Stop Just Using Codex CLI to Generate Code

You might not believe this.

Just 20 days ago, I first heard about Codex CLI. Back then, it was just “an AI coding assistant that writes code”—pretty much like all those other tools out there: Copilot, Cursor, you name it. You give it a requirement, it spits out some code.

Then what?

Then it started updating like crazy. From January 16th to February 5th—20 days, 12 versions. I watched this OpenAI-built AI coding assistant transform from “an intern who can write code” into “a fully autonomous assistant that thinks independently and can rally a team to get things done.” My developer productivity literally doubled.

Honestly, this update pace reminds me of when WeChat launched its red envelope feature—while you’re still figuring out how to use it, they’ve already shipped three more versions.

Today, I want to walk you through the most practical new features across these 12 versions. You don’t need to read everything, but by the end, you’ll at least be able to explain: What makes Codex CLI different? Why is it worth installing a new tool to boost your developer productivity?

From “Solo Warrior” to “Team Player”: The Evolution of Plan Mode

Analogy: Leading a Team vs. Working Alone

Imagine you’re renovating a house.

Working alone: You tell a worker “go paint the living room walls white.” Worker finishes, then you say “go lay the bedroom floor.” Worker finishes, then you say “go install the kitchen cabinets”… By the end of the day, you’ve done nothing but give orders.

Leading a team: You tell the foreman “I want to renovate a three-bedroom apartment.” The foreman handles the rest: painters do walls first, carpenters lay floors next, cabinet installers come in last. You just need to confirm direction at key checkpoints.

Codex CLI’s Plan mode is that “foreman.”

Evolution Timeline

Let me draw you a timeline so you can see how much attention this feature got:

  • v0.88 (Jan 19): Started building the collaborative mode infrastructure. Still “internal testing”—regular users couldn’t access it.
  • v0.90 (Jan 25): Beta released, finally available to everyone. The main agent can dispatch research agents to gather info, then coding agents to write code.
  • v0.91 (Jan 25): Same day release? Yep, that’s the iteration speed. Sub-agent count cut from 12 to 6—preventing it from going wild and dispatching ten minions for one task, accomplishing nothing.
  • v0.93 (Jan 31): Plan generation became a “streaming view.” Before, you’d submit a task and it would think silently, then give you a complete plan. Now? You watch it think step by step, adjusting direction in real-time.
  • v0.94 (Feb 2): Plan mode became the default. New users now see “Tell me what you want to accomplish, and I’ll help you make a plan” as the onboarding message.
  • v0.95 (Feb 4): Finally supports images! You can drop a design mockup screenshot: “Build me a login page that looks like this.” It understands.

When v0.95’s image feature dropped, I tried throwing in a hand-drawn flowchart. It generated an annotated state machine diagram… I thought to myself, if this had existed two years earlier, I could’ve saved three all-nighters on my college “Software Engineering” final project.

When to Use Plan Mode?

My personal rule: When a task takes more than 2 hours, involves multiple steps, or requires research—go straight to Plan mode. Developer productivity goes through the roof.

For example: “Build me a complete user registration flow, backend in Go, frontend in React, PostgreSQL database, with email verification”—throw this into Plan mode, it plans the steps itself, you just confirm direction.

If it’s just “write me a function to remove duplicates from an array”—something that takes 30 seconds—stick with the default chat mode.

Skills System: This Is the Real “Customizable AI Coding Assistant”

Analogy: Phone Apps vs. Built-in Features

Codex CLI itself is a CLI tool, but Skills turn it into a platform.

Think of Skills as “installing plugins for Codex CLI.” It’s not a fixed feature list—it’s an infinitely expandable capability pool.

From “Can Do Work” to “Knows How to Work”

  • v0.86 (Jan 16): Introduced SKILL.toml metadata files. Before this, a skill was just a prompt.txt—no name, no description, no icon. Now with metadata, skills have their own names, icons, colors, and descriptions. The TUI interface shows everything clearly.
  • v0.94 (Feb 2): Supports loading skills from project directory .agents/skills. What does this mean? Your project can have dedicated skills—team members pull the code, skills auto-load. No individual configuration needed.
  • v0.95 (Feb 4): Supports installing skills from remote Git repos. Someone in the community wrote a “database migration assistant”? One command to install: codex skills install https://github.com/xxx/db-migrator.git.
  • v0.97 (Feb 5): Real-time hot reload. Edit prompt.md and save—Codex CLI detects the change and auto-reloads the skill. No app restart, no service restart. Edit AI prompts like editing code.

A Real Example

My team has a data science project that processes raw data weekly. We wrote a data_prep skill:

# SKILL.toml
name = "Data Preprocessing"
description = "Clean and preprocess raw data files"
icon = "📊"

# prompt.md
You are a data preprocessing assistant. When invoked, you will:
1. Read the raw data file specified by the user
2. Fill missing values in numeric columns with median, categorical columns with mode
3. Remove duplicate rows
4. Normalize numeric columns to 0-1 range
5. Save cleaned data with "_cleaned" suffix in filename

Drop this skill into the project’s .agents/skills/data_prep directory—anyone on the team running Codex CLI sees this skill automatically. Others want to use it? Just pull the code.

We used to write documentation, hold training meetings, write scripts for this workflow. Now? One skill file handles everything. Team developer productivity takes off.

Security & Permissions: Not “Can It Do This” but “Ask Before Doing”

Analogy: Going Out Without Money vs. Without ID

Codex CLI, as an AI coding assistant, is essentially a program that can operate your computer. It can read your files, execute commands, access the network—without some limits, it’d be way bolder than you.

So from the start, Codex CLI put security at its core.

Permission System Evolution

  • v0.86 (Jan 16): Supports disabling web search. In sensitive environments where you don’t want AI searching online—one config line handles it.
  • v0.89 (Jan 22): Added /permissions command. /permissions allow all approves all operations—for when you trust it but don’t want to click confirm every time.
  • v0.90 (Jan 25): Network sandbox proxy. Precise control over “which domains can be accessed”—like only allowing GitHub, PyPI, and internal doc sites, nothing else.
  • v0.95 (Feb 4): Enhanced Git command security. Dangerous operations like git push --force and git reset --hard require explicit approval—no more auto-pass.
  • v0.97 (Feb 5): MCP tool approval memory. Click “allow and remember” once, similar tool calls never prompt again.

When to Pay Extra Attention to Security?

If you’re handling these scenarios, enable strict permission mode:

  • Company internal code with trade secrets
  • Production servers with write access
  • Shared development environments
  • Any project involving user data or payment info

My personal approach: Go wild in local development, switch to strict mode before hitting production.

“Use Once, Never Go Back” Productivity Features

Web Search Cache (v0.92)

Ask the same question twice, second time is instant. Because search results are cached by default.

This is perfect for “think-as-you-go” workflows. You ask a question, then remember a related one later—no need for AI to search again, just uses cached results.

Shell Alias Recognition (v0.87)

That k=kubectl in your .bashrc? Codex CLI recognizes it now.

The most painful thing about AI collaboration used to be: you know you have an alias, but AI doesn’t. You either write out the full command or explain what the alias means first.

Now? However you use your terminal, AI uses it the same way.

Thread Filtering & Recovery (v0.92)

/threads --filter "database connection pool" finds historical sessions. Found one that’s archived? thread/unarchive brings it back.

Used to spend ages scrolling through chat history. Now one command handles it.

macOS App Launcher (v0.95)

codex app /path/to/my-project registers an app—then Spotlight search by app name launches directly.

I tried it, registered four frequently-used projects. Now opening projects doesn’t need cd commands—just Cmd+Space, type project name, Enter, straight into project context.

Questions You Might Have

Q: How is Codex CLI different from Copilot and Cursor?

The biggest difference is “autonomy.”

Copilot is “you write a line of code, it completes half.” Cursor is “you describe requirements, it generates a file.” Codex CLI, this OpenAI-built AI coding assistant, is “you give a goal, it plans, researches, codes, tests, and commits on its own.”

It’s more like a “junior programmer assistant”—can complete tasks independently, but needs you to set direction, helping boost your developer productivity.

Q: How hard is this AI coding assistant to learn?

Honestly, harder than Copilot. Copilot is just install a plugin and go. Codex CLI requires installing the CLI, configuring the environment, learning new interaction patterns.

But if you spend 3+ hours daily coding, it’s worth investing 1-2 hours to learn this AI coding assistant. The time it saves will far exceed that investment.

Q: Which version should I start with?

Go with the latest. Currently v0.97, all features are stable.

Q: Is Plan mode too “heavy”?

No. Plan mode can exit to chat mode anytime. Throw in a task, if you think its plan is too complex, just say “too complex, start with something simple.”

Next Steps: From “Trying It Out” to “Using It Daily” for Better Developer Productivity

If you decide to try Codex CLI, here’s how I’d start:

  1. Install it: macOS uses brew install, Linux uses respective package managers
  2. Don’t mess with config yet: Default config works fine, just run codex
  3. Start with a small task: Like “write a Python script to merge two CSV files”
  4. Try Plan mode: Throw in a task that needs research, observe how it works
  5. Understand permission settings: Before using it on real projects, learn /permissions and /debug-config

20 days, 12 versions of iteration speed shows OpenAI is treating Codex CLI as a core product. Getting in now means everything you learn becomes a developer productivity advantage in the future.

Found this article useful?

  1. Like: If it helped, give it a like so more people see it
  2. Share: Pass it along to friends or colleagues who might need it
  3. Follow: Follow for more practical tech articles
  4. Comment: What AI coding tools are you using now? What pain points do you want solved? Want to boost your developer productivity?

Your support is my biggest motivation to keep creating!