Jason Liu shares repeatable Codex skills including /write-like-me

• Codex setup became a personal-ops stack: the episode announcement says Jason Liu used it for a Slack/email chief of staff, session-mined skills, and long-running projects with verifiable goals.
• The core loop is skill mining: the five-takeaways thread starts with asking Codex to review recent sessions and suggest new skills for repeated tasks.
• Voice cloning is treated as a reusable work skill, not a one-off prompt: in the /write-like-me post, Jason has Codex study a week of Slack messages and draft for coworkers, executives, and external users.
• The weirdest demo was a remote video-editing loop: the bike-ride story says Codex edited, exported, posted to Slack, then checked back every 30 minutes for feedback.
• The browser path matters for logged-in tasks: the browser post points to Settings > Browser > Import before asking Codex to handle web tasks like flight check-in.

Jason Liu’s setup is a practical Codex cookbook: Peter Yang’s written recap promises prompts for a Slack-and-email chief of staff, session-mined skills, and long tasks with verifiable goals. OpenAI’s Skills & Plugins docs make the same pattern formal: a skill packages reusable instructions and resources, while a plugin is the installable bundle. The odder part is physical: Computer Use can operate desktop apps, and Record & Replay can watch a Mac workflow once and draft a skill from it.

Jason Liu, described in the episode announcement as DevEx at OpenAI, framed Codex less like a coding chatbot and more like a work operating system. The pattern is recursive: make Codex study the work, turn the repeated parts into skills, then package the useful skills so they travel.

The five mechanics from the takeaways thread and its follow-ups:

1. Ask Codex to mine recent sessions for repeated tasks and propose new skills.
2. Build a /write-like-me skill from a week of Slack messages.
3. Give durable workstreams their own pinned threads.
4. Import browser sign-ins so Codex can complete logged-in web tasks.
5. Bundle related skills into a plugin.

In the /write-like-me post, Jason asks Codex to analyze how he writes to coworkers, executives, and external users, then turn that into a skill for drafting replies in his voice.

OpenAI’s Build skills guide says a skill can include instructions, resources, optional scripts, and references inside a folder with a required SKILL.md. The same guide says Codex initially sees the skill name and description, then loads the full instructions only when it selects the skill.

That progressive loading is the clean part for creative work: the brand voice lives as a callable capability instead of being pasted into every prompt.

Jason’s durable-thread pattern keeps one topic per long-running Codex thread, preserving context, decisions, and open loops, according to the pinned-thread post.

OpenAI’s long-running work docs describe the official version: keep related work in the same chat, give ChatGPT a clear outcome, constraints, and definition of done, then use /goal when the task needs completion criteria.

The docs reduce “done” into three fields:

• Outcome: the result, not only the activity.
• Constraints: required tools, boundaries, compatibility needs, or approaches to avoid.
• Verification: tests, measurements, or review criteria.

The daily-ops version is a scheduled triage loop. In the heartbeat demo, Jason’s prompt asks Codex to check email, Slack, and Linear at 9 am, 1 pm, and 5 pm, then report what he needs to prioritize.

The loop improved by correction. Jason said the first run lacked links, then he asked for links, added Linear, and ended up with pre-drafted Slack replies and emails in the same flow.

OpenAI’s Scheduled tasks docs say scheduled tasks can run inside an existing chat when the task should reuse that chat’s context. The same page lists polling Slack, GitHub, or another connected source as a fit when results should stay in the same chat.

The logged-in browser workflow starts with import. The browser post says Jason imports Chrome cookies and sign-ins through Settings > Browser > Import, then asks Codex to complete logged-in web tasks, including flight check-in.

OpenAI’s Browser docs say the desktop app’s built-in browser uses a separate profile from the user’s regular browser and that Settings > Browser manages browser data and profile-import features when available. The same docs say ChatGPT asks before using a website unless that site has already been allowed.

The launch-video anecdote pushes this beyond web browsing. The bike-ride story says Jason remotely connected from his phone, asked Codex to edit a launch video, export it, post it back to Slack, and check the thread every 30 minutes for V2, V3, and V4 feedback.

OpenAI’s Computer Use docs say Codex in the desktop app can see and operate graphical user interfaces on macOS and Windows after the Computer Use plugin is installed and permissions are granted.

Jason’s packaging example is simple: group write like me, Slack like me, and email like me into one writing plugin, according to the plugin-bundling post.

OpenAI’s Plugins docs define plugins as bundles that can include skills, connectors, MCP servers, browser extensions, hooks, and scheduled task templates. The docs also say plugins are available in ChatGPT Work on the web, ChatGPT Work or Codex in the desktop app, and Codex CLI, but not Chat, the IDE extension, or mobile.

OpenAI’s Record & Replay docs add a second path into the same skill system: demonstrate a workflow on a Mac, then let ChatGPT or Codex inspect the captured actions and draft a skill.

The availability is narrow. Record & Replay requires Computer Use, works on macOS, and initially excludes the European Economic Area, the United Kingdom, and Switzerland.

The official examples include filing an expense, booking parking, creating a configured issue, publishing a video, and downloading a recurring report. That “publish a video” example sits close to Jason’s launch-video loop, where Codex handled editing, export, Slack posting, and repeated feedback checks.