Skillet

Skillet is a curated web home for Claude Code skills — the same skills that normally live in a developer's terminal — packaged as no-install browser experiences. The premise is that Claude Code skills are powerful, but most people will never open a terminal to try one. Skillet bridges that gap with a thin web layer that connects to a real Claude Agent SDK session, not a chatbot pretending to be one.

When a user starts a session, Skillet spins up a Claude Agent SDK instance on Anthropic's Managed Agents service — the same runtime that powers Claude Code locally — with the chosen skill loaded. The browser is just the window into the agent. The agent has the same read, write, and structure tool primitives Claude Code uses, drives the conversation, and writes the user's artifact as it goes.

Skillet launches with /office-hours, the structured-conversation skill from Garry Tan's open-source gstack toolkit, which packages the Y Combinator office hours format into a skill that turns vague ideas into clear written plans. Two more skills are queued: /ce-plan for end-to-end idea planning and /last30days for surfacing what people actually said over the last month. The source skills are loaded directly, so updates upstream flow through to Skillet sessions.

• Real Claude Agent SDK sessions — Each session is a real agent on Anthropic Managed Agents, not a wrapped chatbot, with the same runtime as Claude Code in the terminal.
• No-install browser experience — Click start and run a skill in the browser with no terminal, signup, or API keys required during beta.
• Curated skill library — One skill at a time, each with its own workflow, structure, and deliverable, starting with Garry Tan's /office-hours.
• Live-linked source skills — Skills are loaded from their canonical source repos (e.g. gstack), so upstream updates reach Skillet sessions automatically.
• Shareable doc artifacts — Sessions can produce a publicly shareable document URL capturing the conversation's output.