AI can sound confident even when it is wrong, outdated, or inventing a reference. Doctors need a habit that separates draft writing from clinical verification.

This chapter gives you prompts that make AI label what it knows, what it inferred, and what you must verify from a reliable source.

---

What Problem This Solves

Medical AI outputs can fail in subtle ways:

• outdated guideline recommendations
• fabricated citations
• incorrect drug advice
• missing contraindications
• overconfident wording
• claims that sound evidence-based but are not checked

The solution is not “ask AI to cite sources” and trust the answer. The solution is to force a verification workflow.

---

The Verification Rule

For clinical content, ask AI to divide statements into four buckets:

1. Provided in my prompt: facts directly supplied by you.
2. General medical knowledge: common background information that still needs professional review.
3. Inference: reasonable but not directly provided.
4. Needs verification: guideline, dosage, drug interaction, legal, or time-sensitive claim.

Anything in bucket 4 must be checked by the doctor against a reliable source before use.

---

Prompt 1: Claim Classifier

Review the medical content below.

Create a table with these columns:
1. Claim
2. Category: PROVIDED / GENERAL KNOWLEDGE / INFERENCE / NEEDS VERIFICATION
3. Why it belongs in that category
4. What source or clinician check is needed

Rules:
- Do not invent citations.
- Do not say a claim is verified unless the source text is provided.
- Flag drug dose, contraindication, interaction, guideline, legal, and emergency
  advice as NEEDS VERIFICATION.

Content:
[PASTE AI OUTPUT HERE]

---

Prompt 2: Guideline-Sensitive Rewrite

Use this when patient-facing content might include current clinical recommendations.

Rewrite the content below for patient education, but remove or soften any
claim that depends on current clinical guidelines unless I provide the
source.

For every removed or softened claim, list it under "Doctor must verify".

Keep:
- simple language
- practical patient instructions
- clear red flags
- no medication dose changes

Content:
[PASTE CONTENT]

---

Prompt 3: Source-Aware Research Question Builder

Use this when you need to check guidelines yourself.

Act as a clinical research assistant.

Turn my clinical question into a source-checking plan.

Clinical question:
[WRITE QUESTION]

Return:
- exact question in PICO format if applicable
- what type of source is appropriate: guideline, drug label, systematic review,
  local protocol, textbook, or specialist opinion
- key search terms
- what details must be checked before applying to a patient
- what should not be delegated to AI

Do not provide a final clinical recommendation.

---

Prompt 4: Citation Hygiene Check

Use this when AI has supplied references.

Check the citations or references in this AI output.

For each reference:
- mark as COMPLETE / INCOMPLETE / SUSPICIOUS
- identify missing details such as author, organization, title, year, journal,
  URL, or guideline name
- state whether I need to manually verify the reference before using it

Important:
- Do not assume a reference exists.
- Do not create replacement citations.
- If uncertain, mark SUSPICIOUS.

AI output:
[PASTE OUTPUT WITH REFERENCES]

---

What Doctors Should Verify Manually

Always verify:

• drug dose, frequency, duration, route, renal adjustment, pregnancy safety
• drug interactions and contraindications
• antibiotic choice
• anticoagulation or antiplatelet decisions
• vaccination schedules
• emergency warning signs
• cancer screening recommendations
• consent risks and legal language
• any claim involving a recent guideline
• local availability, cost, or referral pathway

AI can help organize the question. It should not be treated as the source of truth.

---

Mini Case: Bad Verification

Bad prompt:

Give me the latest guideline for hypertension treatment and cite sources.

Why this is weak:

• “latest” may be ambiguous
• AI may fabricate citations
• treatment depends on patient context
• no instruction to separate verified from unverified claims

Better prompt:

I need to update a patient education handout about hypertension lifestyle
care. Do not provide drug treatment recommendations.

List the claims in my handout that may depend on current guidelines and
therefore need manual verification. For each claim, tell me what type of
source I should check. Do not invent citations.

Handout:
[PASTE HANDOUT]

---

Practice Check

Before marking this complete:

• take one AI-generated medical output from your prompt library
• run the Claim Classifier
• identify at least three claims needing verification
• rewrite the content after removing or softening unverified claims
• save the verification prompt in your prompt library