Prompting for Scientific Documents and Data
Scientific work has specific challenges for AI use: the precision requirements are high, the domain vocabulary is specialized, the data formats are heterogeneous, and the distinction between established knowledge and emerging findings matters enormously. Prompting patterns for scientific work encode these requirements.
Working with literature and evidence
When asking AI to synthesize scientific literature, the critical risk is hallucination — confident citation of papers that don't exist or mischaracterization of papers that do. The mitigation is explicit sourcing requirements.
Pattern: "I am providing you with [N] research papers as attachments. Your task is to synthesize these specific papers only — do not add references from your training data that I have not provided. For every finding you cite, include the author name and year in parentheses, drawing only from the provided papers."
This pattern constrains the AI to your provided corpus. It doesn't prevent mischaracterization of the provided papers (that requires your verification), but it prevents fabrication of non-existent sources.
Working with experimental data
When asking AI to analyze or summarize experimental data, precision and uncertainty requirements are paramount.
Pattern: "The following data represents [experiment type] with [N replicates]. Analyze it as follows: 1) Identify the statistically significant findings using the criteria [specify: p<0.05, effect size threshold, etc.]. 2) Distinguish between statistically significant and clinically/biologically meaningful results. 3) Flag any results that are at the boundary of statistical significance (0.05 < p < 0.10). 4) Identify outliers and note whether they were included in the analysis."
This pattern forces the AI to apply specified statistical criteria rather than defaulting to casual language about what the data "shows."
The evidence grading pattern
For summaries that will inform regulatory strategy or clinical decisions, explicit evidence grading prevents conflation of different levels of evidence.
Pattern: "Grade each finding by evidence level: [1a] Systematic review or meta-analysis. [1b] Individual RCT. [2a] Systematic review of cohort studies. [2b] Individual cohort study or low-quality RCT. [3] Case-control study. [4] Case series or cohort without controls. [5] Expert opinion. Clearly mark the evidence level of each finding you include."