Video Summary - Getting the most out of Codex

Summary of Technological Concepts & Product Features (Codex / Claude Code)

The creator previously shared “Claude Code” best practices and notes they generally apply to Codex as well.
Codex is used not only for coding, but also for broader agentic workflows.

On the Codex website, the first prominent option is a desktop app for Mac.
Compared to Claude Code’s desktop app (described as feeling like an “attachment”), Codex’s desktop app is:
- Visually clean
- Fast
- Well-designed
It provides a compact overview of uncommitted changes (rather than a file tree by default).
It supports adding tiny comments to changed files.
Key workflow detail: adding a comment becomes context for the next AI prompt, effectively combining code review + prompt guidance.

A Windows version is planned, but the creator suspects the delay is related to sandboxing.
They suggest the agent relies on native platform features and that the sandboxed approach “doesn’t seem to work on Windows yet.”

The creator avoids “looping” or autopilot-style behavior and emphasizes staying in control.
Plan mode can be triggered:
- In the desktop app via Shift+Tab
- In the command line via Shift+Tab
They stress explicit planning and explicit over implicit context.

While models can find relevant information, the creator recommends being explicit when you already know a file/topic matters.
Example approach:
- Search project docs
- Copy relevant content
- Expand your prompt with it
- (Without overdoing it for every doc page.)

Mentions using MCP (Model Context Protocol) servers, specifically Context 7, to provide a documentation search tool to the agent.
Desktop app setup:
- Settings → Add Server → add MCP details
Without the desktop app:
- Use the MCP page instructions to configure global or project-specific Codex settings.
For Context 7 basic usage, they claim no API key is required.
They also note tradeoffs:
- Making it explicit that MCP-enabled knowledge exists may improve relevance and reduce token usage
- But it may still involve subscription and/or usage limits

Codex “skills” are described as official, standardized agent extensions.
Skill structure:
- A folder containing a Skill MD file including:
  - a name
  - a description (most important for activation behavior)
  - additional skill instructions
Loading behavior:
- Only metadata (including the description) is loaded at the start of a session.
- The full skill content is read only when the agent decides to activate it.
Skills registry:
- skills.sh is mentioned as a place to find community skills (e.g., BetterOff usage).
Main point: constructing the right context and using skills effectively still matters; Codex is described as good at activating the right skills at the right time.

The creator encourages using agent tools but verifying outputs.
Codex includes a code review feature similar to Claude Code:
- Run /review
- Choose review targets, e.g. current branch vs base branch (PR-like comparison)
Still emphasized:
- Humans must review code
- Do not rely on the AI alone

The creator argues developers should not fully hand over control.
They stress:
- You can implement changes directly (e.g., “10 pixels of margin”)
- Avoid becoming “detached” from the codebase (which they say leads to “white coding” and loss of understanding)
- Even if agents reduce how much code you write, you should still be able to edit and understand the system

Mentions an Agents.md file concept common across coding agents (Claude Code / Codex).
Rules:
- Always loaded into every session context
- Keep it concise to avoid bloating prompts and degrading results
Include essentials such as:
- project stack
- key commands
- core code style/rules
- crucially, self-verification instructions
Self-verification suggestions:
- automated tests
- linting / type checking
- use Playwright for browser checks
Behavioral feedback loop:
- If the agent repeatedly makes mistakes, update Agents.md so the fixes apply in future sessions.

Primary speaker: The creator/author of the video (name not shown in subtitles), discussing using Codex alongside Claude Code and referencing their own course/videos.