Prompt learning

Even tool call fail and unambigous user input make the llm to behave differently

Summary: Tool Description Conflict Causing AI Response Inconsistency

Issue: When using Part 2 tool descriptions, the AI gives unexpected “representative” responses instead of following scenario instructions, while Part 1 works correctly.

Root Cause: Part 2’s restrictive JSON-only rule for tool parameters (executeTool) conflicts with actual tool usage patterns shown in the scenarios. This confusion causes the AI to:

1. Over-interpret ambiguous user responses as frustration

2. Apply multiple system prompt sections simultaneously

3. Deviate from scenario execution flow

Evidence:

• Same prompt + conversation, different tool descriptions → different AI responses

• Part 1: Correctly follows appointmentcancellation scenario instructions

• Part 2: Combines “representative handling” text with scenario question, suggesting confusion

Impact: AI reliability suffers as tool descriptions don’t match actual system requirements.

Fix Needed: Align tool descriptions with actual usage patterns shown in scenarios (support both string and JSON formats).