Master 97% of Codex in 1 Hour (full course)

Codex is ChatGPT's desktop super app that can build websites, automate workflows, and control your entire computer—but most users barely scratch the surface. How do you transform a chat interface into a full-stack development environment that creates working dashboards, connects to APIs, and deploys live products while you sleep? And when should you use Codex versus ClaudeCode—or both? This full course walks you from first launch to a deployed, automated YouTube analytics system, revealing the skills, automations, and deployment workflows that separate casual users from power builders.

Nate Herk | AI AutomationTech1 Упомянутых людей 5 Терминов глоссария

Длительность видео: 1:00:59·Опубликовано 6 мая 2026 г.·Язык видео: English

7–8 мин чтения·14,726 произнесённых слов → сжато до 1,530 слов (10x)·

Смотреть на YouTube ↗

1 —

Ключевые выводы

1

Codex is a local super app that goes far beyond web ChatGPT: it reads and writes local files, automates browser tasks, builds reusable skills, and deploys projects to production—making it ideal for pragmatic execution and long-form workflows.

2

Skills are markdown recipe files that codify repeatable workflows; after building something once, you turn it into a skill so Codex can replicate it consistently—and improve it over time with feedback.

3

Automations run scheduled tasks (like weekly data refreshes) inside Codex projects, but require the app to stay open unless moved to the cloud—making them best for supervised iteration before full deployment.

4

Deploy locally built projects to the web by connecting Codex → GitHub → Vercel: changes pushed to GitHub auto-deploy to production, creating a clean separation between local testing and live sites.

5

Use medium intelligence for planning and high for complex builds; extra-high is rarely needed and burns tokens quickly. Plan mode prevents execution until you approve, saving time and session limits.

Вкратце

Codex turns ChatGPT into a local agent that can build, test, deploy, and automate entire projects—from API connections to live dashboards—using skills, browser automation, and GitHub/Vercel deployment, all while keeping you in control of a structured, reusable workspace.

2 —

Codex vs. ChatGPT and ClaudeCode: What's the Difference?

Codex is ChatGPT with hands—it automates locally, builds skills, and deploys products.

WEB CHATGPT

Chat interface with limited connectors

ChatGPT on the web lets you talk to the model and use a few connectors for extended functionality, but it can't touch your local files, automate your computer, or build deployable projects. Codex, by contrast, is a full super app: it reads Excel sheets, navigates local directories, uses mouse and keyboard automation, and can build websites, apps, and video games—then push them live while you sleep.

CLAUDECODE

Exploratory and creative, uses Opus/Sonnet

ClaudeCode (using Opus 4.7 and Sonnet) excels at brainstorming, planning, and creative exploration, but Nate finds Codex (using GPT-5.5) more pragmatic for execution: it follows longer plans better, troubleshoots issues ClaudeCode sometimes can't, and is more token-efficient. He uses both daily, leveraging each model's strengths rather than picking a single winner.

3 —

Setting Up a Codex Project: From Zero to Working Directory

Create a new project, set the working directory, and establish an agents.md onboarding file.

1

Create a new project In the Codex app, click «New Chat» and select «Add New Project.» Choose a local folder (e.g., Desktop/Codex YouTube/YouTube Analytics Demo). This folder becomes your working directory—all assets, code, and files will live here.

2

Set full access or default permissions In Settings → General, toggle «Full Access» to let Codex work autonomously, or leave on «Default Permissions» to manually approve network and file actions. Full access speeds workflows but requires trust; default is safer for beginners.

3

Create an agents.md file Ask Codex to create an agents.md file with context about you and the project goal. This markdown file acts as an onboarding doc: every new chat reads it first to understand your objectives, saving you from repeating context.

4

Familiarize Codex with your work Point Codex to existing folders (e.g., YouTube transcripts) and ask it to read a few files to learn your style and content. This builds contextual memory that persists across chats within the project.

4 —

Connecting to External APIs: YouTube Data Example

Use Google Cloud API keys and .env.local files to securely connect Codex to external services.

Codex has native plugins for many services (Slack, Google Drive, GitHub), but when no plugin exists—like YouTube Data API—you connect manually. Nate demonstrates creating a Google Cloud project, enabling the YouTube Data API v3, and generating an API key. The key is stored in a .env.local file, which Codex automatically excludes from public commits thanks to the dot prefix. Codex then tests the connection by fetching channel data and recent comments, troubleshooting TLS issues autonomously and documenting the fix in agents.md so the error never recurs. This pattern—connect, test, document—applies to any API: if Codex doesn't have a plugin, it will walk you through manual setup step-by-step.

5 —

Building Deliverables: From Raw Data to Excel Dashboards

📊

Data Collection

Codex fetches 200+ recent comments across multiple videos using the YouTube API, storing raw data locally and filtering for quality and relevance.

🔍

Pattern Analysis

It identifies comment types (questions, feedback, tool mentions), sentiment, and recurring themes—surfacing high-priority replies and content ideas for the creator.

📈

Visual Reporting

Codex generates a multi-tab Excel workbook with charts, priority scores, and actionable insights—making raw comment data immediately useful for content strategy.

🎨

Dashboard Design

Using GPT-Image-2, Codex creates UI mockups and a live Next.js dashboard (served on localhost) with tabs for overview, replies, ideas, and questions—complete with a custom logo.

6 —

Skills: Turning One-Off Builds into Reusable Recipes

Skills are markdown files that codify workflows so Codex can repeat them consistently.

After building the YouTube comment analysis flow, Nate asks Codex to «turn that into a skill.» Codex creates a .codex/skills/youtube-comment-insights.md file—a recipe with step-by-step instructions, endpoints, and output formats. Now, anytime Nate types /youtube-comment-insights, Codex executes the entire workflow without re-prompting. Skills live globally (usable across all Codex projects) or locally (scoped to one project). The magic: every time you use a skill, you can refine it based on new edge cases or failures, making your agent smarter over time. Think of skills as cookbooks: consistent results, easy iteration, and portable across projects. Nate emphasizes treating failures as «golden knowledge»—each bug is an opportunity to update the skill so it never happens again.

7 —

Deploying to Production: GitHub → Vercel Workflow

Push local code to GitHub, connect Vercel, and auto-deploy changes to a live URL.

1

Sync to GitHub Ask Codex to create a GitHub repo and push your local project. It initializes Git, excludes .env files, and authenticates via GitHub CLI. Your code is now version-controlled in a private repo.

2

Connect Vercel to GitHub In Vercel, click «Add New Project» and import your GitHub repo. Vercel reads the Next.js config and deploys automatically—no manual setup required.

3

Deploy and test Vercel generates a live URL (e.g., youtube-analytics-demo.vercel.app). Any changes pushed to GitHub trigger automatic redeployment, keeping local testing separate from production.

4

Iterate safely Test changes on localhost first. If you like them, push to GitHub; if not, discard. Production stays clean, and you maintain full version control.

8 —

Automations: Scheduled Workflows That Run Without You

Set up cron-like automations that refresh data and deploy changes on a schedule.

Nate creates a weekly automation: every Sunday at 5 PM, Codex runs the YouTube comment skill, updates the Excel workbook with new comments, merges rows without duplication, and pushes changes to GitHub—which triggers a Vercel redeploy. The automation injects a prompt into a new Codex chat, runs the agentic loop, and commits only the updated files. Nate stresses supervision at first: watch the automation run, catch roadblocks early (like open files blocking writes), and refine the prompt or model settings. He also warns that local automations require the Codex app to stay open—true 24/7 execution requires cloud hosting (similar to ClaudeCode's new cloud routines). The lesson: start supervised, iterate based on failures, and gradually trust the automation as it proves reliable.

9 —

Browser Use: QA Testing and Web Automation

🖱️

Visual QA Testing

Ask Codex to «stress test» your dashboard with browser use. It clicks tabs, filters, and links—documenting bugs like broken external URLs or awkward UI states.

🔐

Authenticated Automation

Browser use can log into sites (e.g., School, admin panels) and automate tasks like downloading reports or changing settings—bypassing the need for APIs.

🧪

Pre-Delivery Checks

Build browser QA into skills: before returning a deliverable, Codex runs visual and functional tests, catching issues before you ever see the output.

10 —

Model Selection and Token Management

Use medium for planning, high for complex builds, and avoid extra-high unless stuck.

💡

Model Selection and Token Management

Nate rarely uses extra-high intelligence—it over-engineers simple tasks and burns tokens fast. Medium handles most planning and execution; high is for large builds or tough bugs. Check Settings → Rate Limits to monitor session usage (resets every 5 hours and weekly). GPT-5.5 is highly token-efficient compared to Opus, making Codex sessions last longer. Use plan mode (slash command) to brainstorm without executing, saving tokens and preventing premature builds.

11 —

Context Management and Project Structure

📄

agents.md

The onboarding doc Codex reads first. Keep it concise—if it grows too large, move detailed knowledge into separate files or skill docs to avoid burning context.

🔒

.env.local

Store API keys here. The dot prefix tells Codex (and Git) to exclude it from commits, keeping secrets safe. Codex will never push .env files to public repos.

📁

Portable Projects

Every Codex project is just a folder. You can open it in ClaudeCode, Cursor, or any agent by adjusting file names (claw.md ↔ agents.md). Mix tools freely.

🧹

Ask for Cleanup

If your directory feels messy, ask Codex to reorganize or create a navigation doc. It can move files, rename folders, and explain the structure—keeping you oriented.

12 —

Key Numbers and Session Stats

Codex has generous limits, efficient token usage, and transparent session tracking.

Context Window

~250,000 tokens (GPT-5.5)

About 1/4 of Claude Opus's 1M context, but auto-compacts and is highly token-efficient in practice.

Session Limits

Resets every 5 hours + weekly

Check Settings → Rate Limits. Nate typically uses 97–99% of his session even after heavy builds, thanks to GPT-5.5 efficiency.

Subscription Cost

$20/month (Plus) or Pro tier

Free tier offers limited Codex access. $20 unlocks full usage and also works with OpenClaw and Hermes Agent.

Comments Analyzed (Demo)

208 total (200 initial + 8 refresh)

Codex automatically avoided duplicates when re-running the YouTube skill, merging only new comments into the Excel sheet.

13 —

Люди

Nate Herc

YouTube Creator / AI Automation Educator

host

Глоссарий

MCP servers / PluginsPre-built connectors that let Codex integrate with external tools (GitHub, Slack, Vercel, etc.) via simple authentication—no manual API setup required.

Plan modeA slash command that makes Codex brainstorm and outline a solution without executing code, saving tokens and letting you approve the approach first.

Full access vs. default permissionsFull access lets Codex work autonomously (network calls, file writes, etc.); default requires manual approval for each action—safer for beginners, faster for experienced users.

Context rotWhen an agent's working memory fills with outdated or conflicting instructions, causing degraded performance—prevented by context management and periodic memory refresh.

Dark codeCode written by an agent that you don't fully understand; while you don't need to read every line, you should understand what it does and why, to catch inefficiencies or risks.

Отказ от ответственности: Это ИИ-сводка видео с YouTube, подготовленная в образовательных и справочных целях. Она не является инвестиционной, финансовой или юридической консультацией. Всегда проверяйте информацию по первоисточникам перед принятием решений. TubeReads не связан с автором контента.