I made every mistake trying to prompt these newer reasoning-first LLMs—O1, O3, DeepSeek.

Nested tasks. Overexplaining. Chain-of-thought overload.

They don’t respond like older models.

They don’t need more words.

They need better direction.

Here are 6 field-tested principles—plus real examples—to help you get consistently high-quality answers from these models for real-world clinical research tasks.

1️⃣ Keep prompts simple and direct

Clarity beats complexity.

Imagine you’re screening a new weight-loss drug trial in type 2 diabetes. You want a quick overview of the endpoints to assess fit.

❌ Don’t do this:

“Could you provide a multi-layered analysis of the protocol, explaining endpoints, statistical methods, criteria, timeline, and condense it into bullets?”

You’ll get clutter. No clarity.

✅ Try this instead:

“Summarize the main endpoints of the new diabetes medication trial in bullet points.”

Clean question. Clean answer.

What to do:

Break it down.

Ask for endpoints first.

Then methods.

Then criteria.

One prompt. One task.

2️⃣ Skip chain-of-thought prompting

These models don’t need your internal monologue.

❌ Don’t do this:

“First, list risk factors for weight gain. Then explain how each causes insulin resistance. Then predict long-term complications.”

That’s three prompts pretending to be one.

✅ Better:

“What are the most significant risk factors for weight gain in diabetic patients, and how do they influence long-term complications?”

You’ll get sharper, cleaner reasoning without the over-directing.

What to do:

Ask the outcome you want.

Let the model figure out the logic path.

3️⃣ Use structure and delimiters

If you want a structured answer, show what structure looks like.

❌ Don’t do this:

“List observed side effects from the pilot study in a structured way.”

“Structured” is vague.

✅ Try this:

“Provide observed side effects in JSON format:

go
CopyEdit
{
  'MildSideEffects': [],
  'ModerateSideEffects': [],
  'SevereSideEffects': []
}
```"*

You’ve given the model a blueprint. It knows the exact containers for each category.

What to do:

Use markdown, bullet points, tables—whatever format you need.

Give the format. Let the model follow the template.

For instance, instead of just saying “structured output,” write:

4️⃣ Start zero-shot. Few-shot only if needed.

Most of the time, you don’t need examples. You just need a clear ask.

❌ Don’t start with this:

“Convert this study conclusion into layman’s terms. Example 1: [original + layman version]. Example 2: [original + layman version]. Now apply to this.”

Too much scaffolding upfront.

✅ Start here:

“Convert the conclusion of this obesity study into everyday language for a non-medical audience.”

Let the model try.

If it’s off?

Then you can show examples.

What to do:

Zero-shot → check result

Few-shot → only if needed

5️⃣ Add context

Long prompting doesn’t work with one exception – these models work better with the context you give them.

❌ Don’t do this:

“Design a study to measure treatment response.”

Too vague. You’ll get a generic, one-size-fits-none output.

✅ Try this:

“We’re researching immunotherapy for stage II breast cancer in women aged 50–65. The cohort includes 150 patients post-partial mastectomy. Design a 12-month study to measure treatment response based on recurrence and quality of life.”

Now the model understands your study population, disease state, and goal.

What to do:

Define your disease.

Describe your cohort.

State your objective.

Even if it makes the prompt longer—

Precision in → Precision out.

6️⃣ Provide constraints

These models perform best when you define the edges.

❌ Don’t do this:

“Design a research plan for a weight-loss intervention in obese patients.”

Too open-ended. You might get a full grant proposal.

✅ Try this:

“Design a research plan for a weight-loss intervention in obese patients, using a $30,000 budget, over 6 months, focused on nutrition counseling and daily physical activity.”

Now the model works within your world—not fantasy-land.

What to do:

Set a budget.

Set a timeline.

Set the scope.

Just like real research.

Prompting smarter = Researching better

These reasoning-first models are powerful—but only if you respect how they think.

Your prompt is your protocol.

Clean in → Clean out.

Confused in → Confused out.

If you’re using these models in your research workflow, experiment with these 5 shifts.

You’ll spend less time editing and more time thinking.

Tried something else that worked better?

I’d love to hear what you’re learning.

P.S: Google’s recently launched Deep Research is free (5 reports/month) – you can try it here by choosing “Deep Research” from the options: https://gemini.google.com/app