Экстракт Знаний

All the Codex CLI News (This Week!)

9:36

All the Codex CLI News (This Week!)

Ray Amjad 25.09.2025 8 621 просмотров 177 лайков обн. 18.02.2026

Смотреть на YouTube

Поделиться Telegram VK Бот

Транскрипт Скачать .md

Анализ с AI

Описание видео

Join AI Startup School & learn to vibe code and get paying customers for your apps ⤵️ https://www.skool.com/ai-startup-school —— MY APPS —— 🎙️HyperWhisper, write 5x faster with your voice: https://www.hyperwhisper.com/ - Use coupon code 772DYJYF for 40% off 💬 MindDeck, an advanced frontend for LLMs: https://minddeck.ai/ - Use coupon code OP6CZ8P7 for 40% off 📲 Tensor AI: Never Miss the AI News - on iOS: https://apps.apple.com/us/app/ai-news-tensor-ai/id6746403746 - on Android: https://play.google.com/store/apps/details?id=app.tensorai.tensorai - 100% FREE —— MY CLASSES —— 👾 Codex CLI Masterclass: https://www.mastercodexcli.com/ - Use coupon code K5LP2NRK for 20% off 🚀 Claude Code Masterclass: https://www.masterclaudecode.com/ - Use coupon code 6OKODFRW for 20% off ————— CONNECT WITH ME 📸 Instagram: https://www.instagram.com/theramjad/ 🐦 X: https://x.com/@theramjad 👨‍💻 LinkedIn: https://www.linkedin.com/in/rayamjad/ 🌍 My website/blog: https://www.rayamjad.com/ ————— Links: Codex CLI Just Fixed Claude Code: https://www.youtube.com/watch?v=GJzfNWK4iHg&vl=en Codex All Advanced Features: https://www.youtube.com/watch?v=KUul2bYAIHo Scale AI Benchmark: https://scale.com/leaderboard/swe_bench_pro_public How OpenAI Uses Codex: https://cdn.openai.com/pdf/6a2631dc-783e-479b-b1a4-af0cfbd38630/how-openai-uses-codex.pdf GPT-5-Codex Prompt Guide: https://cookbook.openai.com/examples/gpt-5-codex_prompting_guide Codex Prompting Guide: https://developers.openai.com/codex/prompting Meta’s Research: https://ai.meta.com/research/publications/are-scaling-up-agent-environments-and-evaluations/ Meta’s Research Results 1: https://huggingface.co/blog/gaia2 Meta’s Research Results 2: https://huggingface.co/spaces/meta-agents-research-environments/leaderboard Timestamps: 00:00 - Intro 00:16 - Usage Limits 00:40 - /review 01:34 - Auto Compaction 01:47 - Output Schema 02:20 - MCP Tool Timeout 03:03 - Codex Login 03:45 - New GPT-5-Codex 04:35 - Scale AI Benchmark 05:46 - How OpenAI Uses Codex 07:06 - GPT-5-Codex Prompting Guide 08:08 - Codex Prompting Guide 08:16 - Meta's Benchmarks

Оглавление (13 сегментов)

0:00 Intro 69 сл.
0:16 Usage Limits 91 сл.
0:40 /review 204 сл.
1:34 Auto Compaction 39 сл.
1:47 Output Schema 139 сл.
2:20 MCP Tool Timeout 154 сл.
3:03 Codex Login 163 сл.
3:45 New GPT-5-Codex 220 сл.
4:35 Scale AI Benchmark 237 сл.
5:46 How OpenAI Uses Codex 263 сл.
7:06 GPT-5-Codex Prompting Guide 201 сл.
8:08 Codex Prompting Guide 31 сл.
8:16 Meta's Benchmarks 310 сл.

Intro

Okay, so Codex CLI released a bunch of brand new updates over the last few days and we're going to be going over some of them. Firstly, you want to make sure you're on the right version by doing codex --version here and you should see that you're on 0. 41. If you're not, then you can run this command and then that should update to the latest version.

Usage Limits

Anyways, the biggest thing that I like about this update is that you can see your usage limits if you do /status. So if you run codex, then allow, then /status, then you can see it says send a message to load the usage data. Say hi. And you can see right over here I have my weekly limits and my five hour limit. So I think my weekly limit just reset, which is pretty good. And I find the usage limits to be quite generous at this stage at least.

/review

Something else I also added in this update is this new /review command. So if I do /review, then you can see I have a couple options here. I can review the uncommitted changes which will trigger a review of any changes that just haven't been committed yet in git. I can also do /review of a particular commit that had previously happened. As you can see right over here I can do /review against a different branch. So maybe before doing a PR or something I can do a /review and then finally with some custom instructions such as saying like can you check the logic for X, Y and Z is like implemented correctly? Or what do you think of A, B and C for example? And I have made a previous video over here called Codex CLI just fixed cloud code that you can watch using the link down below. And basically what I was doing in this video is getting Claude Code to make a bunch of changes and edits to a codebase. And then I got Codex CLI to review those changes. And I think that process is now even easier if you do something like review uncommitted changes.

Auto Compaction

They also have auto compaction triggered automatically if you're using gpt-5-codex when you hit around 222,000 tokens. And you can see the limit is hard coded right over here. But you can also auto compact manually by doing /compact.

Output Schema

Something else they also added is that in exec mode you can now use output schemas. So for example, if you have a schema that kind of looks like this, where you have like name specifications, use cases, for example, then you can run exec mode by quitting Codex CLI and running a command kind of like this, where I'm running GPT-5 with output schema as a. json file that I just defined earlier. And then I can pass in an instructional command such as "What is this project about". Press enter and then the response that this will give will be in the schema format that I defined over here in this file and you can see it's right over here. Bear in mind this only happens in exec mode, so you want to make sure you're running

MCP Tool Timeout

exec mode for that. Something else I also added is a tool_timeout_sec over here. So if you go to your. codex folder, so it should be. codex and then open up the config, this is a file in which you define any MCP servers that you want. If you want to learn more about setting up MCP servers, then I have another video linked down below called Codex CLI: All the Advanced Features and I basically cover setting up MCP servers here. But anyway, they added a new thing which is tool_timeout_sec, which basically means that if a tool does not respond within this certain time then the operation will be cancelled. And this can be useful for MCP servers which like do pretty heavy tasks and you need longer timeouts for. We have the /review command in 0. 39 that we covered earlier and more features were added to. And a breaking change

Codex Login

here is that previously Codex used to read your OpenAI API key from the environment. So you would have to define it such as this by doing export OPENAI_API_KEY and then your key right over here. But now that is no longer the case. So if you do want to log in with your API key, firstly you want to log out of your ChatGPT session if you're using ChatGPT and then do codex login --api-key and then define your key just like this. And this API key will then be stored in your auth. json file. So you can see this file over here where the API key is now stored and then you can run Codex as normal using the API key instead. And I think the same thing also applies if you want to use a different model provider with Codex. And I do cover that in my previous video right over here, which you can watch again. Of course

New GPT-5-Codex

they added their brand new model which is GPT-5-Codex last week and they now have made that default model. So if you want to switch away from that model, then you would do /model. And I've been using the model GPT-5-Codex a lot over the last week and I have found it to be pretty amazing. You can see right over here I have this like really long session running where I'm adding a brand new feature to my application, HyperWhisper, and I've used about 2. 44 million tokens so far and often Codex is just able to run independently for about 30 to 40 minutes. But of course I still have to test the application, give errors back to it, and so forth. But it's kind of nice that I can just like leave it running for a while and then trust that it will still get like 90, 95% of the way there. And if you're interested in what this application is, it's a speech to text application. There's a link down below alongside a coupon code and it's basically the most customizable one on the market right now. So do try it out if you have time and do email me with any feedback that you have because I'm always looking for ways to improve it. Anyways,

Scale AI Benchmark

for some Codex/GPT-5 related news, Scale AI released a brand new benchmark called SWE-Bench Pro and you can see a performance comparison over here. And basically the reason is because I guess SWE-Bench Verified that a lot of model providers were testing themselves against such as this one. Over here we're getting really good scores. So for example like 4 months ago Claude 4 Sonnet got 80. 2% and a lot of them are converging on pretty similar scores. So of course when the benchmarks do get maxed out then we need a new and harder benchmarks. So you can see that many of the models over here have fallen from SWE-Bench Verified to SWE-Bench Pro. But GPT-5 and Opus 4. 1 are both like pretty neck on neck kinda within this margin of error. But of course the GPT-5 is significantly cheaper than Opus 4. 1. One of the things they do mention about the benchmark is that the tasks are more diverse compared to previous ones. So there are problems from consumer facing applications, B2B platforms and developer tools requiring reasoning across varied architecture and development patterns. So it will be pretty interesting to see if this benchmark does get adopted by model providers when they like make their model announcements or releases. And how long before this benchmark is also maxed out as well. OpenAI released a brand new doc called How OpenAI uses

How OpenAI Uses Codex

Codex. There will be a link down below and it kind of reminds me of the previous thing that Anthropic released which is How Anthropic teams use Claude Code about two months ago. So there's like a lot of deja vu happening right now, but anyway, so you can read through the documentation, but like Use case 1 is Code understanding. So, Codex helps our teams get up to speed quickly in unfamiliar parts of the codebase when onboarding, debugging and investigating an incident. During incident response, Codex helps engineers ramp into new areas of the code quickly by surfacing interactions between components and tracing how failure states propagate across systems. And then one of the site reliability engineer's gave a quote over here that it basically helps him jump straight to the right files so I can triage fast. And they also give you some recommended prompts that you can use for this use case, such as like where is the authentication logic implemented in the repo? And then a couple other prompts as well. And then there's information about more use cases like refactoring and migrations, performance optimization, improving test coverage, increasing development velocity, staying in flow. So for example, a prompt could look like Summarize this file so I can pick up where I left off tomorrow, exploration and ideation. And then there's a bunch of best practices right at the bottom over here that you can read through. I may make a separate video about it, but I'm not exactly sure I don't want this to be too long.

GPT-5-Codex Prompting Guide

They also released a GPT-5-Codex Prompting Guide and this is meant for users of the API who are creating developer focused prompts, not for Codex users. So for example, if you're using the Codex CLI, this won't be for you, but if you're making like a vibe coding platform or something like that, then this can be useful. It is quite interesting that the general prompting principle for GPT-5-Codex is "less is more. " So start out with a minimal prompt inspired by the Codex CLI system prompt, then only add the essential guidance you truly need. Remove any prompting for preambles, because the model does not support them. And that includes stuff like saying oh, you are an expert typescript engineer who specializes in blah blah like you should not need to have them anymore. Reduce the number of tools available, and then make tool descriptions as concise as possible. So overall it seems that for GPT-5-Codex you don't have to prompt as much as previously, because for GPT-5, for example, the system prompt for Codex CLI was about 310 lines, whereas a new system prompt for GPT-5-Codex is 104 lines. So it's about one-third the length, for better performance.

Codex Prompting Guide

And if you're interested in prompting Codex in general and not the API GPT-5-Codex, then there's a separate guide right over here that should be linked down below. And Meta also

Meta's Benchmarks

did release a new agentic benchmark a couple of days ago and you can see that GPT-5 (high) performs really well across all these different categories. So when it comes to execution, search, ambiguity and adaptability and you can read the full paper to understand what all of these mean and then also run the agent environment yourself to test different models too. And I guess that because GPT-5-Codex is a variation of GPT-5 it also explains why Codex CLI has suddenly become really good. And yeah, you can see the leaderboard here with all their different scores and unfortunately they don't seem to have Claude 4 Opus here. But I imagine based on other benchmarks it should be pretty similar to GPT-5. But of course it's much more expensive than GPT-5 though. But yeah, overall this space seems pretty exciting. I think it will push Anthropic to release their like 4. 5 or 5 model pretty soon to basically catch up with GPT-5 and also win back many of the customers who are switching over from Claude Code to Codex CLI. Anyways, I will be making more videos about Codex-CLI-related news, especially as they add more features. So do subscribe to the channel if you do want to see more of that kind of stuff. Anyways, if you do want to improve your vibe coding and vibe marketing skills and I do cover a lot to do with that in my community over here, a bunch of people have been part of it for a while and have seen pretty great success with their own mobile applications and web applications as well. The biggest value added for you may be that you have personal help from me for whatever you may be stuck with. So for those of you who are interested there will be a link down below.

Ещё от Ray Amjad

How to Make Anki Flashcards 10x Faster with AI (for free!)

Ray Amjad | 23.04.2025 | 6 сегм. | 267 982

Staying Organised for A-levels // Cambridge Student

Ray Amjad | 05.09.2021 | 11 сегм. | 117 382

How I Got an A* in Physics A-level (Cambridge Student)

Ray Amjad | 28.09.2021 | 14 сегм. | 68 131

Mistakes I Made in Sixth Form

Ray Amjad | 17.09.2021 | 10 сегм. | 68 002

Brutally Honest Replit Review (Not Sponsored)

Ray Amjad | 12.09.2025 | 16 сегм. | 49 405

How I Got an A* in Further Maths A-level (Cambridge Student)

Ray Amjad | 02.10.2021 | 14 сегм. | 47 938

Ctrl+V

Экстракт Знаний в Telegram

Транскрипты, идеи, методички — всё самое полезное из лучших YouTube-каналов.

Подписаться