改造 ChatGPT 聊天/工作模式的七点提案

Simon Smith @_simonsmith原推发布于 2026.09.20

在 X 上打开原推

Here's a proposal for fixing ChatGPT's Chat/Work issue: 1. Eliminate the need to choose chat or work. 2. Bill usage strictly by model, not type of task. 3. Make Luna-level usage effectively infinite. 4. Give me all the features of the current work mode always, both in text mode and voice mode. 5. Have a chat-based "chief of staff" default thread in which I have one long conversation over time, as in Muse. Have the chief of staff start new threads for anything complex. Allow users to tell the chief of staff what default models it wants for different types of tasks (e.g. "for coding tasks, always use Astra") and to ask for different models at the time of delegation. Have delegated threads report back to the chief of staff when they're done, and the chief of staff tells the user. 6. Let me talk to the chief of staff everywhere. Email, Slack, Messages, phone. Wherever, however, whenever I want. 7. But still let power users start complex threads on their own, as well as, of course, create projects and work in projects. Projects probably also should have this chief of staff model, as Claude just rolled out. So, like: 1. I talk to ChatGPT. There's one thread. Just like talking to a human. It's just a dude I talk to. Friendly, conversational, knows me really well. I can reach it via multiple channels. 2. I tell this dude to do something for me. It spins up a thread to execute that task and decides, based in part on what I've told it in the past, which model to assign the job to. It then monitors that thread and lets me know when it's complete, brings back the result, and provides me a link to the thread if I want to work directly in that thread later. 3. After it fires off a task in a delegated thread, we can keep talking. I can ask for many more tasks. It just keeps firing them off. 4. None of this requires me to select "chat" or "work." When we're chatting, we're chatting. When we're working, we're working, usually in a delegated thread. 5. My usage gets consumed according to model usage. Not whether I'm in chat or work mode. 6. I can still link into individual chats somehow. Like, maybe in the sidebar based on timestamp or something. Or, maybe we have the concept of main chat and side chat like in Muse. Not sure. Anyway, those are just some ideas that came to mind, and I haven't pressure-tested them, so they're quite half-baked right now. But it feels directionally right to me. It's also kind of how you can get things to work in ChatGPT on the desktop right now, if you want to. But you can't really do it right now elsewhere.