OpenAI Expert Teaches How to Max Out Codex

Jason Liu, the developer behind the 13k-star open-source library Instructor, has been hired by OpenAI to work on the Codex team. True to his reputation, he hasn't held back. Not only did he give away API credits on social media, but he also published a guide called 'Codex-maxxing,' revealing his personal strategies for getting the most out of Codex. His approach transforms Codex from a simple chatbot into a persistent, autonomous work system that can handle tasks around the clock.

Liu's core philosophy is to keep Codex sessions alive for as long as possible. Instead of closing conversations after a single query, he maintains multiple threads that last for months, each dedicated to a specific workflow: one for scheduling, one for open-source projects, one for social media monitoring, and so on. He assigns each thread a permanent spot and jumps between them with Command-1 through Command-9. These threads accumulate conversation history, preferences, and decisions over time, allowing the agent to pick up where it left off without needing a fresh context. He largely abandons typing in favor of voice commands, arguing that speaking preserves raw thought processes without the need to optimize prompts. With Codex's Steering feature, he can interrupt the agent mid-task to add new instructions and walk away without waiting.

What truly elevates Codex from a tool to an employee, however, is the combination of Heartbeats and @computer. Heartbeats add a periodic task scheduler to the agent. Liu has a 'Chief of Staff' thread that runs every 30 minutes, checking Slack and Gmail for messages, prioritizing them, and drafting replies for human approval. In a more complex example, while working on an animation project, he would post a video to a Slack review thread and have Codex check it every 15 minutes. If a colleague gave feedback, Codex would render a new version and upload it back, even using @computer to click the 'Add file' button when the Slack MCP didn't support file uploads. Another time, he had Codex monitor Amazon customer service queue while he showered; by the time he finished, the refund had already been processed.

Central to Liu's workflow is a verification mechanism that determines when a task is complete. He once tasked Codex with porting Python's Rich library to Rust, with the hard requirement that the Rust version pass all of the original Python library's unit tests. The agent iterated until the tests succeeded. Liu emphasizes that 'without a verification mechanism, ambition is just a wish.' OpenAI has since promoted Goal Mode from experimental to official, allowing users to set a final objective and acceptance criteria, with Codex autonomously working towards it over hours or days, while users can check progress or adjust direction.

On memory management, Liu takes a contrarian stance: he doesn't trust Codex's built-in memory system. Instead, he stores all his long-term working memory in an Obsidian vault, organized into directories like TODO, people, projects, agent, and notes. His top-level AGENTS.md file instructs the agent to update the knowledge base whenever information changes. This approach keeps his data in local, version-controllable files that he can manually edit, compare, or roll back, avoiding vendor lock-in. He acknowledges Codex's Chronicle feature for screen-based context but finds it still immature in terms of permissions, rate limits, and privacy.

Codex itself has been evolving. The sidebar now supports rendering Markdown, filtering tables, and viewing PDFs and PPTs. Agents can control web pages via JavaScript through a built-in browser, and users can annotate without switching windows. Liu often works with multiple side panels open simultaneously, such as Storybook for UI review, Remotion Studio for animations, and Slidev for presentations. His favorite deliverable is a single index.html file with JS and CSS, which requires no deployment. He also packages successful workflows as reusable Connectors and Skills, so Codex can reuse them without relearning. Recent updates include lock-screen capabilities—Codex can continue working when the computer is locked, and users can monitor, approve, or take over tasks from their phones. The message is clear: while you clock out, Codex clocks in. But as AI increasingly takes over sustained work, humans might find themselves with more leisure time.