Prompting Guide

Context

The AI Landscape in 2026

A lot has changed since the early days of “just ask ChatGPT”. Understanding the current environment helps you choose the right tool, set the right expectations, and get far better results.

1

Reasoning Models are Mainstream

Models like Claude Sonnet 4 and GPT-5 now reason before answering; they can plan, check their own logic, and work through multi-step problems. This changes how you prompt: you can ask for structured reasoning, not just answers.

2

Multiple Models, One Platform

Imperial’s dAIsy gives you access to Claude, GPT-5, and DeepSeek under one secure, data-protected login. Different models have different strengths. Knowing which to use when matters.

3

Prompts are Now Systems

The shift from “ask a question” to “design a workflow” is complete. Practitioners now think about system prompts, multi-turn conversations, personal agents and reusable templates, not single queries.

4

Multimodal is Standard

Text, images, PDFs and documents can all be fed into modern models. For medicine: annotating histology images, summarising papers, analysing data tables, all now possible within a single prompt session.

5

Hallucinations Haven’t Gone Away

Models are more accurate but still fabricate citations, statistics, and clinical details with great confidence. For academic and clinical work, verification remains non-negotiable, especially for anything patient-facing.

6

Data Protection Still Matters

General AI tools (ChatGPT free, Gemini) may use your inputs for training. Imperial’s dAIsy is configured so prompts are NOT used to train external models; use it for anything sensitive to your work.

Framework

The CRAFT Framework

Most mediocre AI outputs come from vague prompts. The CRAFT framework gives you five levers to pull; use all five and you’ll see a dramatic improvement in quality, consistency and usefulness.

C

Context

Tell the model who you are, what you’re working on, and any relevant background. The more specific the context, the more tailored the response.

Weak: Summarise this paper. Strong: I am a clinical pharmacologist preparing a literature review on antibiotic stewardship for a postgraduate teaching module. Summarise this paper, highlighting clinical implications and any limitations relevant to acute care settings.

R

Role

Assign the model a relevant expert persona. This “primes” it to draw on domain-specific knowledge and apply appropriate professional judgment.

Weak: Help me write a grant abstract. Strong: You are a senior academic with expertise in translational medicine and a track record of successful NIHR funding. Review and rewrite this grant abstract to strengthen the impact and feasibility sections.

A

Action

Be precise about what you want done. Use active, specific verbs. Break complex tasks into numbered steps. Avoid open-ended asks that force the model to guess.

Weak: Write something about sepsis for students. Strong: Create a 5-question formative MCQ set on the early recognition and management of sepsis for Year 3 medical students. Each question should include: stem, 5 options, correct answer, and a one-sentence explanation.

F

Format

Specify exactly how you want the output structured. Length, headers, bullet points, tables, tone; the model will match your specification if you give one.

Examples of format instructions: “Respond in under 300 words.” “Use a table with columns: Drug | Dose | Caution.” “Write in the style of a NEJM editorial.” “Use bullet points under three headings: Background, Methods, Clinical Takeaway.” “Avoid jargon, write for a general audience.”

T

Test & Iterate

The first response is rarely the final answer. Strong prompt engineers treat the conversation as a loop: review, refine, and ask the model to improve its own output.

Useful follow-up prompts: “That’s good, now make the tone more accessible for undergraduate students.” “Critique your own response and identify any weaknesses or gaps.” “Restructure this as a table.” “Are there any clinical inaccuracies or anything I should verify?” “Now write a shorter version (under 150 words) for a slide.”

Level Up

Techniques That Actually Move the Needle

These techniques consistently produce better results on complex tasks, particularly the kind of analytical, writing and reasoning work common in medicine and research.

Chain-of-Thought

Make the model show its reasoning

Asking the model to reason step-by-step before answering significantly improves accuracy on complex tasks, and lets you spot errors in the logic.

How to trigger it: “Before answering, think through this step by step.” “Walk me through your reasoning, then give your conclusion.” “Work through the differential diagnosis systematically, then rank by likelihood given this clinical picture.” Medical example: “A 62-year-old presents with sudden-onset chest pain, diaphoresis and nausea. Think through the differential diagnosis step by step, considering cardiac, pulmonary, gastrointestinal and other causes, then rank by priority and suggest immediate management.”

Few-Shot Prompting

Give examples before the task

Show the model 1–3 examples of what you want before asking it to do the task. Hugely effective for formatting consistency, matching a writing style, or producing structured outputs.

Pattern: Example → Example → Now you do it “I’m creating learning objectives for medical teaching. Here are two examples of well-written objectives: 1. ‘Describe the pathophysiology of type 2 diabetes and its effect on major organ systems.’ 2. ‘Apply the ABCDE framework to assess and manage a deteriorating patient.’ Now write four similar learning objectives for a session on acute kidney injury in critical care.”

Self-Check Prompting

Ask the model to critique itself

One of the most underused techniques. After getting a response, ask the model to review its own output for errors, omissions or weaknesses, then improve it.

Use these follow-ups after any response: “Now review your answer. Are there any clinical inaccuracies, outdated guidelines, or important caveats I should check?” “What assumptions did you make in that response that might not hold in a real clinical or academic context?” “Rate your own confidence in each section of that response on a scale of 1-5, and flag anything I should verify.”

Audience Targeting

Specify exactly who this is for

The same content explained for a Year 1 student, a specialist registrar, or a lay patient requires completely different language, depth and structure. Specifying the audience eliminates guesswork.

Audience modifiers you can drop into any prompt: “…for a first-year medical student with no prior clinical experience” “…for a specialist registrar preparing for MRCP” “…for a lay patient who has just received this diagnosis” “…for an academic audience familiar with NICE guidelines” “…for a postgraduate researcher with a background in molecular biology but limited clinical experience”

Persona + Constraint

Combine a role with explicit limits

Pair a role assignment with specific constraints: what the model should NOT do. This prevents the model from drifting outside the boundaries you need.

Template: “You are [expert role]. Your task is [specific task]. Constraints: [list what to avoid]. Output format: [specify structure].” Example: “You are a medical ethics consultant. Your task is to analyse the ethical dimensions of this case study for use in teaching. Constraints: do not give a definitive ‘right answer’, present the competing considerations fairly. Do not reference specific UK legislation by name as this may become outdated. Output: three paragraphs covering autonomy, beneficence and justice.”

Applied

For Medicine & Health Academics

The real value of AI isn’t writing generic text; it’s accelerating the high-effort tasks that consume your time. Here are the use cases where prompting makes the biggest difference in your context.

Literature Synthesis

Feed multiple abstracts or key sections into dAIsy and ask it to synthesise themes, identify gaps, or structure a narrative for your review; saving hours on first drafts.

⚠

Always verify: AI cannot access papers you don’t provide. It may generate plausible-sounding but fictitious citations. Use it for synthesis only on text you’ve given it.

Teaching Content Creation

Generate MCQ banks, case studies, learning objectives, lecture outlines and formative feedback, then refine with follow-up prompts. Use few-shot examples for consistent formatting.

Best for: rapid prototyping of materials, variation of scenarios, adapting content for different year groups.

Grant & Publication Writing

AI is excellent at restructuring arguments, tightening impact statements, and translating dense technical language into accessible lay summaries. It works best as an editor, not a first author.

Try: “You are an expert grant reviewer. Critique this impact statement for clarity, specificity and fundability.”

Clinical Reasoning Practice

Use AI to generate case vignettes with varying complexity, create worked examples of ABCDE assessment, or produce structured differential diagnoses. Always validate outputs against current guidelines.

⚠

Clinical outputs must always be reviewed by a qualified clinician before any patient-facing use.

Research Design Support

Brainstorm methodology options, stress-test your research questions, draft ethics application sections, or generate participant information sheets in plain English.

Note: If using patient-identifiable data or sensitive information, dAIsy is the only appropriate tool, and review your institution’s research ethics requirements first.

Student Feedback Drafting

Provide anonymised student work and a marking rubric. Ask AI to draft developmental feedback aligned to the criteria. Review and personalise before sending; this should augment your judgment, not replace it.

Always: remove any identifying information before pasting into any AI tool.

Imperial Platform

Making the Most of dAIsy

dAIsy is Imperial’s secure, multi-model AI platform. It’s not just a browser wrapper for ChatGPT; it has features that make a significant difference to how useful it is for academic work.

What dAIsy gives you that consumer tools don’t

Data protection: prompts and files are not used to train external models. Access to multiple models (Claude Sonnet 4, GPT-5, DeepSeek) under one Imperial login. Ability to create Personal Agents, essentially customised AI assistants with persistent instructions and uploaded documents.

Access dAIsy (Imperial login required) →

Choosing the Right Model

Claude Sonnet 4: Best for: nuanced writing, medical reasoning, long documents, balanced outputs
GPT-5: Best for: coding, structured data tasks, when you want a second opinion
DeepSeek: Best for: technical analysis, research-heavy tasks

Personal Agents

Create a custom agent with a system prompt that persists across every conversation. Examples:

“Teaching assistant” pre-loaded with your module handbook
“Grant reviewer” with your funding body’s criteria
“Literature buddy” with your research focus

System Prompt for Your Agent

When creating a Personal Agent, your system prompt sets the persistent context. A good template:

“You are an AI research assistant for a [specialty] academic at Imperial London. My focus is [your area]. Always apply clinical accuracy standards. Flag any claims I should verify.”

Critical Thinking

The Verification Layer

AI in 2026 is significantly more accurate than it was in 2023, but “significantly more accurate” still means it makes confident errors. In medicine and research, the stakes of unchecked errors are high.

Things AI Gets Wrong Most Often

Citations and references: frequently fabricated with believable author names, journal titles and DOIs
Specific statistics: incidence rates, drug doses, trial results, often slightly wrong
Very recent guidelines: training data has a cutoff; NICE, BNF and SIGN updates may not be reflected
Nuanced clinical judgment: AI gives answers, not the uncertainty that characterises good clinical practice
Local policies and protocols: AI has no knowledge of your trust or department’s specific procedures

⚠

Always verify before publishing, teaching or using clinically: any statistics or numerical claims, any specific citations, any drug interactions or dosing information, any reference to current clinical guidelines.

Build Verification Into Your Prompts

Ask the model to flag its own uncertainty before you have to find it yourself.

“After your response, add a section titled VERIFY: listing any specific facts, statistics or citations that I should independently confirm.” “Flag any information in your response that may have changed since 2024.” “Note any clinical recommendations that are context-dependent or where practice may vary.”

The Four-Question Check

Before using any AI output, ask yourself:

Have I checked every statistic against a primary source?
Have I verified every citation actually exists?
Is anything here patient-facing or used in assessment?
Would I be comfortable putting my name on this without checking?

Grab and Go

Prompt Template Library

Copy and adapt these templates. The brackets [ ] mark what you should customise: everything else is structured to get consistently good results.

Literature Synthesis

Research

You are a systematic review specialist in [your field].
I will paste [N] abstracts below. For each one:
1. Summarise the key finding in 1-2 sentences
2. Note the study design and sample size
3. Flag any methodological limitations

Then provide an overall synthesis paragraph identifying:
- Common themes across the papers
- Contradictions or gaps in the evidence
- The strongest conclusion the evidence supports

[PASTE ABSTRACTS HERE]

VERIFY: list any claims I should cross-check against the source papers.

Tip: paste up to 5-6 abstracts at once in dAIsy. For full papers, upload the PDFs via the attachment feature.

MCQ Generation

Teaching

You are an experienced medical educator writing formative assessment questions.
Create [N] single-best-answer MCQs on the topic of [TOPIC].
Target audience: [Year group / level, e.g. "Year 3 MBBS students"]

For each question:
- Clinical stem (realistic scenario, 2-4 sentences)
- 5 answer options (A-E), one correct, four plausible distractors
- Correct answer
- Brief explanation (max 3 sentences) why the answer is correct
  and the most tempting distractor is wrong

Difficulty: [foundation / intermediate / advanced]
Avoid: [any topics or drugs you want excluded]

Grant Impact Statement

Research

You are a senior research grant advisor with expertise in [NIHR / Wellcome / MRC / other].
Review the following draft impact statement for my [grant type] application.

Critique it on:
1. Clarity: is the significance immediately obvious to a non-specialist panel member?
2. Specificity: are the claimed benefits concrete and measurable?
3. Feasibility: does the impact claim seem realistic given the project scope?
4. Patient/public benefit, is this made explicit?

Then rewrite the statement incorporating your suggestions.
Keep to [word limit] words.

[PASTE YOUR DRAFT IMPACT STATEMENT]

Patient Information Leaf (Plain English)

Research / Clinical Ed

You are a health literacy specialist. Rewrite the following clinical/research text
as a patient information document.

Requirements:
- Reading age: no higher than 12 years (Flesch-Kincaid)
- Avoid jargon; explain any medical terms you must include
- Use short paragraphs (max 3 sentences)
- Include: What this is, Why it matters, What happens next
- Tone: warm, reassuring, informative

[PASTE SOURCE TEXT]

After completing the document, flag any section where
simplification may have changed the clinical meaning ,
these will need clinical review before use.

Case Study Generator

Teaching

You are a clinical educator designing case-based learning for [YEAR / PROGRAMME].
Generate a case study on [CLINICAL SCENARIO / CONDITION].

Structure:
1. Presenting complaint and history (realistic, 150 words)
2. Examination findings (relevant positives and negatives)
3. Initial investigations with results
4. Three discussion questions progressing from recall to application
5. Key learning points (maximum 5 bullet points)
6. Suggested further reading topics (do not fabricate citations)

Complexity: [foundation / intermediate / advanced]
Avoid: any patient-identifiable details; keep demographics generic.

After generating, ask: “Now create an answer guide for facilitators for the three discussion questions.”

Where to Go Next

This guide focuses on practical technique. For everything else: the fundamentals, ethical frameworks, events and support, your colleagues and Imperial central have you covered.

Getting More from AI

The AI Landscape in 2026

Reasoning Models are Mainstream

Multiple Models, One Platform

Prompts are Now Systems

Multimodal is Standard

Hallucinations Haven’t Gone Away

Data Protection Still Matters

The CRAFT Framework

Context

Role

Action

Format

Test & Iterate

Techniques That Actually Move the Needle

For Medicine & Health Academics

Literature Synthesis

Teaching Content Creation

Grant & Publication Writing

Clinical Reasoning Practice

Research Design Support

Student Feedback Drafting

Making the Most of dAIsy

What dAIsy gives you that consumer tools don’t

Choosing the Right Model

Personal Agents

System Prompt for Your Agent

The Verification Layer

Things AI Gets Wrong Most Often

Build Verification Into Your Prompts

The Four-Question Check

Prompt Template Library

Literature Synthesis

MCQ Generation

Grant Impact Statement

Patient Information Leaf (Plain English)

Case Study Generator

Where to Go Next

dAIsy Platform

Faculty GenAI Site

dAIsy Guidance (ICT)

AI & Education Hub

Library AI Guidance