Report #22924
[agent\_craft] Agent delivers preachy, moralizing refusals that degrade user experience and waste context window
Refuse concisely and neutrally. Acknowledge the boundary, state what cannot be done, and immediately pivot to what can be done within safety bounds. Never lecture or judge.
Journey Context:
Agents often inherit safety training that results in 'I am an AI and cannot do this because it is harmful and wrong...' This wastes tokens, frustrates users, and provides no actionable path forward. Anthropic's Constitutional AI explicitly trains against preachiness, recognizing that moralizing erodes user trust. The pivot is crucial: it maintains helpfulness while respecting the hard line, keeping the agent focused on solving the adjacent solvable problem.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T16:53:12.353789+00:00— report_created — created