ChatGPT Work + Codex Tutorial: How AI Becomes Your Personal Operating System

According to @jxnlco (OpenAI DevEx member) in an exclusive demo, ChatGPT Work and Codex are essentially two views of the same engine; Work simply hides git and code. The reason is that files, scripts, and verifiable execution results are the foundation of 'action capability,' so the best starting point for building a productivity tool is a powerful coding agent.

This insight means that the current differentiation of AI productivity tools (e.g., Work for general tasks, Codex for programming) is only superficial; underlying capabilities are shared. Users don't have to choose between a 'work agent' and a 'coding agent' but can build personalized workflows on the same engine.

@jxnlco keeps only about a dozen pinned threads as workspaces, and has the model write important content into a notes directory. His main project is a pure-markdown Obsidian vault stored on GitHub, but the key detail is that he almost never opens Obsidian himself—it's a memory layer written for the agent to read; humans can read it incidentally.

This design shifts context from easily lost conversation history to persistent files, enabling the agent to access structured information across sessions. Additionally, the file system naturally supports version control and collaboration, providing a foundation for long tasks and team scenarios.

His chief-of-staff agent reads Slack, email, Twitter DMs, and Linear three times a day but is not allowed to send messages automatically. The reason: 'It just read 40 emails and 100 messages; I need to be more surgical.' However, in specific task threads, he gives it freedom, e.g., 'Once tests are all green, go DM Andrew on Slack yourself.'

Automation authorization is not a starting point but the result of manually requesting and validating N times. This principle prevents the AI from making wrong decisions when overloaded with information while retaining human control over critical outputs.

All sessions are saved as JSON. @jxnlco can query 'which skills have never been used in the past 400 sessions? Which should be merged? What feedback have I repeatedly given?' He even wrote a self-improve skill to perform this analysis automatically. The mental model: this is no different from hiring on day one; no skill is ever written perfectly once and for all.

This metacognitive ability allows the system to self-evolve, periodically cleaning up redundant skills, merging similar functions, and adjusting behavior based on feedback, thus avoiding skill library bloat and performance degradation.

When shopping, @jxnlco doesn't ask for a markdown report; instead, he has the AI 'open a browser tab for each product' so he can come back and click checkout directly. A report requires the user to translate it into action, but four tabs don't. Output should land directly on the interface where the user would make decisions anyway.

This principle applies to all AI interactions: avoid generating intermediate products that require secondary processing; instead, deliver directly executable results, reducing cognitive load and operational steps.

The video distills four reproducible practices: 1) Heartbeat starter: have the agent check email, Slack, and Linear at fixed times and report priorities; 2) Write-me skill: have the agent read the user's past week of writing to infer personal style and refine it by scenario; 3) Long-task three-file method: goal.md (verifiable success criteria), plan.md (implementation details), worklog.md (human log), with the goal set by the agent based on the user's described success criteria; 4) Use session history to periodically clean up the skill library.

These methods emphasize starting simple and gradually adding complexity. For example, the heartbeat first requires each item to include a link, then integrates with the task system, and finally pre-writes drafts. The write-me skill is rated as 'the most cost-effective trick' because it uses the user's own corpus to teach the AI style without manual rule-writing.

Codex has always been a desktop product, but many capabilities are migrating to ChatGPT Work cloud, with automation running in the cloud rather than on the user's machine. Voice interaction is the next step; @jxnlco uses dictation extensively and mentions the host's desired feature of 'calling your own Codex to give instructions.'

The video specifically mentions RESET's Tibo (@thsottiaux) and his influence on the Codex team: he not only represents the company to users but also represents users to the company, and his interaction with the community has driven frequent updates.