AI Guides › Playbooks
By Nigel Guy · 6 min read
Most people ask an AI about a health worry the way they would type it into a search box: three words, no context, then judge the answer by whether it sounds reassuring. A vague question gets a vague, hedged reply, and the hedging gets blamed on the tool. The better habit is to give Claude what a sensible friend with medical knowledge would need, and then check what comes back.
The rule: Ask Claude to help you understand and prepare, never to diagnose or decide, and give it a brief that makes a specific answer possible.
The durable, verifiable part is this. Anthropic's published constitution for Claude lists, among behaviours it does not want, giving poor responses to "medical, legal, financial, psychological, or other questions out of excessive caution", and adding warnings or caveats that are not needed. It describes the ideal as a brilliant friend with a doctor's knowledge who gives "real information based on our specific situation rather than overly cautious advice driven by fear of liability". That is a stated aim, not a guarantee of any single answer.
In January 2026 Anthropic also announced consumer health features: in the US, Pro and Max subscribers can connect health records and lab results through HealthEx and Function (in beta), and Apple Health and Android Health Connect integrations were announced as rolling out in beta through the mobile apps. Claude can summarise medical history, explain test results in plain language, spot patterns in fitness and health metrics, and help you prepare questions for appointments. Anthropic says you opt in, choose what to share, can disconnect at any time, and that health data is not used to train models.
What we could not verify: a specific "admission" by Anthropic that earlier answers were useless, or any measured percentage for how often Claude refused or hedged on health questions before. We found no primary source for either, so this guide states no number. If you have seen one, ask where it came from.
| Reader | What applies |
|---|---|
| Anyone on any plan, anywhere | The general behaviour: Claude should engage with health questions rather than deflect. Free to try. |
| US Pro or Max subscribers | The connectors for records, labs and phone health data (beta). |
| UK readers | We found no official announcement that the record and phone-health integrations are available outside the US. Check Settings in your own account rather than assuming. |
Cost at time of writing: Pro is $20 a month (about £15 at time of writing), Max starts at $100 a month (about £75 at time of writing). Check the £ price at checkout. One wrinkle: Anthropic's announcement says Pro and Max, but its HealthEx tutorial page describes setup for Max subscribers. Confirm which plan your account needs before paying.
Fill in five fields before you ask. It takes two minutes and is the whole mechanism.
| Field | What to write | Why it matters |
|---|---|---|
| 1. Situation | Age band, relevant conditions, medicines, what you have noticed and for how long | Without it, Claude can only give a generic answer |
| 2. Question | One question, phrased as "help me understand" or "help me prepare" | Keeps the answer on one job |
| 3. Evidence | Paste the actual result, leaflet text or reading, with units and reference ranges | Claude explains what is there instead of guessing |
| 4. Output | Plain-English explanation, a list of questions for your clinician, or both | Gives you something to carry into an appointment |
| 5. Red lines | Ask it to say which symptoms mean you should seek urgent help | Makes the safety information specific rather than a boilerplate disclaimer |
You are a careful, plain-speaking health information assistant helping an adult in the UK prepare for a conversation with their own clinician. You are not their doctor and you will not diagnose.
My situation: [AGE_BAND, RELEVANT_CONDITIONS, CURRENT_MEDICINES, SYMPTOMS_AND_DURATION]
My question: [ONE_QUESTION]
The evidence I have: [PASTED_TEST_RESULT_OR_TEXT_WITH_UNITS_AND_REFERENCE_RANGE]
Please do the following, in order:
1. Restate my situation in two sentences so I can check you have understood it.
2. Explain what the evidence means in plain English, separating what is well established from what is uncertain.
3. List the most common explanations, ordered from most to least likely, and say what would make each more or less likely.
4. Tell me which specific symptoms or changes would mean I should contact NHS 111 or emergency services rather than wait.
5. Give me five questions to ask my GP or pharmacist.
Rules: If anything I have given you is missing or ambiguous, ask me before answering rather than guessing. Do not give medicine doses or tell me to start or stop a medicine. Say clearly when something depends on a test or examination you cannot do.
Before you answer, check that every claim is tied to something I told you or to general medical knowledge, and flag anything you are unsure about.
Fill in the square-bracket fields. Remove anything identifying, such as your name, NHS number or address, unless you have deliberately connected a record.
Imagine a 45-year-old with a routine blood test showing a cholesterol figure marked "high" and no symptoms. The bare question "is my cholesterol bad?" invites a generic paragraph. With the card, they paste the full lipid panel with units and ranges, state they take no medicines and that a parent had a heart attack in their fifties, and ask for an explanation plus questions for the GP. A good reply separates the numbers that are flagged from the ones that are not, explains that a single reading is not a risk score, notes that family history matters and that clinicians use wider risk tools, and ends with questions such as which risk estimate the surgery uses and whether a repeat test is sensible. That is preparation, not a verdict, and it is what you can actually use.