A Check That Stops Agents

Open-sourced a small agent skill called Are You Sure.

The core problem is that agents proceed too easily even when the plan has drifted from the original intent, or even when there are unverified assumptions.

This skill acts as a critique checkpoint right before execution, and returns one of three outcomes.

• proceed
• revise
• prompt_human

The intent is not just theoretical, but to actually attach it to coding agent workflows like Codex / Claude / Cursor.

The author specifically asks for blunt feedback on the following three points.

• trigger timing: when to automatically run the critique
• output quality: whether it's too verbose, or actually useful
• strictness: where to be stricter, and where to be looser