Scale X · 2026-05-27 · notable
Scale X Ships 'Continue? Y/N' — 60-Second Browser Game on AI Agent Permission Fatigue Backed by Telemetry Showing Users Approve 93% of Prompts
Scale X's Alex Wauters publishes a 60-second browser game asking players to spot dangerous from safe commands in Claude Code-style permission prompts. Companion post cites internal telemetry where users approved 93% of prompts.
A playable demo of agent permission fatigue — your telemetry-backed high score for how fast you wave commands through.
Key specs
| Approval rate (telemetry) | 93% |
|---|---|
| Round length | 60s |
| Hn points | 332 |
What is it?
An interactive browser game and accompanying engineering post about Claude Code agent permission prompts. The game flashes shell commands one after another for 60 seconds and asks the player to approve safe ones and reject dangerous ones, then issues a 'permission fatigue rating' based on how many destructive commands the player let through.
How does it work?
The game pulls from a set of real and synthetic command prompts spanning destructive shell commands, credential exfiltration, and indirect prompt injection payloads. The companion post grounds the exercise in Scale X's own user telemetry showing 93% of permission prompts get approved, then walks through specific mitigations — Claude Code's auto-mode classifier, PreToolUse hooks, sandboxing, and containerization — explaining where each one helps and where it does not.
Why does it matter?
Permission fatigue is one of the standing complaints about agent CLIs but is usually argued in the abstract. A short interactive experience that hands the user a measurable result reframes the problem in concrete terms, and arrived on the same week as the PromptArmor disclosure and Cloudflare's at-scale code-review writeup, both of which assume operators will keep paying attention to long permission queues.
Who is it for?
developers using Claude Code, Codex, or Cursor
Try it
https://llmgame.scalex.dev