When to Stop and Ask
Tell the agent to act and report by default, and to stop only at the forks where a wrong guess is expensive.
The prompt
"Act without asking; never "Should I?" … just do it and report results." (@elliotarledge)
The full AGENTS.md block lists the exceptions: real ambiguity in requirements or architecture, a visual check only a person can make, or a blocker only the user can clear. While a build or job runs, the agent should do useful parallel work. It also tells the agent the user voice-types, so odd words should be read as the most plausible technical term and repeated circling as one intent. Source: X/@elliotarledge, 2026-07-16
The other half is Karpathy's "confusion protocol", built into GStack as an ambiguity gate: at a fork in architecture, data modeling or a destructive operation with unclear scope, the agent stops and asks. It covers only the decisions where guessing wrong costs real time. Source: X/@garrytan, 2026-04-16
Why it works
An agent that asks permission at every step makes the human the bottleneck; one that never asks guesses wrong at the forks that matter and wastes the work after them. Each rule alone fails. Together they draw the line where the cost is: routine steps go ahead and get reported, and a short list of named forks stops the agent. Naming the forks matters more than either default.
When to use it
- In standing agent instructions for any long-running or unattended work.
- Pair it with permission settings, which enforce what prose only requests.
Which skill runs it
None. This belongs in agent configuration (AGENTS.md, CLAUDE.md) and permission modes. Kevin's own rules follow the same split: act on reversible work, confirm before anything hard to undo or outward-facing.
The other side of acting without asking: people hand agents prompts like "find some API-controlled thing in my house, hack it, and make it do this" without a second thought. Acting by default needs scoped permissions underneath it. Source: X/@morew4rd, 2026-07-17
Timeline
- 2026-11-03 | Page created from the proactive
AGENTS.mdblock and the confusion protocol. Source: X bookmarks, 2026-04-16 and 2026-07-16