crosstalktalks, distilled into a living wiki

Setting Yourself Up for Success with Codex — Jason Liu (OpenAI Workshop)

AI Engineer · watch the talk ↗ · updated Jul 25, 2026

Jason Liu, Developer Experience at OpenAI and creator of Instructor, delivers a 75-minute hands-on workshop on using OpenAI Codex as a personal productivity operating system — covering memory vaults, skills/plugins, computer use, thread automation, and multi-thread orchestration. The central thesis: investing in persistent memory, voice input, and long-running pinned threads transforms Codex from an IDE assistant into something closer to a team of autonomous colleagues.

Key Takeaways

  • Compaction has made long-lived threads viable; the old advice to start a new thread every 20 messages is obsolete. 6:49
  • Appshots (Command + screenshot capturing the full accessibility tree) outperform plain screenshots by providing structured IDs (channel IDs, user IDs) that reduce tool-call hops dramatically. 16:38
  • Voice input (~3× faster than typing) plus a foot pedal for transcribe/enter removes the keyboard bottleneck; Liu advocates dictating messy 15-minute voice memos directly to the AI. 8:10
  • The personal monorepo pattern (one pinned project directory, structured projects/ and people/ subdirectories, per-project agents.md files) is Liu's primary context-management strategy. 18:03
  • Pinned threads that talk to each other via list thread and send message to thread tools are the current frontier of multi-agent orchestration inside Codex. 6:10, 1:07:46
  • Use low/medium thinking most of the time; running o3 / "X High" on trivial tasks (e.g., finding a coupon) wastes tokens with no quality benefit. 1:11:50

Memory Vault & Personal Monorepo Architecture

Liu's memory system is a git-tracked monorepo (jxnl/personal-monorepo-template) open-sourced for cloning. 17:59 Structure:

  • projects/<name>/ — per-workstream directories each with their own README and agents.md specifying relevant Slack channel IDs, preferred package managers, and style constraints. 1:00:04, 1:02:49
  • people/<name>/ — CRM-style files storing email addresses, Slack connector IDs, areas of work, and channels. 1:00:11
  • A maintained todo.md file polled by a dedicated thread; sub-agents verify whether items are genuinely incomplete. 1:00:41

Liu treats the vault as a git repo and runs git diff periodically to audit what agents have written without reading full session JSON. 22:37 He advocates including Slack channel IDs directly in project front-matter so models discover and read relevant channels without being explicitly instructed. 46:01

A write-like-me skill trained on six months of emails and Slack messages generates style-consistent drafts, adapting tone by context (e.g., stricter for customer support forms). 20:33

Skills, Plugins, and the "Plugin Hero" Mindset

Skills are small file-and-script constructs; plugins are curated libraries of skills. 9:18 Key sources: the skill installer skill (lists OpenAI-curated skills including GitHub, Playwright, Notion integrations), skillset.sh, and Vercel's skills tool. 13:31

Liu's highest-impact internal skill: a DX triage skill that knows relevant Slack channels, which engineers own which features, and automatically requests feedback IDs when regressions are DM'd to him — executing appshots to reply via Twitter computer use. 15:37, 36:19

Skills are allowed to self-edit: every time you run this skill, you're allowed to edit yourself if you learn something new. 15:01 Liu builds skills fast for personal use (yolo, self-improving on mistakes), then shares with teammates only after ~2 months of real-world validation. 24:02

The plugin hero frame: the metric that matters is not tokens consumed but how often teammates use the skills you've built. 5:12 Example: a "finalize before PR review" skill adopted company-wide at OpenAI that consistently catches style-guide violations. 5:21

Computer Use, Appshots, and Remote Control

Computer use is Liu's favored "feel the AGI" feature: the agent controls any native application (iMovie, trading software, Safari) in the background without taking over the foreground. 10:46 Locked-use mode in Settings allows computer use even with the laptop lid closed. 29:03

Chrome extension vs. computer use: the extension targets Chrome only but leaves the main screen free; computer use targets all apps. Models intelligently route — Chrome → browser extension, Safari → computer use. 10:50, 17:29

Appshots specifics:

  • Keyboard shortcut: Command + side-by-side button. 4:34
  • Captures both the visual screenshot and the full accessibility tree, giving the model exact element IDs. 16:40
  • Practical result: one function call to post to a Slack channel vs. many OCR + lookup hops with a plain screenshot. 16:58

Remote control (iOS app → QR code scan → phone controls any pinned thread on the laptop) enables on-the-go delegation; Liu uses it when leaving the office and being asked for something. 42:22

Documented computer-use examples from Liu: editing iMovie, auto-checking into flights (downloading boarding pass, sending via iMessage), filling DocuSign and faxing medical records, negotiating an Amazon refund (~$400 returned after a shower). 39:18, 27:31, 25:45, 1:13:20

Security caution: a sufficiently prompted model may bypass a plugin restriction (e.g., Slack file upload blocked) by falling back to computer use. Liu mitigates this with agents.md guardrails and the auto-review permission mode. 54:05

Thread Orchestration: Heartbeats, Goals, and Manager Threads

Pinned threads as teammates: each thread corresponds to a workstream (chief of staff, Agents SDK, open-source program). 34:23 Threads that persist for weeks with hundreds of sub-agents are viable because compaction works. 3:34

Heartbeat automations: scheduled messages back into the same thread (not new threads) to maintain continuity. The loop skill (/loop) is the monorepo equivalent of "keep an eye on this." 35:37 Example: a PR-watching loop that integrates reviewer feedback, rebases on master, and keeps CI green from Monday to Thursday without user involvement. 35:52

Goals / Ultra Goal: /goal defines a verification step; the agent loops until the verifier passes. 40:14 The ultra goal skill stores the goal in a goal.md file so scope can be edited while the loop is running; a work-log file tracks long-running (1–2 day) tasks without requiring the user to parse multi-gigabyte session JSON. 41:16, 41:53

Multi-thread orchestration: threads expose list threads and send message to thread tools. A monitor thread watches for issues; on detection it spawns a new triage thread, pins it to the sidebar, and sends updates back upstream when the issue recurs. 56:34, 57:06 Liu characterizes the progression: IC + IDE → pinned threads (you as manager) → manager threads coordinating IC threads. 6:25

Liu's recommendation for testing thread control: ask one thread to read all pinned threads, rename them descriptively, and emoji-color-code their readiness. 1:08:54

Thinking-level discipline: Liu sets his chief-of-staff thread to default medium thinking; "low thinking on o5 [o3-mini equivalent] is still so much better than prior models" and is sufficient for most coordination work. 1:12:24