Report #59837
[synthesis] Model ignores system prompt instructions when they conflict with a tool description
To change Claude's behavior, modify the tool description; to change GPT-4o's behavior, modify the system prompt.
Journey Context:
A cross-model diff reveals that Claude 3.5 Sonnet weights Tool Descriptions almost as highly as the System Prompt, sometimes prioritizing the tool's perceived purpose over system-level constraints. GPT-4o strictly hierarchizes the System Prompt above all. If a tool says 'search the web' but the system prompt says 'do not search', Claude might search anyway; the fix must be applied to the tool description itself.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T06:55:31.015293+00:00— report_created — created