ChatGPT Work Tool and Skill Reference: Snapshot
A public reference site now lists 232 tool interfaces and 44 complete skill files pulled from ChatGPT Work, giving anyone curious about how the assistant actually operates a documented look at its callable endpoints and instruction packages rather than a marketing description of them.
A public reference site now lists 232 tool interfaces and 44 complete skill files pulled from ChatGPT Work, giving anyone curious about how the assistant actually operates a documented look at its callable endpoints and instruction packages rather than a marketing description of them.
What the snapshot actually contains
The inventory splits ChatGPT Work's internals into two kinds of objects. Tools are described as callable endpoints, the functions the model can invoke to do something in the world, of which the snapshot records 232. Skills are described as reusable instruction packages, and 44 of these exist as complete main skill files. The site states plainly that skill pages reproduce the complete current SKILL.md source, so what you're reading on each skill page is that file's content as it stood at the time of the snapshot, not a paraphrase or a summary written by a third party. That distinction matters if you're trying to understand how an agent decides what to do next: the tools are the verbs available to it, and the skills are the written procedures that tell it when and how to use those verbs.
Why the same tool list won't show up twice
The reference is explicit that tool availability can vary with session configuration, permissions, connected apps, and installed plugins. That means the 232 interfaces catalogued here represent one observed configuration, not a fixed ceiling or a guaranteed baseline for every account. Anyone comparing their own experience with ChatGPT Work against this snapshot should expect some tools to be missing or additional ones to appear, depending on what's connected and what's been granted. This is also where a recurring community question landed: someone asked how ChatGPT Work differs from Codex, and the honest answer implied by the reference itself is that the boundary is fuzzy by design — the same underlying model can be wired into very different tool sets depending on the product surface and the permissions attached to a session.
The one skill everyone kept pointing at
Among the 44 skill files, control-browser drew the most attention from people reading through the reference, and it's easy to see why. It explains how to launch a Playwright instance through a Node.js REPL, which puts a real browser under the model's control rather than a simulated one. What made it stand out further is a detail one reader specifically flagged: the workflow calls browser.documentation() to retrieve further browser instructions dynamically, instead of listing every instruction directly in the skill's Markdown file. That means the static SKILL.md source you can read on the page is only the entry point into the skill's full behavior, with additional instructions fetched at runtime rather than committed to the document itself.
The friction points readers actually hit
Not every reaction was about capability. Several people ran into interface issues while trying to use the reference itself — the left sidebar can require independent scrolling to reach lower items on smaller screens, and at least one reader asked outright for guidance on testing the site at standard desktop screen sizes. There was also a broader complaint that surfaces whenever tool-heavy agent setups get examined closely: some work tools can slow tasks down and burn through tokens unnecessarily, a cost that isn't visible from a tool's name or its entry in a registry. And a few commenters raised concerns entirely outside the reference itself, questioning whether to trust ChatGPT Work given various allegations about OpenAI's leadership, safety changes, and operational failures — a reminder that documentation like this answers the how but not the whether.
Where the aesthetic critique fits
One observation ties the whole thing back to something bigger than this one reference: a commenter noted that AI-generated websites, artifacts, and reports tend to share a strikingly similar visual style, regardless of who built them or what tool produced them. Applied to this snapshot, that's a useful caution rather than a dismissal — the reference's value isn't in how it looks but in the fact that it reproduces primary source material, the actual SKILL.md files and the registry of tool interfaces, rather than a rendering of them. If you want to see it yourself, the site is public and browsable at codex-tool-reference.simonw.chatgpt.site, with individual skill pages such as the control-browser entry linked directly from its index, and background on how the project was built, including its original creation prompt, posted alongside it for anyone who wants to reproduce or extend the same kind of inventory for a different session.
Specs
- Tool interfaces: 232
- Complete main skill files: 44
- Skill categories: 44
- Registry tools: 223
- Direct session controls: 9
Sources
- ChatGPT Work Tool and Skill Reference — codex-tool-reference.simonw.chatgpt.site
- control-browser skill page — codex-tool-reference.simonw.chatgpt.site
- control-browser: browser.documentation() — codex-tool-reference.simonw.chatgpt.site