OpenAI Employee Shares How to Build an AI Work Operating System with Codex

When AI evolves from a "code-writing tool" into a "work operating system," how exactly does the human role change? In an in-depth interview, Jason Liu from OpenAI's Codex team demonstrated how he integrates the ChatGPT Workbench and Codex into every aspect of his daily work and life. His core insight is thought-provoking: "The only work left is figuring out where you're dissatisfied with something, translating that into language, and telling the AI."

The Fundamental Difference Between ChatGPT Workbench and Codex

Many people confuse ChatGPT Work (the workbench) with Codex. Jason offers a clear explanation: they are essentially different UI views of the same Codex experience.

Codex is OpenAI's autonomous programming agent that runs in a cloud sandbox environment, capable of reading and writing files, executing code, accessing the internet, and outputting results through Git workflows. The ChatGPT Workbench is an interface layer for enterprise and power users that packages Codex's underlying capabilities into a more accessible task execution environment. Both share the same reasoning engine and tool-calling capabilities — the only difference is the frontend presentation. Codex shows the complete technical process (terminal output, file diffs), while the workbench abstracts these into a "task completed" results view.

A counterintuitive insight: if you want to build an excellent work productivity tool, the best starting point is actually a powerful coding agent. Codex shows you Git history, Pull Requests, and code changes; the workbench hides all of this behind the scenes. "If I'm making slides, I don't really need to see that I wrote 20 lines of Python to accomplish something."

Jason admits that most of his work now happens in the "workbench" because he "doesn't need to review code at all." This confirms a current trend — more and more people have stopped reviewing AI-generated code line by line.

Regarding model selection, his advice is highly practical: for everyday tasks (reading Slack, scheduling, managing calendars), the default Sol medium is sufficient — these tasks don't require maximum intelligence. Only when building complex prototype applications does he switch to extra high or ultra levels, setting ambitious goals and letting Codex work continuously for hours. This tiered strategy not only saves compute resources but also reflects a practical principle: match task complexity with model capability level to avoid overkill.

Persistent Threads: Managing All Context with an Obsidian Vault

Jason's thread organization approach upends most people's habits. He doesn't have hundreds of threads running in the background — instead, he treats each pinned thread as an independent workspace.

Thanks to Codex's excellent "compaction" capability, these persistent threads don't lose historical context. In long conversation scenarios, limited context windows are a core challenge for large language models. Compaction is a technique that distills lengthy conversation history into key information summaries, ensuring the agent still remembers core objectives, completed steps, and key decisions after hundreds of interaction rounds. This is similar to human working memory — you don't need to remember every word of every conversation, just retain structured key points. Thanks to this mechanism, Jason's persistent threads can run continuously for weeks or even months without losing critical context.

He has automated threads dedicated to reading Reddit, Twitter, and LinkedIn to collect user feedback; a "Chief of Staff" thread that tells him daily/weekly priorities; and various threads for preparing Dev Day, producing vision videos, personal drumming projects, and more.

More crucially, his knowledge management philosophy: all context is deposited in an Obsidian Vault, which is essentially just a collection of Markdown files. Obsidian is a knowledge management tool based on local Markdown files, with the core advantage that all data is stored in plain text without relying on any proprietary format or cloud service. This characteristic makes it naturally suited as a knowledge source for AI agents — Markdown files are both human-readable and editable, and extremely easy for language models to parse and reference. "I rarely actually open Obsidian — it mainly provides context for my agents." This Vault contains categorized projects, people, notes, daily records, and personal preferences. Since it's hosted on GitHub, whether switching to the cloud or other environments, it can be pulled and rebuilt at any time — any agent instance with repository access can obtain the complete knowledge base via git clone, achieving a "portable brain" effect.

Notably, when a task needs execution, Jason doesn't let the main thread handle it directly — instead, he has it "spawn a subagent." A subagent is an architectural pattern for decomposing and executing complex tasks: the main thread (like a "manager") handles understanding goals, splitting tasks, and aggregating results, while subagents (like "executors") process specific subtasks in independent sandbox environments. This design prevents the main thread's context from being polluted by execution details, allows multiple subtasks to run in parallel, and ensures that subagent failures don't affect the main thread's state. This can be done automatically in ultra mode, but Jason prefers more careful manual control.

The Chief of Staff System: Having AI Proactively Prioritize for You

The most practical part of Jason's system is what he calls the "Chief of Staff" — a scheduled task.

Every day at 9 AM, 1 PM, and 5 PM, this automation reads all his Slack messages, uses computer use to read Twitter DMs, checks unreplied emails and his Linear board, then provides a clear overview of "what needs attention right now." If there are statuses on Linear that need updating, Codex handles them directly. This heartbeat task design borrows from distributed systems concepts — in software engineering, a heartbeat is a periodic status signal sent between system components to confirm they're alive and synchronized. Jason transplants this concept into a personal productivity system, having Codex automatically check various information sources at preset intervals and proactively report, freeing humans from the cognitive burden of "pulling information."

Even better are the small features refined over time through accumulated use: because he travels frequently, whenever Codex detects booking information and reference numbers, it automatically checks him in for flights and texts the boarding pass to his personal phone. When he's in New York, he'll suddenly receive a boarding pass from Codex.

For ordinary users wanting to build this system, Jason offers an extremely simple starter prompt:

"Turn this thread into a heartbeat. I want you to check my email, Slack, and Linear at appropriate times — like 9 AM, 1 PM, and 5 PM — then tell me what to prioritize."

Then it's continuous iteration: initially it doesn't include links, so you say "add links to each item"; later it can pre-draft Slack replies and emails that you only need to edit and polish before sending. "It's just like hiring someone — on day one you give them tons of information and constant feedback, and they grow over time."

Skills and Plugins: Building a Personal AI Voice Library and Capability Packages

Jason's favorite category of capabilities are various "writing skills." He has skills like write me, tweet me, email me, plus skills that turn voice transcripts into blog posts and videos into video essays.

The creation process is equally simple: "I want you to use the Slack connector to read the past week's messages, then write a skill to figure out how to mimic the way I talk." As usage deepens, he also teaches it to distinguish contexts — using different tones for external users, team members, executives, and colleagues.

Regarding concept clarification, Jason is very clear: a skill is a component of a plugin. A plugin may contain several MCP servers, multiple skills, and some resources and scripts. MCP (Model Context Protocol) is an open protocol that standardizes how large language models interact with external tools and data sources. An MCP server is essentially a lightweight service endpoint that exposes specific capabilities to AI agents (such as reading databases, calling APIs, operating file systems). In Codex's plugin system, this layered architecture makes capability reuse and sharing modular. If he wants to share his writing capabilities with the team, he packages them into a plugin called "Better Writing." Currently, anyone can submit their own plugins to the Codex plugin directory — pure skill plugins get fast-tracked through review, while those containing MCP require additional security audits.

To combat skill bloat, he has a meta-skill called "self-improve": having Codex review the past 400 sessions, identify never-used skills and duplicate skills that can be merged, then automatically clean up. "It's no different from managing a team."

Browser and Computer Use: Work Continues Even After You Close Your Laptop

Jason considers browser and computer use to be Codex's most impressive capabilities. Computer Use refers to an AI agent's ability to control graphical interfaces through screenshot recognition and simulated mouse/keyboard operations. The technical path is typically: the agent captures a screenshot → identifies interface elements through visual understanding models → plans the next action → executes clicks, typing, etc. → takes another screenshot to verify results. The breakthrough significance of this capability is that it lets AI operate any software without an API — video editors, web forms, desktop applications — thereby vastly expanding the boundaries of automation.

He shared a classic example: while cycling, he received a colleague's request to re-export a video. Using his phone to remotely control his MacBook, he had Codex use computer use to find the video, identify the export tool, edit it, and send it back to Slack. Then he set it to check for feedback every 30 minutes, automatically producing V2, V3, V4. By the time he got home, the video had been approved and became an officially published video.

Two key tips for using the browser: first, understand that Codex's built-in browser supports login state (with cookie authentication), enabling access to more information; second, it can control multiple tabs. He gave a shopping example — instead of having AI produce a Markdown recommendation list, he has it open each candidate product in a separate tab. After he finishes his coffee and comes back, he directly compares four tabs and clicks to place the order himself.

Building Apps from Scratch: The Goal, Plan, and Work Log Methodology

The most impressive part of the interview was Jason demonstrating how he built a custom learning app for his "learn to play drums" project.

His methodology revolves around three files, all generated by Codex itself:

Goal.md: Defining Verifiable Success Criteria

For example, "able to paste a YouTube link and extract separate drum kit voices." The key is that goals must be verifiable — he even specifies test songs, video IDs, and requires Codex to use computer use to open the app, upload a video, and retrieve data to verify core functionality. "If you want to set good goals, don't write the goals yourself — tell Codex what you want and what success looks like, then let it set its own goals to accomplish it." This approach embodies the software engineering concept of Acceptance Test-Driven Development (ATDD) — define what "done" means first, then work backward to determine the implementation path.

Plan.md: Defining the Technical Implementation Path

Implementation details: built with React, two tabs, using Tailwind and shadcn as the tech stack — if unfamiliar, read the documentation. React is Meta's frontend UI framework, Tailwind is a utility-first CSS framework, and shadcn is a copyable UI component library — these tech stack choices reflect the mainstream pattern for AI-generated applications: choosing tools with rich documentation and active communities to ensure the AI can draw on sufficient reference knowledge from its training data.

Work Log: Tracking the Execution Process

Due to the compaction mechanism, he doesn't read messages one by one, but the work log helps him quickly understand where things got stuck — as a DX (Developer Experience) team member, this is his window into understanding tool limitations. The work log also serves another important purpose: when Codex's compaction summaries miss certain details, the log file provides a complete audit trail as persistent storage.

The benefit of separating goals into their own file: even while a task is still running, he can edit Goal.md at any time to expand or narrow the scope.

Core Philosophy: Learning to Be "Dissatisfied" Is the Real Competitive Edge

A phrase Jason repeatedly emphasizes is the soul of the entire interview:

"If you want taste, you have to eat. Our job now is to come up with the language, to understand more deeply what we actually want."

When asked "what should people learn now that programming has been solved," his answer is: learn to be dissatisfied with the status quo, and learn to articulate that dissatisfaction with precise vocabulary. Four years ago the question was "how do I write this code" — now the question becomes "why does the panel flicker when switching" — you ask Codex, it tells you it's because there's no preloading, and then you say "fix it." This shift means that technical literacy has moved from "knowing how to implement" to "knowing what good looks like" — aesthetic judgment and problem diagnosis abilities have replaced coding skills as the core competitive advantage.

He explicitly rejects the self-identification as a "manager." At OpenAI, everyone works hard. What makes you stand out is how much you care about whether something is done well enough, how much you care about whether it produces real results. For him, the success criterion for the drumming app isn't whether the code looks good, but whether he actually learned to play drums and whether the tool can be shared with others.

Looking ahead, Jason says Codex is migrating heavily from the desktop to the cloud-based ChatGPT Workbench experience. The direction he's most excited about is — "Work shouldn't stop just because you close your laptop." The technical foundation for this vision: when the agent runs entirely in the cloud, it no longer depends on the user's local device being online. It can execute tasks 24/7 and notify the user upon completion. Additionally, voice interaction (such as calling your Codex to give instructions) will also be a key focus area.

For everyone looking to boost productivity with AI, Jason's system offers a clear path: don't expect AI to solve your problems in one shot. Instead, treat it as a partner that needs continuous nurturing and will grow over time. What you truly need to cultivate is the ability to "know what you want and articulate it clearly."