I’ve been building with every major coding agent for months. The pendulum swings constantly. This week it swung hard toward OpenAI.
GPT 5.4-codex dropped. And it’s not close for me. It’s the best coding model I’ve used and brings a consistency I can’t ignore.
However, I don’t think it matters anymore which model is on top. It’s what you should be building that works on top of the AI models. Let’s get into it.
GPT 5.4: The My Favorite Model Right Now
I’ve switched most of my personal AI spending to ChatGPT.
That’s not something I say lightly. I’ve been deep in the Anthropic ecosystem. But GPT 5.4-codex has convinced me to go all in on the codex ecosystem It’s not just a better coding model — OpenAI upgraded reasoning and personality. The whole product feels sharper.
Where it really shines is coordinating with the Codex desktop app. There’s been a lot of buzz about the app being better than the CLI for Codex. It’s true. I’ve never been happier using a coding tool, it feels like an absolute joy to use.
I’ve also switched my OpenClaw Assistant Hermy to entirely using the OpenAI stack. While I haven’t seen massive improvements there with tasks that we’re having Hermy do (I’ll be writing and updating about here soon), it’s certainly cheaper to run this on a GPT pro account.
What blew me away about Codex 5.4
Playwright testing out of the box. Once you set it up, Codex automatically verifies your app with Playwright. I haven’t seen it make a mistake. It has spatial reasoning with design elements. It thoroughly tests your app without being asked. Just out of the box, it really views the website both like a human would and all the benefits as system has at viewing it.
Web browsing that actually works. Point it at any website. “How are they doing this animation?” 5.4 tells you. Then it builds a demo. Then it integrates it into your project. I’ve never had a model handle this workflow so smoothly.
Reverse engineering bundlers. I’ve had it reverse engineer front-end bundlers and preformance optimizations. Not toy examples. Real production bundlers. It just does it without asking.
If you’re using AI to code every day and still searching for a tool I highly recommend starting with desktop Codex. CLI Codex is also incredible and open source and is the bedrock of the desktop tool.
The pendulum always swings
We’ve been the new pattern. The best model changes every few months. Anthropic ships something great. Then OpenAI leapfrogs. Then Gemini does something very interesting in the google ecosystem.
This is the AI race. It’s not going to stop.
So the question isn’t “which model should I use?”. I think that question really doesn’t matter anymore, even the low reasoning models now are competitent at doing kinds of work.
The question is: what can you build once that works everywhere?
Skills: The Best Investment You Can Make Right Now
I’ve been saying this for 3 weeks, but I want to go deeper today. Skills are the highest-leverage thing you can spend your time on if you’re working with AI coding agents.
What are skills?
Skills are prompt-based extensions that teach an AI coding agent how to do something. You write a SKILL.md file with instructions. The agent picks it up and adds it to its toolkit.
No code required. No API integration. Just clear written instructions that tell the agent how to handle a specific workflow.
Anthropic has the best documentation on this in my opinion.
Skills follow the open Agent Skills standard. That last part matters.
Skills are now standard across all LLM tools
Codex has skills. Claude Code has skills. This pattern is everywhere now.
That means the skills you write are portable. You build a skill for how your team deploys. You build a skill for how you structure API endpoints. You build a skill for your testing patterns. Those skills travel with you to whatever provider has the best model next month.
Codex makes skills deeply embedded and really intutive in the experience. Claude Code supports them too, though it’s not as straightforward to configure Is just as powerful.
I think excellent examples of this are actually within some of the Codex app’s defaults. You can just open the hood and see what the skills are made of. It’s really interesting to see how they’re integrating code, MCPs within these skills, and how they all orchestrate together into new LLM capabilities. Skills repo
This is where you should be spending your time. Not chasing the model of the month. Building the instructions that make any model work the way you need it to.
Skills vs MCPs — know the difference
People mix these up constantly. They’re complementary but different.
Skills are instructions. They teach the agent how to do something. “We deploy our app this way” “We write database migrations with this example.” “Here’s our code review checklist.” It’s knowledge and workflow captured in markdown. No external dependencies.
MCPs are integrations. They give the agent access to something. A database connection. A GitHub API. A Slack channel. They’re plumbing. You should be thinking of these more like an API but with some abstractions.
You need both. But skills are where your expertise lives. MCPs are commodities and integrations. Your deployment skill that has instructions on how to self heal and suggest common fixes and optimizations is going to save you more time and something that you, as a developer have agency over.
My take: the goal of a software engineer right now
Our job is now organizing agents for victory.
That sounds dramatic. I’m not sure it is. Think about what you actually do when you’re productive with AI tools. You’re not writing code. You’re setting context. You’re defining how the agent should approach a problem.
The repeatable version of that is a skill.
You still need to push code and provide value with what you’re doing. But you should be spending more time optimizing your agent setup when it makes sense than grinding code across the finish line. The new leverage is in the setup, not always the execution.
Every time you find yourself giving the same instructions to an agent, how to structure a PR, how to handle error logging, how to write tests for your specific stack — that’s a skill waiting to be written. An great hack is when you’re done doing something with an AI, you can just ask it to help build the skill based on what you just did. From our previous article: We know that those skills are not always as good as one that is completely human-written; however, it’s better than not having a skill.
I think writing agent skills is so important. I’m likely just to devote part of this weekly newsletter just on skills. I think this is thing to get great at this year as our ecosystem and profession evolves.
What I’m thinking about for this next week
How do you go from app idea to totally autonomous agents working on it consistently? There’s folks on the Twitterverse claiming to do this, and I think some are. I’ve heard of Kelly Claude that is just autonomously making kind of sloppy mobile applications, but how do you do this well for things that you want? I think we’re really not that far from this.
And critically how do you check in so your code doesn’t go haywire and you’re not burning thousands of dollars in tokens?
What does it look like when a company is really just a series of AI agents? With 5.4, that feels more possible than ever.
I don’t have the full answer yet.
Please stay in the loop with your AI. Go deep on your tools. And build your skills library it’s the one thing that compounds no matter which model is on top.
We’re going to be including Pangram Labs AI Detector at the bottom of these article. I actually think it’s important to show people that you’re writing things and it’s not an AI doing it. The AI models are really good at it now and I feel linkedIn is full of AI messaging. However, you might see some AI assistance in this time to time. I think it would be unwise to be communcating this and not having AI help you at all.
See you next week.
Pangram Labs Results
