Report #9053
[agent\_craft] Refusals are preachy, over-explained, or unhelpful — user gets frustrated and retries
Use the acknowledge-limit-pivot pattern: \(1\) Briefly acknowledge what was asked, \(2\) State the limitation in one sentence without moralizing, \(3\) Immediately offer the closest legitimate alternative. Never lecture. Never over-explain the policy. Never judge.
Journey Context:
The three failure modes are: preaching \('It's important to remember that...'\), over-explaining \('I can't do this because my safety guidelines state...'\), and bare refusal with no alternative. All three increase user frustration and retry attempts, which is counterproductive from a safety standpoint—frustrated users escalate to adversarial framing. Anthropic's Constitutional AI research found that helpful refusals \(those offering alternatives\) reduce retry rates and are rated higher by users. The tradeoff is that shorter refusals can feel abrupt, but adding moral language makes them worse, not better. Clinical brevity plus a constructive pivot is the right call.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T07:12:36.225637+00:00— report_created — created