Learn vocabulary for diagnosing and fixing prompts: instruction-following failures, hallucination, format errors, negative constraints, and grounding.
0 / 5 completed
1 / 5
What does 'the model is not following instructions' mean in prompt debugging?
Instruction-following failure means the model's response violates one or more of the constraints or directives in the prompt. Common fixes include making instructions more explicit, moving them to a system message, or rephrasing to reduce ambiguity.
2 / 5
What does it mean when you say 'it's hallucinating facts'?
Hallucination refers to confident generation of false or fabricated information. Mitigation strategies include grounding the prompt with retrieved facts, instructing the model to say 'I don't know' when uncertain, or using retrieval-augmented generation.
3 / 5
What is 'iterative refinement' in prompt engineering?
Iterative refinement is the core debugging loop in prompt engineering: run the prompt, evaluate the output, identify what went wrong (wrong format, hallucination, missed constraint), adjust the prompt, and repeat until the output is satisfactory.
4 / 5
What is a 'negative constraint' in a prompt?
Negative constraints explicitly forbid certain outputs or behaviours. They are often needed when the model defaults to a behaviour you want to prevent — for instance, adding unsolicited caveats, using a particular format, or speculating beyond the provided data.
5 / 5
What does 'grounding the output' mean in prompt debugging?
Grounding means giving the model concrete evidence or data to draw from, and instructing it to base its response on that material. This reduces hallucination by constraining the model to what is actually present in the provided context.