Most people use AI serially: ask, wait, read, ask again. The model works at machine speed; the workflow runs at human-waiting speed. If a reply takes ninety seconds and you send forty prompts a day, you spend an hour watching progress indicators.
The fix is old. Telephone switchboard operators never waited for one call to end before patching the next — they kept a board of lines, each with a state, and touched only the line that needed a human. That is exactly the right mental model for AI chats.
Design your board
- One session per task, not per topic. “Refactor auth module”, “draft launch email”, “review contract” — each gets its own tab with its own context. Mixing tasks in one chat destroys both.
- Pick the model per task. Long agentic coding runs to Claude, quick drafts to ChatGPT, document crunching to Gemini's context window. Parallelism means you stop caring which model is “best overall”.
- Prompt for autonomy. End prompts with the decisions pre-made (“prefer X over Y, don't ask, note assumptions at the end”). Every clarifying question the model doesn't need to ask is a blocked state you never enter.
Run the loop
The operating loop has two rules. Rule one: when a session finishes, respond to it before it cools — review the output, send the next prompt, move on. Rule two: blocked beats finished. A tab waiting on a permission or a question is dead time compounding; a finished tab is just inventory.
The loop only works if state changes reach you. That's the job of AI Tab Switch: a live status list of every AI tab, notifications on finished and blocked states, a badge with the ready count, and prompt staging so your next instruction is queued before the current reply even lands.
How many sessions is too many
In practice the ceiling isn't the tooling, it's review bandwidth. Four to six heavy sessions (code, long documents) or eight to ten light ones (drafts, lookups) keeps review quality honest. Past that, outputs pile up unread and you've just moved the queue from the model's side to yours. Start with three, add one at a time, stop when review starts feeling rushed.
The payoff is not subtle. Serial prompting keeps one model busy while you idle; a switchboard keeps you busy while five models idle — and their idling is free.