Top tips for prompting – A snapshot from The AI for Medical Affairs Course

First, understand what you are working with

The deck opens with a line that does more work than it appears to: AI is your new intern who spent their summer reading the entire internet, but who gets replaced after every task.

Both halves matter. The first explains why it can draft a congress summary or rewrite an abstract in plain language without any training from you: it was built on a very large body of text, medical writing included. The second explains why the same request can work well on Monday and badly on Wednesday, and why nothing you carefully explained last week carries over into today. Every new conversation starts from nothing.

Hold both halves in mind and the rest of the advice follows from them.
 

Give more context than feels necessary

Almost every disappointing output comes from a prompt that was too thin. People who get good results do the opposite of what feels efficient. They paste in the source document, the previous version, the strategy, the constraints they are working to, and their own half-formed thinking about what they want.

The deck shows real examples of this, and they are longer and messier than most people expect a prompt to be. That is the point. More is more: context, detail and examples.
 

Then be specific in the prompt itself

Once the context is there, six things are worth naming explicitly. The deck sets them out like this.

Most people specify one or two of these and leave the other four to chance. Naming all six takes an extra thirty seconds and gets you far closer to what you had in mind than any other adjustment.
 

Keep going, and break the work up

The first answer is a draft, not a result. The deck shows the kind of follow-up that works: "make it shorter", "more like this example", "make it friendlier". It also shows the kind that does not: "make it better", which gives the model nothing to act on.

For anything substantial, split the task. Asking for a full piece of web content with figures and graphs in one go produces something generic. Asking to go step by step, and checking each step, produces something you can use. The deck works through this with a real example.
 

Do not expect perfection

Two things in the deck are worth more than any prompting trick.

The first is what the authors call the jagged frontier: AI is unpredictably good and unpredictably bad at things that look equally difficult, and what fails today may work tomorrow. So, testing matters more than theorising.

The second is the 80:20 principle. If you wait for output you can publish untouched, you will conclude the tools do not work. If you accept 80 per cent and finish it yourself, you save real time on most Medical Affairs tasks. You remain the person who decides what is correct for your job, which is not a limitation of the tool so much as a description of your role.
 

What is in the deck
 

• The four-step framework in full, with real prompt screenshots rather than invented examples
• A worked example of breaking a content task into steps
• The six-part prompt anatomy as a reference slide
• Ten examples of AI applications in Medical Affairs, from literature monitoring and KOL insight analysis to compliance review and slide deck development
• 12 slides, from The AI for Medical Affairs Course, updated June 2026
   

Download the Deck