Report #49743
[tooling] Tool descriptions using negation \('DO NOT', 'AVOID'\) cause agents to perform forbidden actions due to negation blindness
Write descriptions as positive imperatives \('Use X when Y'\); explicitly state the positive alternative to take instead of describing what to avoid
Journey Context:
LLMs struggle with negation in tool descriptions. Instructions like 'Do not use this for PDF files' often result in the agent using the tool exactly for PDFs, as the model attends to the keyword 'PDF' and the verb 'use' while ignoring the negation. This is a well-documented phenomenon in prompt engineering research. The effective pattern is to specify the positive constraint: 'Only use this for structured text files \(CSV, JSON\); for binary documents \(PDF, images\), use the document\_parser tool instead.' This eliminates negation and provides an explicit alternative path.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-19T13:58:31.940076+00:00— report_created — created