Jason Liu's Codex Workflow for Work, Skills and a Drum-Practice App

Jason Liu of OpenAI's Codex team shows Peter Yang how he organizes work, maintains skills and builds a personal drum-practice app. The discussion reflects his workflow and product experience at the time of recording. The episode is sponsored by Google AI Studio.

Liu describes ChatGPT Work as a work-oriented view of a similar agent experience, while Codex exposes more coding details such as changes and pull requests. For routine organization he chooses a moderate effort level; for more complicated application work he says he may choose a higher one. These are personal preferences rather than permanent product recommendations.

He organizes ongoing work around pinned conversations and a Markdown-based Obsidian vault. The vault contains projects, people, notes and preferences; he asks the agent to save important information there. He describes it as a starting point from which he can access other repositories and projects.

His daily chief-of-staff automation summarizes information from sources such as Slack, email and a task board at scheduled times. He says he added capabilities gradually, including links to sources and prepared drafts. Some actions can be automated in his setup, but he distinguishes broad summaries from situations where he chooses a more specific instruction.

A simple starting instruction, Liu suggests, is to ask for a recurring check of selected sources and a list of priorities. Then refine the result: request links, clarify which tasks matter and adjust how drafts are prepared. He says his own workflow developed over time rather than appearing fully configured from the start.

His writing skills use examples of his own messages to capture voice, with differences for teammates, external users and executives. He periodically reviews and consolidates skills, sometimes asking the agent to examine prior sessions for repeated feedback. Those changes are iterative and depend on having relevant material available.

For browser work, Liu describes using a signed-in browser and multiple tabs for research. In one shopping example, the agent opens candidate products and he compares them and checks out himself. He also uses computer interaction for forms and testing apps; the available integrations and interfaces can change over time.

Liu's drum-practice app grew over roughly a week as he learned what he wanted. It includes practice patterns and adjustable tempo; he demonstrates a four-limb grid. He describes further ambitions for breaking a drum solo into practice sections, which should be read as a goal rather than a completed capability.

For long-running work, he distinguishes a goal file defining verifiable success, a plan describing implementation choices and a work log recording progress and obstacles. In one example, success requires opening the app, supplying a specific video and checking the resulting data. He discusses refining goals as the project evolves.

Liu argues that a builder can learn while making a tool: notice a specific problem, such as flickering panels or missing keyboard shortcuts, describe it and iterate. He contrasts actionable feedback with repeatedly asking an agent to make something better without explaining what is wrong.

He judges the drum app by whether it helps him learn and can be shared with others. He also expresses interest in cloud-based work that can continue after a laptop is closed. That is the direction he discusses, not a guarantee about every scheduled workflow or product configuration.