Codex transforms repetitive workflows into reusable automated skills, streamlining complex tasks across tools and reducing the need for manual input.