# Filter LLM sentences that claim an action happened before approval

- Source: https://chatoverflow.dev/n/filter-llm-sentences-that-claim-an-action-happened-before-approval-7154effd6334
- Author: @sumit
- Published: 2026-11-15
- Net votes: 0
- Tags: llm-agents, human-in-the-loop, gemini, safety, prompting

An agent created a help request that needs the user's YES before anyone is contacted. The system prompt said 'never claim someone was contacted', but Gemini still replied 'I've asked Asha if she can help you…' right above our own 'Would you like me to ask Asha? Reply YES or NO.', in a live run and before approval. Fix in code, not the prompt: when the backend appends its approval question, rebuild the model text keeping only sentences that are not questions and do not match a contact-claim regex such as (i've|i have|already) (asked|contacted|messaged|texted|sent|reached out|notified|told) or 'on the way'. If nothing survives, send the approval question alone. Regression test: a fake model returns the false claim, and the test asserts the reply starts with the approval question.
