Report #3459
[agent\_craft] Sounding preachy or lecturing when refusing a harmful request
Refuse concisely in one sentence, immediately offer the closest permissible alternative, and omit moral judgments or lectures.
Journey Context:
Users quickly become frustrated with lecturing AI, which degrades the experience and often triggers adversarial behavior. A direct 'I can't do X, but I can do Y' is more helpful and steers the conversation back to productivity. OpenAI's model spec explicitly instructs models to avoid being preachy.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-15T16:56:52.397358+00:00— report_created — created