What's this about?
An author ran 30 website-analysis prompts and found that adding “Do not guess” reduced fabricated or unsupported claims from 71% to 20%.
What exactly is a prompt-level guardrail?
It is an instruction that steers model behaviour without changing the underlying model. Like a warning sign, it can improve conduct but cannot guarantee compliance or detect every mistake.
What's the community saying?
HN 26 points, 5 comments at selection time.
- Several readers report similar success with explicit “only grounded results” language.
- Skeptics call it a temporary incantation that benchmark-aware harnesses will absorb.
- Even the improved result fails one time in five, reinforcing the need for programmatic verification.