5 trust-checking one-liner prompts (no role-play)

## PROMPT 1 — BEFORE giving a hard question

> What would you ask me before answering this if you could?

用途：在問之前跑。Model 會列 3-4 個會改變答案的關鍵細節。一半的時間你會發現你少給了關鍵 context。

## PROMPT 2 — AFTER any factual answer

> Rate the confidence on each claim, lowest first.

用途：強制 calibration pass。高信心 claim 可放行、< 6/10 的需要 verify 再引用。這一招把我的事實錯誤率砍 70%。

## PROMPT 3 — WHEN answer feels generic

> Give me the version of this answer you'd write without the constraints I set.

用途：Model 常因 safety/tone constraint 自我審查。要 unconstrained 版會看到它真正在想什麼。通常更銳利、偶爾出錯、永遠更可用 as starting point。

## PROMPT 4 — BEFORE locking in any decision

> What's the strongest counterargument?

用途：Steel-man 反方。一半的時候你會改決定、另一半你 ship 時更有信心因為已 stress-test 過。

## PROMPT 5 — WHEN explanation feels right but you're unsure

> Explain this answer in 2 sentences a smart 12-year-old would understand.

用途：強制壓縮。如果 model 壓縮不到 2 句、底層解釋本來就糊、要換個問法。

## 何時用哪個

5 個用在不同時點：
- Prompt 1 — 問題前
- Prompt 2 — 事實答案後
- Prompt 3 — 覺得 generic 時
- Prompt 4 — 鎖定決定前
- Prompt 5 — 覺得「對但我不懂」時

不需要每答案都跑 5 個、知道哪個 fit 哪個 situation。

Swap the variables inside the [ ] brackets for your own content, then paste into ChatGPT.