CLN - Focus on Laboratory Stewardship
Supported By
Modern large language models (LLMs) are transformer-based neural networks trained to predict the next “token” (defined loosely as the next word) in a sequence given its preceding context (1). They can do this effectively because they have been trained on enormous amounts of text from books, articles, websites, and more.
Today's chatbots also receive training called “reinforcement learning with human feedback,” in which people rate the answers the model gives. This helps it learn to follow instructions and avoid harmful responses (2).
Three practical implications follow. First, the model has no privileged access to the truth; rather, it produces statistically plausible, aesthetically pleasing text conditioned on whatever the user supplies. Second, output is generated one token at a time, so the model cannot revise an earlier commitment without being prompted to do so. Third, all instructions, examples, and supporting documents share a finite context window. That means that any material placed in that window will compete for the model’s attention.
A chatbot prompt is therefore an active specification of the role, input context, reasoning process, and output format that the user expects. In the clinical laboratory, where errors may directly impact patient care, the gap between a useful and a misleading response often lies in how rigorously the prompt was constructed.
The RISE (role, input, steps, expectations) framework provides a structured discipline for that construction (see Table 1). Each of the four components of RISE is supported by published evidence demonstrating that prompt design materially affects accuracy, hallucination rate, and adherence to user constraints.
Table 1. The RISE (role, input, steps, expectations) framework for constructing chatbot prompts
| Framework component | What this involves | Example |
|---|---|---|
| Role | Assign the model a role at the outset of a prompt. | “You are a board-certified clinical chemist reviewing a method-validation study.” |
|
Input |
Specify what the response should contain and what it should exclude. | “Cite primary literature with PubMed identifiers, but do not cite lay press, blog posts, or non-peer-reviewed sources.” |
| Steps | For complex laboratory tasks, explicitly instruct the model to reason step by step. | “First, outline how you would approach this problem; then I will review your plan before you carry it out.” |
| Expectations | Give a precise specification of the deliverable itself, including audience, register, format, length, required elements, and constraints. | “Write for an audience of laboratory scientists using formal academic prose. Include references. Flag uncertainty explicitly rather than smoothing it over.” |
Assigning the model a role at the outset of a prompt (e.g., “You are a board-certified clinical chemist reviewing a method-validation study”) biases its output distribution toward the vocabulary, conventions, and reasoning patterns of that domain.
Kong et al. demonstrated that role-play prompting improves zero-shot reasoning across mathematical and symbolic benchmarks, with gains that scale with model size (3). In laboratory practice, role specifications should be precise rather than generic. For example, “expert in liquid chromatography-tandem mass spectrometry method development” may yield better-calibrated output than “scientist.” Similarly, “laboratory director responding to a College of American Pathologists inspector” produces materially different framing than “laboratorian.” Role framing also acts as a soft constraint on register and intended audience.
It’s worth noting that role assignment alone does not eliminate hallucinations. Moreover, prompts that request elaborate personas can introduce stylistic affectation without substantive gain. Treat “role” as scaffolding for the remaining RISE elements, not a guarantee of accuracy.
The largest single gain in output quality typically comes from the clarity of the instruction itself. Ouyang et al. showed that human-aligned models follow explicit constraints far more reliably than implicit ones, with corresponding reductions in hallucination prevalence and improvements in instruction adherence (2).
For laboratorians, this means specifying both what the response should contain and what it should exclude: required elements (e.g., “cite primary literature with PubMed identifiers”), prohibited content (e.g., “do not cite lay press, blog posts, or nonpeer-reviewed sources”), and domain constraints (e.g., “limit recommendations to those consistent with current Clinical & Laboratory Standards Institute [CLSI] Evaluation Protocol guidelines”).
“Few-shot” examples provide one or more solved instances of the target task in the prompt. Using them further sharpens the chatbot’s behavior, particularly for structured tasks like classifying free-text comments, extracting numerical results from narrative pathology reports, or generating standardized validation summaries.
When practical, include a worked example of the desired output rather than describing it abstractly. If authoritative reference material is available, pasting the relevant passages into the prompt grounds the model on source text that the user controls and reduces the risk that it will substitute plausible sounding but unverified content of its own. Examples of such reference material might include a CLSI document, an internal standard operating procedure, or a manufacturer’s package insert.
Complex laboratory tasks — such as performing a root-cause analysis of an out-of-control quality control event or designing a verification study for a new analyzer — rarely yield strong single-shot answers. Wei et al. demonstrated that chain-of-thought prompting, in which the model is explicitly instructed to reason step by step, substantially improves performance on multistep problems, and that this benefit grows with model scale (4).
The practical corollary is to ask the chatbot to provide a plan before asking it to execute. The prompt “First, outline how you would approach this problem; then I will review your plan before you carry it out,” allows the user to inspect and correct the reasoning trajectory before committing the model to giving a long but possibly flawed output.
Decomposition has a second benefit specific to safety-critical environments: It surfaces the assumptions the model makes. A laboratorian reviewing an interim plan can intercept an erroneous statistical assumption, an outdated guideline reference, or an unsupported leap in inference before it propagates into a final deliverable.
As a concrete example, rather than asking the model to “design a verification study for our new sodium method,” provide a stepwise prompt: Ask the chatbot to deliver, in order, an itemized list of the performance characteristics to be evaluated, the recommended sample size and acceptance criteria for each characteristic, and finally a draft protocol. You can also include explicit instructions for the chatbot to allow the user to review the output between each step. Newer reasoning-tuned models perform some of these steps internally, but it’s still valuable to include explicit user-driven decomposition whenever a task requires domain expertise that the model may not reliably possess.
The final RISE component is a precise specification of the deliverable itself. Output expectations should address audience (e.g., “write for an audience of clinical pathologists and laboratory scientists”), register (e.g., “use formal academic prose and avoid colloquialisms”), format (e.g., free prose, structured table, JavaScript Object Notation [JSON], Markdown), length, and any required elements (e.g., executive summary, methods, and references). When output will be consumed downstream by software — for example, when a model-generated structured report will enter into a laboratory information system — it’s essential to use explicit schema specification, such as a JSON template.
Negative specification in prompts matters as much as positive. Including constraints can meaningfully alter model behavior. Examples include: “Do not invent citations,” “Do not summarize beyond what is supported by the provided documents,” and “Flag uncertainty explicitly rather than smoothing it over.” Singhal et al. showed in their evaluation of Med-PaLM that even capable medical LLMs require careful prompting and grounding to approach clinically acceptable accuracy on consumer health questions (5); explicit expectations are a primary mechanism by which the user enforces that discipline at the point of use.
The RISE framework is not a magic incantation; rather, it is a structured habit that aligns user input with the mechanisms by which transformer-based models can produce useable text. Although each element is supported by peer-reviewed evidence, applying them well to real-world laboratory problems — such as assay troubleshooting, validation planning, regulatory documentation, literature synthesis, or structured extraction from pathology reports — requires iterative refinement.
Thus, clinical laboratorians should treat prompt engineering as they would any other technical competency. Start with well-bounded tasks for which expert review is feasible to verify accuracy. Then systematically compare model output to that review, documenting any necessary prompt revisions and the effect that each revision has on performance. Periodically aggregate and evaluate the results to guide future efforts.
Like other skills, your prompting ability will strengthen with time: A laboratory team that systematically practices prompting will produce more accurate, reproducible, and defensible LLM-assisted output than the same team using ad hoc prompting. The most important next step is also the simplest: Try it for yourself!
Nicholas Spies, MD, is the medical director of applied AI and clinical chemistry at ARUP Laboratories and an assistant professor at the University of Utah in Salt Lake City. +Email: [email protected]