Simon Willison Made ChatGPT Work Publish Its Own Manual
Simon Willison pulled a very Simon Willison move: he asked ChatGPT Work to inventory every tool and skill it ships with, format the whole thing as a technical docs site, and host it on ChatGPT Work's own site-hosting feature. The result is a public snapshot of OpenAI's agent harness as of August 31: 232 tool interfaces and 44 skill definitions, live at https://codex-tool-reference.simonw.chatgpt.site/. He calls it a "vibe-coded slop site." It hit 157 points on Hacker News anyway, because the content is something OpenAI itself never published.
The numbers tell you where OpenAI thinks agents earn their keep. Of 232 tools, 89 are GitHub operations. There are six tools just for sub-agent coordination: spawning, messaging, interrupting. Add scheduled automations, webhooks, Gmail, Calendar, Sheets. Desktop ChatGPT Work, as the HN thread put it, is effectively Codex with a different skin.
The 44 skills are the more interesting half. They sit above the tools: document production, data analytics, browser control, even a skill for creating new skills. The design detail HN engineers liked most: the browser-control skill doesn't embed its Playwright documentation upfront, it defers it to runtime via a browser.documentation() call, so the context window only pays for it when the agent actually browses. That's a context-engineering pattern worth stealing.
Step back and this is the second time in a month we've gotten an unsanctioned x-ray of a frontier lab's harness. Anthropic's came via the Project Glasswing release notes (https://clauday.com/article/7c67757e-d195-4a61-80d1-2ae6a837528e). And the skills layer here looks structurally identical to Anthropic's Agent Skills, the same standard K-Dense just rebranded around (https://clauday.com/article/8997635e-7386-4639-b2f4-cc3f394d86ac). Two labs, one packaging format for expertise. The convergence is no longer a prediction, it's shipped product on both sides.
← Back to all articles
The numbers tell you where OpenAI thinks agents earn their keep. Of 232 tools, 89 are GitHub operations. There are six tools just for sub-agent coordination: spawning, messaging, interrupting. Add scheduled automations, webhooks, Gmail, Calendar, Sheets. Desktop ChatGPT Work, as the HN thread put it, is effectively Codex with a different skin.
The 44 skills are the more interesting half. They sit above the tools: document production, data analytics, browser control, even a skill for creating new skills. The design detail HN engineers liked most: the browser-control skill doesn't embed its Playwright documentation upfront, it defers it to runtime via a browser.documentation() call, so the context window only pays for it when the agent actually browses. That's a context-engineering pattern worth stealing.
Step back and this is the second time in a month we've gotten an unsanctioned x-ray of a frontier lab's harness. Anthropic's came via the Project Glasswing release notes (https://clauday.com/article/7c67757e-d195-4a61-80d1-2ae6a837528e). And the skills layer here looks structurally identical to Anthropic's Agent Skills, the same standard K-Dense just rebranded around (https://clauday.com/article/8997635e-7386-4639-b2f4-cc3f394d86ac). Two labs, one packaging format for expertise. The convergence is no longer a prediction, it's shipped product on both sides.
Comments