ChatGPT Work and Codex Usage Guide

A tutorial and Q&A guide detail ChatGPT Work and Codex usage. ChatGPT Work is described as a coding agent with the git and code layers hidden, operating across connected apps (Slack, Gmail, Drive, etc.) to deliver finished artifacts like documents and web apps. Codex operates on code repositories for code changes. Key power-user practices include: storing context in files (an Obsidian vault on GitHub) rather than chat history, so agents read from persistent memory; restricting write access for agents with broad read permissions, granting auto-send only after repeated manual verification; using session history to audit and improve skills; and demanding directly actionable outputs (e.g., opening browser tabs) instead of reports. Replicable workflows include a 'heartbeat' routine for triaging communications at set times, a 'write-me' skill that learns the user's writing style from past messages, and a three-file method (goal.md, plan.md, worklog.md) for long-running tasks. Work and Codex share an agentic usage quota separate from Chat. Pricing ranges from Free to Enterprise, with GPT-5.6 sub-models Sol, Terra, and Luna at different capability and cost tiers. Work is available on web, mobile, and desktop; Codex is desktop-only, with the original Codex app merged into the ChatGPT desktop client.
Key insights
Work Is a Coding Agent with the Code Layer Hidden
ChatGPT Work and Codex share the same underlying agent engine; Work simply hides the git and code interface. This design choice makes agentic capabilities accessible to non-developers while preserving the file-based execution model that makes coding agents reliable.
Persistent File-Based Context Outperforms Chat History
Storing important information in markdown files within an Obsidian vault on GitHub, rather than relying on chat context, creates a durable memory layer that agents can read across sessions. The human rarely opens these files—they are written for agent consumption.
Automation Trust Should Be Earned Through Repeated Verification
Agents with broad read access (email, Slack, DMs) should not be given write/send permissions initially. Auto-sending is granted only after manually verifying the agent's actions many times, treating automation as an earned privilege rather than a default.
Product Differentiation Reflects a Spectrum from Information to Action
Chat handles Q&A, Work handles cross-app task completion, and Codex handles code changes. The shared agentic quota pool between Work and Codex, separate from Chat, reflects the higher resource cost of action-oriented tasks.