How to Get Useful Answers from AI: Prompting and Context

LLMs will always get you an answer. This is a great thing in principle but a big pitfall in practice if you do not focus on your prompt. As with all things in technology, this follows the same principle of garbage in, garbage out (GIGO). Poor AI answers almost always trace back to poor questions, incorrect or missing context, or just vague instructions. A lot of other things factor into this as well, such as the model’s quality and weights, but as the AI wave has neared half a decade now, we have to agree that the models are all strong enough that there is little room for error on most simple tasks. How to prompt AI well is less about special syntax and more about one habit: give the AI the context it would need to actually help you. In practice we tend to just send a short, half-baked sentence and expect it all to work quickly. But you need to set a tool up for success, and through this post we will discuss how you can do it.

This is the third post in our series on using AI in pharma, after Part 1 on what an AI agent is and Part 2 on the four ways to use AI.

What is a prompt?

A prompt is not just the question you type. It is everything you hand the AI: the instruction, the relevant documents, the format you want back, the constraints it must respect, and any examples of what good looks like. Think of how you would talk to someone in real life when handing over a task. You wouldn’t just say, “write an essay on such-and-such a topic.” It is precisely how you ought to talk to AI. Say “summarise this protocol deviation” and you get a guess, but hand over the deviation text, the governing SOP, the escalation criteria, and your summary format, and you get something usable. The AI is not reading your mind and it does not have your institutional knowledge. It has exactly what you give it, so write the prompt like the briefing it is.

What is context?

Context is the specific material the AI needs to do the task correctly. Three things carry the most weight: the relevant documents, the output requirements, and the constraints. Paste the deviation, the SOP section, or the regulation instead of describing them. Tell the AI whether you want a three-paragraph summary, a bullet list, or a structured QA form, since those are not interchangeable. And spell out the constraints: the regulatory rules, the escalation criteria, and the terminology that separates your context from the general case. The more of this you provide, the less the AI has to guess, and the less room there is for it to invent something plausible but wrong.

Why vague asks get vague answers

A vague ask forces the AI to fill in the blanks with assumptions. Sometimes those assumptions are reasonable. Often they are not, especially in a specialised domain like clinical research, where the difference between a protocol deviation and a protocol amendment matters, where “escalation” means something specific, and where the output format may need to satisfy a QA reviewer. “Summarise this protocol deviation” could mean a hundred different things. Summarise for what audience? For what purpose? Does the summary need to flag whether escalation is required? Does it follow a specific format? What SOP governs this? The AI does not know any of that unless you tell it. So it will produce something that looks like a summary, in whatever format seems sensible to a general-purpose system. That output may be wrong for your context even if it is technically correct in isolation. Vague inputs also raise the risk of the AI hallucinating, filling gaps in its knowledge with plausible-sounding content. The more specific the context, the less room there is for the AI to invent.

A quick test: if you would not hand the task to a new colleague with only that one sentence, do not hand it to the AI either.

Before and after: a real prompt comparison

Here is the same task done two ways.

Weak prompt:

“Summarise this protocol deviation.”

The AI produces a general summary. It may be 100% there or it may be grossly incorrect. There is no way to make this even slightly deterministic.

Strong prompt:

“Here is the protocol deviation: [text]. Here is the relevant section of our SOP governing deviation escalation: [text]. The output format we use for deviation summaries is: [format]. Please summarise the deviation, flag whether it requires escalation per the SOP criteria, and output in the format provided.”

The AI now has the deviation, the governing SOP, the escalation criteria, and the required format. You go from hoping for the right answer to something you can use with minimal changes. That is the strength of spending a little extra time on adding context.

Here is that same structure as a skeleton you can copy and adapt to any task:

Reusable prompt skeleton
Here is the [document or task]: [paste the text].
Here is the rule or standard that applies: [paste the SOP, regulation, or guideline].
The output format I need is: [describe or paste the format].
Please [do the task], flag [any conditions to check for], and output in the format above.

Three simple rules of thumb

1. Lead with the task, then the material. State what you want before you paste documents. The AI reads top to bottom, so put the instruction where it will be found first.
2. Paste the source, don’t describe it. Descriptions carry your interpretation; the source material gives the AI the exact text to reason against.
3. Show what good looks like. One solid example of a finished summary is often worth more than a paragraph describing the format.

At Appsilon, we help you meet LLMs and AI tools where you are with our four pillars. Follow the link to learn more.