Video summary
Master 96% of Codex in under 35 minutes
Main summary
Key takeaways
Summary
The video is a practical introduction to 18 Codex concepts. It shows how the presenter uses Codex as a project-aware assistant for both technical and nontechnical work.
Projects and agent behavior
- Projects: A project is a collection of files and folders that gives Codex context about ongoing work. The presenter’s example project contains business information, content, and past work, allowing Codex to answer questions about previous activity.
AGENTS.md: A project-level instruction file that sets ground rules and points Codex to relevant information, such as a wiki, project notes, or writing guidance. The presenter describes it as a way to avoid re-explaining context in each chat.- The agent loop: Codex can plan, inspect files, run commands or use other tools, and then report back. The presenter recommends observing its actions to understand how it approaches a request.
- Goals: A goal lets Codex keep working toward a defined outcome. The presenter recommends making goals measurable when possible and says that giving the agent room to work can be more effective than directing every step.
The presenter also explains how he adapted a project from Claude Code: he copied its instructions and skills into Codex-compatible locations, then had Codex help with the conversion.
Environments and configuration
- Local, cloud, and remote access: Local work uses files on the computer, while cloud environments are sandboxed. The presenter demonstrates using the mobile app to access Codex chats on his desktop remotely, provided the computer is on.
- Worktrees: A worktree is a separate copy of a project for parallel or experimental work. It can help avoid disrupting the main version, especially during software development.
.codex: This folder can contain project-specific settings, while user-level settings apply across projects. The presenter says users need to know where it is but generally need not manage it constantly.
Models, usage, and customization
- Models and tokens: Models differ in capability and cost. The video distinguishes Codex subscription usage from API billing, which is priced by input and output tokens. The presenter says subscriptions can provide more usage for the price, though extra credits may be needed after limits are reached.
- Effort and speed: Higher effort can improve results but uses more resources. The presenter recommends matching the model and effort to the task rather than using the most powerful setting for everything. Fast mode is described as quicker but more usage-intensive.
- Permissions: Codex can be configured to ask before taking actions, approve routine work automatically while stopping for risky actions, or run with broad access. The presenter suggests starting with more confirmation when learning how the system behaves.
- Skills: Skills are reusable workflows or instructions for repeat tasks. They can be invoked in natural language or with a slash command, and improved over time with feedback.
.agents: Skills are stored here, at either the project or user level. A skill is typically a Markdown file containing metadata and instructions, and it can refer to other skills or supporting files.- Plugins: Plugins connect Codex to services such as email, storage, project management, and creative tools. The presenter says plugins are convenient because they can handle sign-in; services without a plugin may still be connected through an API key or environment configuration.
Tools and scaling
- Browser: Codex can interact with websites through its browser, including sites where the user is signed in. The presenter demonstrates this with a community site and a Canva task, and praises the browser experience.
- Sites: Codex can publish projects for sharing, with options such as access controls, analytics, custom domains, and database connections. The presenter presents this as a way to host work created locally.
- Subagents: A main agent can delegate research or other tasks to several agents, potentially using different models. They work in parallel and report their results back to the main session.
- Scheduled tasks: Codex can run prompts automatically on a schedule, using a selected project and its associated context and skills. Schedules can be configured for recurring work or reminders.
- Voice mode: Voice can be used to start tasks, invoke workflows, and coordinate chats. In the demonstration, the presenter asks Codex to create a thumbnail in a separate thread and send it to an existing research thread. He also notes that voice can be used remotely from a phone.
Guides and demonstrations
The video includes demonstrations of:
- Setting up project context with
AGENTS.md. - Converting a Claude Code project for use with Codex.
- Asking Codex to review past work and retrieve YouTube comments.
- Using a goal to create and visually check motion graphics.
- Understanding local versus cloud work and remote desktop access.
- Choosing models, effort levels, and permissions.
- Building and refining reusable skills for content tasks.
- Connecting services with plugins and using the browser.
- Publishing a project as a site.
- Delegating research to subagents, scheduling tasks, and coordinating work by voice.
Overall assessment
The presenter favors Codex as a flexible, project-aware assistant and emphasizes using context, reusable workflows, delegation, and automation. He also advises choosing permissions and model effort according to the task and its risks.
Main speaker/source: Nate Herk.
Rate this summary
Your feedback will help improve summaries.
Improve this summary
Reprocess with a stronger model when the summary feels incomplete or inaccurate.
Translate summary in another language
Ask questions to this video
Chat for follow-up questions, clarifications, and source-backed answers.