Summary
Peter Yang’s interview with Jason Liu — a member of OpenAI’s Codex developer experience team — doubles as a detailed live demonstration of how Liu has built his entire work life around ChatGPT Work and Codex. Liu walks through his actual setup: dozens of pinned threads functioning as persistent workspaces (covering everything from Reddit/Twitter sentiment monitoring to dev day planning and personal projects like a drumming app), automated chief-of-staff loops that surface daily priorities, and Slack-integrated skills that pre-draft responses and close the loop on requests without prompting.
The video addresses practical questions many power users face: when to use ChatGPT Work versus Codex (primarily a UX difference, with Codex exposing git history and pull requests), which model and effort level to default to (Sol medium for operational tasks, ultra for complex prototyping), and how to manage skill drift over time. Liu’s approach to skill self-improvement is particularly detailed — a “self-improve” skill that reads recent Slack messages, tweets, and blog posts to update its writing style, and a system that reviews session logs for consistent feedback patterns across all invoked skills.
For developers and power users building personal AI operating systems on top of OpenAI’s platform, this is one of the most concrete and reproducible demonstrations available of how Codex’s long-running pinned threads, compaction, and sub-agent spawning work together in daily practice — delivered by someone who ships the product.
📺 Source: Peter Yang · Published July 26, 2026
🏷️ Format: Tutorial Demo
