How Scientists Can Use AI Effectively: Practical Prompting Strategies for Research

In late 2024, artificial intelligence achieved PhD-level performance on scientific-reasoning benchmarks. By 2025, it was approaching the measurement ceiling on one of those tests. These advances are changing how researchers approach literature reviews, grant writing, data interpretation and scientific training.

In response, I developed an artificial intelligence module for incoming biomedical PhD students at Vanderbilt University in Nashville, Tennessee, where I direct AI initiatives for an interdisciplinary graduate pathway. A survey conducted as part of the course found that 81% of students had used AI tools for science. However, only 5% demonstrated baseline proficiency in writing prompts for AI. After completing the module, that figure rose to 48%.

Since then, I have taught courses on AI-assisted grant writing, integrated AI into scientific training in my laboratory and shared preprints on scientific AI use¹ and AI teaching methods² for others to adapt and reuse.

These experiences have shown me that you do not need a computational background to use AI effectively. My own research focuses on the biochemistry of DNA replication and repair. AI is a tool, and researchers can get the most from it by applying the scientific training they already possess.

The most useful strategy I have developed is to treat every prompt as an experiment and every AI response as a result. The following strategies can help researchers design more effective, reliable and reproducible AI experiments.

1. Connect AI Tools Before You Write a Prompt

An AI language model can reason only about the information it can access. Web search is the minimum requirement, but biomedical researchers should also connect AI systems to scientific literature databases such as PubMed and literature-analysis tools such as Scite and Elicit.

One way to do this is through the Model Context Protocol (MCP), an open standard that provides a consistent method for connecting AI applications to external systems. In different platforms, these connections may appear as “connectors” in Claude, “apps” in ChatGPT or MCP servers in other tools.

Choosing and configuring the right tools is part of designing the experiment. Providing an AI model with relevant, high-quality information gives it a better chance of producing useful scientific results.

2. Add Structure to Your AI Prompts

AI models generally process structured text more reliably than long, unorganized instructions. For many scientific tasks, Markdown formatting is sufficient and works across major AI models.

For more complex requests, XML can help organize information, particularly when working with Claude. JSON provides additional structure when precise formatting or data exchange is required.

Researchers use formatting conventions to navigate complex scientific documents, and AI models benefit from similar organization. Clear headings, numbered instructions, defined objectives and explicit output requirements can make prompts easier for models to interpret.

I have also developed a free tool to help researchers learn how to structure prompts for AI-assisted scientific work.

3. Let the AI Model Write the Prompt

When beginning a new laboratory assay, researchers often start with the manufacturer’s protocol. The equivalent approach for an AI system is to ask the model to create an effective prompt before using it to complete the task.

Describe your goal in plain language, ask the AI model to generate an optimized prompt and then run that prompt in a fresh conversation. This approach, known as meta-prompting, is often one of the fastest ways to improve AI performance.

For example, to summarize a scientific paper, you could ask an AI model:

“Generate a prompt optimized for [your AI tool] that asks it to summarize a scientific manuscript provided by the user. The prompt should instruct the AI tool to summarize the introduction, results and discussion separately, identify limitations acknowledged by the authors and critically evaluate whether the manuscript’s conclusions are supported by the evidence.”

You can also use a general-purpose meta-prompt or an interactive tool designed to create prompts for specific research tasks.

4. Start With Validated AI Prompts

Scientists typically rely on published protocols rather than reinventing every method from scratch. The same principle applies to AI prompting. Starting with prompts that have already been tested can save time and reduce unnecessary trial and error.

Effective prompts vary between AI models and can change as models are updated. A validated starting point helps researchers avoid rediscovering techniques that are already known to work.

I maintain a collection of prompts for common scholarly tasks, including literature summaries, manuscript critiques and writing feedback. These prompts are available through Zenodo, a general-purpose research repository, and an online prompt library. They are free to use and adapt.

5. Keep AI Conversations Short

AI systems have limitations that researchers should understand. Although conversational interfaces are convenient for humans, models can become less reliable as a conversation grows longer³. They also cannot simply forget information that has already been introduced into a conversation.

A more reliable approach is to develop a task within a conversation, then ask the AI model to condense the successful result into a single, well-structured prompt. Run that prompt in a new conversation rather than continuing indefinitely in the original thread.

6. Return to the Last Working Version

AI conversations can degrade during a single thread. The model may begin ignoring instructions, repeat an error or require the same correction several times. One reason is that AI models often give greater weight to the beginning and end of a conversation than to its middle⁴.

Treat this problem like troubleshooting a laboratory protocol. If the conversation begins producing unreliable results, branch from a point before the errors appeared by editing an earlier message. Another option is to summarize the important information and start a new conversation.

7. Make AI Outputs Reproducible

AI responses can vary between runs because language models generate outputs probabilistically. Running the same prompt several times in separate conversations can reveal which findings are consistent and which may be artifacts of a single response.

This is similar to conducting biological replicates. Repeated runs can also improve completeness by revealing different parts of the relevant answer space. Do not assume that an AI response is complete until repeated runs stop producing substantially new information.

For important research tasks, save the prompt, model name, date, connected tools and resulting outputs. This creates an audit trail and makes the process easier to reproduce or evaluate later.

8. Verify the Accuracy of AI-Generated Information

AI-generated content should be treated as a preliminary result, not as an authoritative source. Models can produce inaccurate statements, misinterpret scientific findings, invent citations or present uncertain conclusions with excessive confidence.

Researchers should verify important claims against primary literature, trusted databases and original experimental evidence. When using AI for literature reviews, check that cited papers exist and that they actually support the claims attributed to them.

For scientific writing, compare AI-generated summaries with the original manuscript. For grant proposals and research plans, independently review factual statements, statistics, references and descriptions of previous work.

Use Scientific Thinking to Improve AI Research Workflows

The most effective way to use AI in science is not to treat it as an infallible expert. Instead, approach each interaction as a research experiment:

• Define the objective: Clarify exactly what you want the AI system to accomplish.
• Control the inputs: Provide relevant literature, background information and formatting instructions.
• Run repeated trials: Test the prompt across multiple conversations or model runs.
• Evaluate the results: Check accuracy, completeness, relevance and scientific reasoning.
• Document the workflow: Save prompts, outputs and model details for future reference.
• Refine the method: Improve the prompt based on observed failures and successes.

Researchers do not need advanced programming skills to benefit from AI. They need the same habits that support good science: clear experimental design, careful controls, critical evaluation, reproducibility and independent validation.

Used in this way, AI can become a powerful research assistant for biomedical scientists, graduate students and other researchers. The goal is not to replace scientific judgment, but to extend what researchers can accomplish with the knowledge and methods they already possess.