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

Tech concepts / product features covered

• Codex as an “AI super app” with local project automation

  • Unlike web chat, Codex works inside projects tied to a local directory on your computer.
  • It can operate on your files (including Excel), read local assets, and even control the mouse/keyboard and automate browsers.

• Workspace structure: projects, chats, and context

  • UI overview: projects + chats panels, model selection, and settings like speed and “intelligence” level (low / medium / high / extra high).
  • Emphasis on the interface not being overwhelming and learning settings early.
  • Mentions context window management (a status/context bar; Codex auto-compacts like other tools).

• Models and operational knobs

  • Switching models (examples mentioned: GPT 5.5 / 5.4 / other, plus “extra high” etc.).
  • Guidance: avoid extra high for simple tasks to reduce over-engineering risk and higher token cost.
  • Token/rate limits shown via “rate limits remaining” and session reset cadence.

• Agents onboarding file: agents.md

  • Codex uses agents.md (compared to Claude/Cloud Code’s claw.md) as an onboarding/context document.
  • Codex reads it to understand user identity, project goal, and product direction.
  • Also discusses keeping agents.md from becoming too large (token cost).

• Plugins / connectors / MCP-like integration

  • Codex supports a plugin-like interface for connecting external services (e.g., Versel, GitHub, design tools, productivity tools, etc.).
  • Example: no default YouTube plugin, so Codex asks how to connect:
    • API key vs OAuth vs both, and which comment set to analyze.

• OAuth/API setup workflow with Codex-guided plans

  • Codex can generate a plan, then guide you through implementation steps:
    • Create a Google Cloud project
    • Enable YouTube Data API v3
    • Create an API key with restrictions
    • Store secrets in env.local so they’re excluded from public commits
  • Codex tests connectivity and data retrieval; notes a PowerShell TLS issue but Node/Python work.

• Permission modes: Default vs Full access

  • Codex may ask for approvals under default permissions.
  • A Full access / auto-review option allows more autonomous actions (with safety cautions).
  • Advice: start cautious, then move to full access after you trust/understand the automation.

• Plan mode

  • Used before execution to brainstorm and structure tasks.
  • Plan mode prevents actual changes while outlining deliverables and execution steps.

• Slash commands, tagging, side chats

  • Uses / to invoke built-in capabilities (e.g., plan mode, browser use, memories, MCP, etc.).
  • Uses @ to tag specific plugins/files.
  • Side chat answers relevant questions without interrupting the main automation.
  • Mentions a “pet” UI widget that shows what Codex is working on; can change the pet appearance in settings.

---

Tutorial flow: YouTube Comment Intelligence System (0 → deployed dashboard)

1) Create project + ingest context (local files)

• Create a new Codex project directory (e.g., “Codeex YouTube/YouTube Analytics Demo”).
• Load context by prompting Codex to find and read YouTube transcript files from a local folder (demonstrates local searching).
• Warning: searching can be token-inefficient; better to provide exact folder paths.

2) Write agents.md and define end goal

• Codex creates/updates agents.md as onboarding.
• Project goal:
  • pull YouTube comments
  • analyze them
  • generate an Excel workbook
  • build a web dashboard hosted online

3) Connect to YouTube Data API and fetch comments

• Codex guides the connection using:
  • Google Cloud API key setup (YouTube Data API v3)
  • Store API key in env.local
• Tests:
  • works for channel data
  • works for comment retrieval from recent videos
• Captures failure knowledge:
  • Codex is asked to record the TLS issue so it won’t recur
  • warns that storing too much in agents.md can become costly

4) Plan + execute Excel analytics deliverable

• In plan mode, Codex structures an Excel deliverable:
  • pull ~200 recent comments across recent videos
  • analyze patterns, classify topics, and generate visualizations
• Execution outputs:
  • Excel file created under outputs/…/YouTube comment insights
  • tabs include:
    • “creator insights”
    • “frequent questions”
    • “reply opportunities”
    • “raw comments”
  • visual breakdowns such as:
    • comment category mix
    • “top mentioned tools”

5) Turn the workflow into a reusable Skill

• Defines a skill as a repeatable recipe.
• Codex converts the workflow into a global skill file:
  • invoked via / + skill name (example: /YouTube comment insights)
• Skills can be global (available across projects) or local to a single project.

6) Build a dashboard UI using the Excel output

• New plan: create a localhost dashboard with UI, charts, and visuals.
• Uses GPT Image 2 for dashboard UI/logo concepts.
• Codex verification loops:
  • automated browser/server verification passed
  • visual scan found issues → fix → re-verify with screenshots
• Dashboard features:
  • tabs like overview/replies/ideas/questions/explorer
  • interactivity suggests comment links (some may not be “synced” immediately)
  • charts + “AI insights” such as reply priority and common themes

7) Deploy from localhost → public URL (GitHub + Vercel)

• Deployment approach:
  • sync Codex project files to a GitHub repository
  • connect/import that repo into Vercel and deploy
• Safety implementation:
  • avoid committing secrets; keep keys in env.local (dotfile excluded from public commit)
• Result:
  • dashboard available at a Vercel domain (optionally customizable with DNS)

8) Create a scheduled Automation (weekly refresh)

• Schedule: every Sunday at 5:00 p.m.
  • re-run the YouTube comment insights skill
  • refresh Excel data (merge/append while avoiding duplicates)
  • update dashboard code/output
  • commit and push to GitHub so Vercel redeploys
• Notes on runtime:
  • behaves like a local cron tied to the Codex app machine
  • for always-on, suggests using cloud scheduling (compared to Cloud Code’s cloud routines)

9) QA with browser automation (“browser use”)

• Codex uses browser use to stress-test the dashboard UI:
  • click around tabs/filters and try to “break” the app
  • detect issues like external YouTube links not opening visibly
• Then generates fixes via a new plan/execution cycle.

10) Iterative improvements + handling automation failures

• Two setbacks described:
  • dashboard/Excel file was open → Codex couldn’t overwrite → delayed run
  • automation configured with the wrong model (GPT 5.2 instead of desired GPT 5.5) → slower execution
• Fixes:
  • re-trigger automation with corrected model + reasoning settings
• Demonstrates incremental refresh:
  • after refresh, the dashboard reflects only new additional comments (deduping occurred)

11) Optimizing automation (“dark code” concept)

• After automation runs, Codex is asked to:
  • explain what it did
  • identify streamlining opportunities
  • recommend guardrails to make it safer and less “heavy”
• Mentions “dark code” as the risk of autogenerated code you don’t fully understand.

---

Key “review/guide/tutorial” takeaways explicitly highlighted

• Interface guide: get familiar with Codex UI, model toggles, permissions, and plan mode.
• Integration guide: connect YouTube via YouTube Data API v3, use an API key, and store secrets in env.local.
• Skill-building guide: convert a one-off workflow into a reusable skill (global or project-local).
• Deployment guide: localhost dashboard → GitHub + Vercel continuous deployment.
• Automation guide: scheduled weekly refresh, token/session considerations, and local-cron limitations.
• QA guide: use browser use to run UI tests and identify bugs before redeploying.

---

Main speakers / sources

• Primary speaker: the channel creator narrating the tutorial (no name given in subtitles, but refers to himself as “Nate Herk” regarding the project/channel content).
• Referenced systems/tools (not speakers): Codex, ChatGPT models, Claude/Claude Code, GitHub, Vercel, Google Cloud (YouTube Data API v3).