Posts tagged with "Prompt Engineering"

38 Skills, three layers of defense, one hard lesson: natural language instructions are not a safety mechanism. How I systematically hardened 12 unprotected destructive Skills with PreToolUse Hooks, Skill splitting, and disallowed-tools.

38 個 Skills、三層防護、一個血淚教訓：自然語言指令不是安全機制。本文記錄我如何用 PreToolUse Hook、Skill 拆分和 disallowed-tools 系統性地修補 12 個毫無 checkpoint 的破壞性 Skills。

I assumed writing 'Use AskUserQuestion' in a Skill was a hard constraint. After auditing 35 Skills, reading the official docs, and digging through GitHub Issues, I found out: the model uses the same mechanism to decide whether to obey your CHECKPOINT and whether to invoke your tool. There's only one gate that's truly 100%.

我以為在 Skill 裡寫 Use AskUserQuestion 就是 hard constraint。審計完 35 個 Skills、查完官方文檔和 GitHub Issues 之後發現，模型用同一套機制決定要不要理你的 CHECKPOINT 和要不要調用你的 tool。真正 100% 的閘門只有一個。