Your Codex tutorial is ready

    OpenAI Codex
    field guide

    Work through the sections in order on one real task. The prompts are designed to produce evidence you can inspect, not just confident text.

    Start here

    1. 1. Choose one small task with a clear definition of done.
    2. 2. Open the relevant repo or workspace and read its current instructions first.
    3. 3. Run the framing prompt, then let the task’s evidence decide the next step.
    4. 4. Save the durable lesson in the right instruction, skill or team policy.

    Table of contents

    1. 01Choose the right surface
    2. 02Write the outcome contract
    3. 03Make the first run safe
    4. 04Create durable repo context
    5. 05Plan and isolate work
    6. 06Turn behavior into skills
    7. 07Connect MCP deliberately
    8. 08Use subagents with control
    9. 09Ship evidence
    10. 10Teams and enterprise governance
    11. 11The non-developer route
    12. 12Handoff and maintenance

    The truth about Codex workflows

    Codex is powerful when the task has a legible context boundary and a verifiable finish. The model does not replace repository rules, access control, tests or human ownership. Treat current product availability and workspace settings as things to verify in the OpenAI docs and your own account.

    Start by deciding whether the work belongs in a normal chat, ChatGPT Work, or Codex operating on a repository. Chat is useful for thinking and drafting. Work is useful for connected knowledge-work contexts. Codex is useful when the source of truth is code, files and a testable change.\n\nAsk where the source of truth lives.\nName the data and permission boundary.\nChoose the smallest surface that can finish the task.\n\nClassify this task as Chat, ChatGPT Work or Codex. Explain the source of truth, permissions, expected artifact and verification required. If Codex is appropriate, state the repo and first safe read-only checks.

    Prompt to use

    Start by deciding whether the work belongs in a normal chat, ChatGPT Work, or Codex operating on a repository. Chat is useful for thinking and drafting. Work is useful for connected knowledge-work contexts. Codex is useful when the source of truth is code, files and a testable change.\n\nAsk where the source of truth lives.\nName the data and permission boundary.\nChoose the smallest surface that can finish the task.\n\nClassify this task as Chat, ChatGPT Work or Codex. Explain the source of truth, permissions, expected artifact and verification required. If Codex is appropriate, state the repo and first safe read-only checks.

    For a team rollout, separate capability from authorization. OpenAI documents ChatGPT seats and Codex-only seats as different access paths, and workspace admins can manage members, groups, permissions, apps and analytics. Confirm the current plan and controls in your workspace before promising a capability.\n\nPilot with approved repositories, limited groups and named reviewers.\nCheck ChatGPT versus Codex-only seat requirements, app allowlists, custom MCP and read/write action controls.\nDefine human approval, audit evidence, data boundaries and exit criteria before production use.\n\nDraft an enterprise Codex pilot brief with workspace, seat, group, app/MCP, repo, data, reviewer, audit and rollback fields. Link every product-specific assumption to current OpenAI admin documentation and mark unknowns for verification.

    Prompt to use

    For a team rollout, separate capability from authorization. OpenAI documents ChatGPT seats and Codex-only seats as different access paths, and workspace admins can manage members, groups, permissions, apps and analytics. Confirm the current plan and controls in your workspace before promising a capability.\n\nPilot with approved repositories, limited groups and named reviewers.\nCheck ChatGPT versus Codex-only seat requirements, app allowlists, custom MCP and read/write action controls.\nDefine human approval, audit evidence, data boundaries and exit criteria before production use.\n\nDraft an enterprise Codex pilot brief with workspace, seat, group, app/MCP, repo, data, reviewer, audit and rollback fields. Link every product-specific assumption to current OpenAI admin documentation and mark unknowns for verification.

    Current source checks

    Use the official product docs for changing plan, seat and workspace facts. Use the related site content for context, then verify it against the live product before publishing.

    OpenAI CodexOpenAI Academy: What is Codex?Codex for teamsCodex vs Claude Code context

    Run this on one real task today.

    The tutorial becomes useful when it changes a workflow. Start small, inspect the result and keep the lesson that survives contact with reality.

    Back to the resource