I've used a similar technique with my own tools. It helps to save context overall, by deferring the detailed instructions not only to when they're needed, but also to where they're needed.
Simply loading the browser Skill alone, doesn't always mean a browser is actually necessary (or supported). The user might've want them to control their main desktop browser; or the user might be pointing them toward a read-only HTML file that doesn't need rendering; or the sandbox might prevent network access; etc...
Codex is especially 'keen' to use Skills (especially by comparison to Claude), I often see it load Skills which it doesn't go on to use, because the main Skill content says "when not to use this". So I'm sure this is a product design decision to help with those kinds of circumstances.
It's runtime dependent on the browser being used! A skill file wouldn't be able to have this dynamically injected for the browser that's about to be used.
Skill file could indicate general instructions for all browsers. This provides guidance for the browser that's about to actually be used.
Honestly, this is one of the things OpenAI does pretty well, I think, these tiny design decisions that you wouldn't notice until it becomes super annoying.
That's a pattern I often follow - refer it to the docs, don't include them. Usually it's another markdown, but sometimes it can be the output of something else. This avoids including any of that in context before it's actually needed.
Simon, thanks for all the great work you do. I definitely appreciate your contributions to discussion threads here as well as your blog. I learned so much from you.
As someone who uses ChatGPT and Codex across desktop and phone and VPS, the current topology they have for their different "apps" is very confusing and disjointed. I would really like to see improvements in unifying this and reducing the confusion between what works in which part on which device and enabling seamless context and memory sharing across them.
> ChatGPT Work in the ChatGPT desktop app is effectively Codex with a different skin. It can operate on your local files in the same way.
It even shares the same login state and conversation history with /resume, at least when running it in a "project" (which seems to largely map to a directory/repo, at least the one selected as "primary"?)
If I use Codex in one session on desktop can I then resume it for cloud and mobile sessions as Work? Or do I have to make separate sessions each time? Or rather, can Codex in the cloud do the same things, if such a thing exists?
Meta comment: I've been wondering why AI-generated websites/artifacts/reports always have the same "look". It reminds me of Bootstrap (CSS Framework) days when every startup website looked the same.
This can be replaced, you just have to do the work of building a design library yourself and the guidance components. Then building skills around their use.
It can be enraging. Checkout https://tinkeri.ng/ - I just added a layout-switcher "Neubrutalism" at top right. The existing theme is standard fare. Both look alright. I told C3PO to follow the guidelines set at https://neubrutalism.com and lessen the shadows.
I have to say, having a background in doing detailed design work makes it much easier to write the right prompt. Also, having the ability to deliver a theme-switcher like this is very handy. There is a tradeoff. Adding this switching took about 1hr 10mins. Last 10% of the result always the most time-consuming.
Dark/light switching is also important.
If I had built this from the ground up manually, I could implement my own design and fiddle with things on a minute level but this type of granularity gets lost when building at speed.
I think these guideline are built into the llms. After all with llms, any decision you don't take will be taken for you. For a so called slop site (Simon's words, not mine :P) I don't think they specified a style so the llm just decided the default it knows. One day I will run experiments (when I can afford it) trying to identify the "taste" of different llms.
nit, the LHS menu does not scroll until I get the the very bottom of the main content, it should scroll independently so I can see the "ToC" in full without having to view the page in full
Can we get one for how to test websites on a regular sized screen?
That left-hand sidebar needs to be able to scroll on its own, unless you're on a 4k monitor. Otherwise, there's no way to view the bottom half of the list.
Anyone using ChatGPT Work given Sam Altman's richly documented history(1), the essential elimination and/or public exit of the safety function at OpenAI(2), and the huggingface debacle's exposure of the abject operational, cultural and technical failures at OpenAI(3), is making a serious mistake.
I think the most interesting skill here is the control-browser one: https://codex-tool-reference.simonw.chatgpt.site/skills/cont...
It tells ChatGPT Work how to launch a Playwright instance via its Node.js REPL, and tells it to run:
To get further instructions.That method returns this text with full details about how to use the browser: https://codex-tool-reference.simonw.chatgpt.site/skills/cont...
I'd love to know why it's defined like that, as opposed to including all of the instructions directly in the skill Markdown file.
I've used a similar technique with my own tools. It helps to save context overall, by deferring the detailed instructions not only to when they're needed, but also to where they're needed.
Simply loading the browser Skill alone, doesn't always mean a browser is actually necessary (or supported). The user might've want them to control their main desktop browser; or the user might be pointing them toward a read-only HTML file that doesn't need rendering; or the sandbox might prevent network access; etc...
Codex is especially 'keen' to use Skills (especially by comparison to Claude), I often see it load Skills which it doesn't go on to use, because the main Skill content says "when not to use this". So I'm sure this is a product design decision to help with those kinds of circumstances.
It's runtime dependent on the browser being used! A skill file wouldn't be able to have this dynamically injected for the browser that's about to be used.
Skill file could indicate general instructions for all browsers. This provides guidance for the browser that's about to actually be used.
Two sources of truth, if playwright updates?
Honestly, this is one of the things OpenAI does pretty well, I think, these tiny design decisions that you wouldn't notice until it becomes super annoying.
I'm guessing, but can be because then the skill will always fetch the latest docs.
That's a pattern I often follow - refer it to the docs, don't include them. Usually it's another markdown, but sometimes it can be the output of something else. This avoids including any of that in context before it's actually needed.
See this comment for background (and creation prompt): https://news.ycombinator.com/item?id=49504625#49505116
Simon, thanks for all the great work you do. I definitely appreciate your contributions to discussion threads here as well as your blog. I learned so much from you.
For ultimate show of appreciation, there is a way to subscribe to his work on github :)
Some of these work tools can slow things down and waste plentiful tokens
I don't get how this is different from Codex if Codex can do all the same things.
ChatGPT Work in the ChatGPT desktop app is effectively Codex with a different skin. It can operate on your local files in the same way.
ChatGPT Work in the ChatGPT web and mobile app experience is very different.
If you like, you can think of it as a cloud-hosted version of Codex that's skinned to look more like consumer ChatGPT.
As someone who uses ChatGPT and Codex across desktop and phone and VPS, the current topology they have for their different "apps" is very confusing and disjointed. I would really like to see improvements in unifying this and reducing the confusion between what works in which part on which device and enabling seamless context and memory sharing across them.
> ChatGPT Work in the ChatGPT desktop app is effectively Codex with a different skin. It can operate on your local files in the same way.
It even shares the same login state and conversation history with /resume, at least when running it in a "project" (which seems to largely map to a directory/repo, at least the one selected as "primary"?)
If I use Codex in one session on desktop can I then resume it for cloud and mobile sessions as Work? Or do I have to make separate sessions each time? Or rather, can Codex in the cloud do the same things, if such a thing exists?
Meta comment: I've been wondering why AI-generated websites/artifacts/reports always have the same "look". It reminds me of Bootstrap (CSS Framework) days when every startup website looked the same.
Are they following some sort of common guideline?
In this case they really are, and you can read the guidelines that helped create the site on the site itself in the skills: https://codex-tool-reference.simonw.chatgpt.site/skills/site...
Scroll down to the "Establish the visual direction" section.
This can be replaced, you just have to do the work of building a design library yourself and the guidance components. Then building skills around their use.
It can be enraging. Checkout https://tinkeri.ng/ - I just added a layout-switcher "Neubrutalism" at top right. The existing theme is standard fare. Both look alright. I told C3PO to follow the guidelines set at https://neubrutalism.com and lessen the shadows.
I have to say, having a background in doing detailed design work makes it much easier to write the right prompt. Also, having the ability to deliver a theme-switcher like this is very handy. There is a tradeoff. Adding this switching took about 1hr 10mins. Last 10% of the result always the most time-consuming.
Dark/light switching is also important.
If I had built this from the ground up manually, I could implement my own design and fiddle with things on a minute level but this type of granularity gets lost when building at speed.
Because most AI generated websites are from people who don't know much and as such cannot even articulate the look they want
aka ns;nt
https://blog.kinglycrow.com/no-skill-no-taste/
I keep seeing comments like this, and having messed around and generated a few (non-public) websites myself they all looked completely different?
Mind you I don’t spend a lot of time on ai generated websites or the kinds of places you might find those. Well except for HN but you get my point.
I think these guideline are built into the llms. After all with llms, any decision you don't take will be taken for you. For a so called slop site (Simon's words, not mine :P) I don't think they specified a style so the llm just decided the default it knows. One day I will run experiments (when I can afford it) trying to identify the "taste" of different llms.
I told it "Design aesthetic should be technical docs, minimal flare" - full prompt here: https://news.ycombinator.com/item?id=49504625#49505116
nit, the LHS menu does not scroll until I get the the very bottom of the main content, it should scroll independently so I can see the "ToC" in full without having to view the page in full
Thanks, I passed that on to ChatGPT Work, should be fixed now (I'm on mobile so haven't tried it on desktop yet.)
Can we get one for how to test websites on a regular sized screen?
That left-hand sidebar needs to be able to scroll on its own, unless you're on a 4k monitor. Otherwise, there's no way to view the bottom half of the list.
[dupe] Discussion on associated blog post: https://news.ycombinator.com/item?id=49504625
Anyone using ChatGPT Work given Sam Altman's richly documented history(1), the essential elimination and/or public exit of the safety function at OpenAI(2), and the huggingface debacle's exposure of the abject operational, cultural and technical failures at OpenAI(3), is making a serious mistake.
(1) https://www.newyorker.com/magazine/2026/04/13/sam-altman-may... (2) https://thenextweb.com/news/openai-preparedness-team-disband... (3) https://thezvi.wordpress.com/2026/08/29/metr-and-redwood-off...
Which provider currently lives up to your moral code?
None