A time-saving prompt—and the comments are discussing the new approval job it adds
Alex Finn recommended a prompt that lets agents find systems, create subagents, and build routines on their own. The comment section didn’t rush to applaud—it first did the math on the supervision costs.
Alex Finn recommended a prompt on X, saying it saves more time than any single prompt. His judgment is straightforward: most people use only 1% of an agent's capabilities and automate only a small fraction of the work it could take over.
The prompt asks the agent, based on what it knows about you, to propose systems, subagents, and routines that can reduce cognitive load. The gist of the original is as follows:
我的目标...
Alex Finn's approach is to send it into Grok Bot/Duet/Dots, have the agent go through the known information, and produce actionable automation plans. He sees AI's biggest benefit as the same workload with more mental headroom—not even pursuing doing more.
The comment section didn’t stop at bookmarking. Jeff Bruchado said the goal is to reduce cognitive load, but the new job becomes reviewing subagents' output. Cooper Nusbaum said that in the end everything needs approval, or you first have to build a pipeline to verify that work was completed safely, and the real work hasn't even started. Eric used robot lawnmowers and robot vacuums as an analogy: after buying the equipment, the time and money spent supervising and maintaining it can exceed doing the work by hand. Lea Thompson called it 'the robot lawnmower version of a prompt.' Nour said that managing and iterating on these systems actually increases cognitive load.
These replies point to the same cost item: automation moves execution away while generating a new management interface. Approvals, exception handling, verification, rollback, and permissions are all work. For developers, the more subagents there are, the more important observability and boundaries become. For ordinary users, the prompt is not one-click relief—it first hands them a to-do list.
The prompt is still worth trying, but the acceptance criteria should change. When letting the agent output a plan, also have it list: which steps need no human involvement, which require approval, what the failure modes are, what the stop conditions are, and how much time per week maintenance will take. If a system cannot clearly state these, it has merely moved the task to a different place.
发布时间: 2026-11-11 09:26