Anthropic measured its own staff (report of 16 October 2026): in interactive, multi-turn sessions, "Claude sessions amount to around 80% idly waiting for human input." That's Anthropic engineers, not everyone. But any AI that has to ask before every small choice stalls the same way, and you end up answering questions one at a time.
These 3 lines are my own brief, not Anthropic's advice. They tell the AI what it can decide, what has to come to you, and to save the rest.
What you need
A longer task for Claude, ChatGPT, Copilot or another AI agent
30 seconds to edit line 1 for your task
The 3 lines
Paste these above the task:
Decide these yourself: file names, formatting, the order of sections, which source to read first.
Ask me before you send, delete, pay or share anything.
Save every other question for one list at the end. Don't stop to ask mid-task.
Change the examples in line 1 to the small choices in your task. Keep line 2 as it is.
Then
Check in twice a day, not every 10 minutes. Answer the list at the end in one go.
The mistake to avoid
Cutting line 2 to make it faster. Anything that sends, deletes, pays or shares should still come to you first.