In late 2024, artificial intelligence reached PhD-level performance on scientific-reasoning benchmarks, and in 2025 it approached the measurement ceiling for one such test. In response, I developed an AI module and taught it to incoming biomedical PhD students at Vanderbilt University in Nashville, Tennessee, where I direct AI initiatives for the interdisciplinary graduate pathway.
As part of the course, I surveyed the students and found that 81% had used AI tools for science. Yet, only 5% met a baseline level of proficiency in writing prompts for AI; after the module, 48% did. Since then, I have taught classes on AI-assisted grant writing, made AI a core part of the scientific training in my laboratory and shared preprints on scientific AI use1 and teaching methods2 for others to adapt and reuse. All of this taught me that you don’t need a computational background to use AI well — my research focuses on the biochemistry of DNA replication and repair. AI is a tool: to get the most from it, you need to apply the scientific training that you already have.
The single most useful strategy I have developed is to treat each prompt as an experiment and the output as a result. Here are ten tips I’ve learnt for running those experiments effectively.
Connect tools before you prompt
An AI language model can reason only about what it can reach. Web search is the minimum; for biomedical work, granting the model access to the literature, using PubMed or through literature tools such as Scite and Elicit, grounds outputs in published work. This can be done by using an open standard called the Model Context Protocol (MCP), which provides a standardized way to connect applications to external systems. It appears as ‘connectors’ in Claude, as ‘apps’ in ChatGPT and as MCP servers in other tools. Setting up the right tools is part of designing the experiment, so you can give it the best chance of being successful by arming the model with as much information as possible.
Add structure to your inputs
Models parse unstructured text less cleanly than structured text, so sometimes it helps to write prompts in a structured format. For example, Markdown formatting works across all major models and is usually enough for prompting; XML helps to structure more-complex inputs (and is particularly effective with Claude); JSON adds more rigour when you need it. Just as scientists use formatting conventions to navigate complex text, so do models. I’ve built a tool to help teach people how to do this, which is free to use.
Let the model write the prompt for you
When running a new assay, you start from the manufacturer’s protocol. The equivalent for an AI system is built into the model itself: you can describe what you want in plain language and ask the model to write the prompt that will achieve it, then run that prompt in a fresh conversation. This is called meta-prompting, and because the model knows how to construct effective prompts, it is the fastest way to get a model working productively. For instance, to summarize a paper, you might ask the model to: “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.” Alternatively, you can use a general-purpose meta-prompt, or an interactive tool that writes one for you.
Use validated prompts
Just as researchers lean on published protocols rather than reinventing every method, you can start from prompts that others have validated already. The types of prompt that are effective differ across models and shift over time, so a vetted starting point saves you from having to rediscover what already works. I maintain a set of prompts1 for common scholarly tasks (literature summaries, manuscript critique and feedback on writing) on Zenodo — a general-purpose research repository — and as an online library, which is free to use and adapt.
Keep conversations short
Like any experimental system, this one has limits that are worth recognizing. The conversational interface exists because it suits humans, but models become less reliable as a conversation grows3, and they cannot forget anything that’s already been said. A useful strategy is to develop a task across a conversation, then get the model to condense the result into a single prompt that you then run in a fresh conversation.
Return to the last working version
Conversations can also degrade mid-thread: for instance, the model might seem to ignore instructions, or you could have to correct the same error repeatedly. This happens because models prioritize the start and end of a conversation over the middle4. Treat this like troubleshooting a protocol. Branch from a point before the errors appeared by editing a message from earlier in the conversation. Alternatively, summarize the stuff that matters and start afresh.
Make sure outputs are reproducible
Outputs vary across runs because AI models sample probabilistically. Running the same prompt several times in separate conversations reveals what is robust and what is an artefact. This is the AI-model equivalent of conducting biological replicates, and it improves completeness: the more times that you run a prompt, the more of the relevant answer space you see. Do not assume completeness until repeated runs stop surfacing new answers.
