Grounding Instructions

Grounding instructions anchor the model to a specific set of facts, documents, or data as the sole acceptable source for its output. Without grounding, the model draws on its training data — which is vast, mostly accurate, occasionally wrong, and always generalized. When your task requires precision about specific facts, figures, or documents, grounding constraints are not optional. They are the primary mechanism for preventing hallucination in factual tasks.

Pattern 1: Document-grounded
"Answer questions based only on the information in the document provided below. Do not use information from your training data. If the answer is not in the document, respond with 'This information is not available in the provided document.'"

Pattern 2: Data-grounded
"Generate the analysis using only the data in the table below. Do not extrapolate beyond the provided figures. Do not include industry benchmarks or external comparisons unless they appear in the provided data."

Pattern 3: Fact-grounded
"All factual claims in your response must be supported by the source material provided. For each claim, cite the specific section it comes from. Do not include claims that cannot be attributed to the source material."

Pattern 4: Scope-grounded
"Base your recommendations only on the constraints listed in Section 2 of the attached brief. Do not consider factors outside those constraints, even if they would be relevant in a broader analysis."