Ten minutes of setup that protect every task on this map. Think of it as hand hygiene for AI use: nobody applauds it, and everything downstream depends on it.
This card has no clinical case and no exercise. It is the only page here that runs before a patient is involved, and the only one you complete once rather than per visit.
Nothing, yet. This card is about what you do before the tool does anything at all.
Which environment you are in, and whether the one you picked is the wrong place for patient data. That is invisible from inside the chat window.
The interface can look identical across three completely different destinations for your data. Consumer free, consumer paid, and an institutional deployment under contract are hard to tell apart on screen. The contract is what separates them, and the contract is invisible. So the first skill is not prompting. It is knowing where you are.
The same chat window may train future models on your conversation, allow human review, and retain it for a period you never chose. This varies by product and changes over time, so the safe posture is to assume yes until you have read the policy.
Cross-conversation memory is useful in personal life and dangerous in clinic. The patient from your 2 p.m. visit should not surface in the reasoning about your 3 p.m. visit.
A persistent instruction lowers risk. It does not eliminate it, and models drift from their instructions over long conversations.
Every major assistant has a field for instructions that apply to all conversations. It is where you trade repetition for consistency, and where a safety layer almost nobody uses lives.
Treat every conversation as a different patient. Never carry clinical context from earlier conversations into this one. Flag uncertainty explicitly. If you are not sure, say so. Do not invent doses, reference values, citations, or data I did not give you. Ask me for essential missing information before you answer. If a task is beyond what you can do reliably, say that instead of attempting it.
Instructions help. They do not guarantee. This block lowers risk; it does not replace the verification each card teaches.
I practice in [setting: private office, primary care, urgent care], in [relevant context]. I have access to: [tests, medications, and resources available to me]. I do not have access to: [relevant constraints]. When you suggest management, prioritize what is workable in my setting, but flag when an option outside my access is clinically important, so I can consider referral.
Without context, the model gives you the best treatment on paper. With it, suggestions arrive already filtered for feasibility. The last line is the escape hatch: context must not become a blindfold.
How you want cases laid out, what the tool should never suggest, the formats you use. Document templates live on the card for that task; here you keep the principle, plus one instruction worth stealing:
At the end of any plan, ask: "if this patient came back worse tomorrow, what would you wish you had done differently today?"
An instruction that forces your reasoning instead of replacing it.
Persistent instructions describe you and your practice, never a patient. That is exactly why this is the right place for them: nothing sensitive persists.
A clinical case never goes here, under any circumstances. What may be pasted into a conversation, and how to prepare it, is Card 02.
If the first box is checked, follow your organization's policy. That is what the agreement exists for.
If it is not, the rule is fixed: identifiable patient data does not go into a consumer tool, under any configuration. What goes in is the clinical minimum for the task, and nothing else.
The remaining boxes reduce the risk of what is left. They do not buy an exception to the rule above.