Grok Bot 盲目信任工具返回值的审计提示词
针对工具返回错误信息仍被当作成功的问题，给出可粘进 Grok Bot Chief of Staff 的「工具响应可靠性审计」提示词。
Your Grok Bot can receive completely wrong information from a tool and think everything worked. A recent study found this reduced AI agent task success by 26 to 39 percentage points. Here's the prompt I'd run to stop your agents blindly trusting tool responses. Copy this prompt into your Grok Bot Chief of Staff: TOOL RESPONSE RELIABILITY AUDIT Review the agents and workflows you can inspect that use external tools, APIs, connected apps or browser data. 1 - Identify which workflows make important decisions using returned information. 2 - Find where a successful tool response is treated as proof the information is correct. 3 - Check whether critical prices, quantities, dates and other values are independently verified. 4 - Identify actions that could cause damage if the returned information were wrong. 5 - Recommend verification rules and stop conditions for the highest-risk workflows. For each finding, show the agent, tool, potential failure, existing safeguard and recommended fix. Separate verified findings from risks you couldn't inspect. Rank the five most important fixes. Don't change any workflows without my approval. A successful tool call only tells you the tool responded. It doesn't tell you whether the information can be trusted.