::industry track · healthcare ai

Healthcare AI. Cut the paperwork. Never cut the judgment.

Healthcare is the field where AI hype is loudest and the danger is highest. The same tool that can save a nurse two hours of charting per shift can also get a license pulled, trigger a HIPAA breach letter, or — worst case — contribute to a patient death. So we start here with honesty: AI in healthcare today is a documentation engine, a translator, a research librarian, and a draft-writer. It is not a diagnostician, not a prescriber, not a co-signer on your clinical judgment, and not a defense in a malpractice deposition. The phrase "the AI said so" has zero standing with your malpractice carrier, your state board, or a plaintiff's attorney. Where AI earns its keep right now: drafting SBAR handoffs you then verify, translating discharge instructions into plain language a patient can actually use, synthesizing the last six months of literature on a niche diagnosis before your shift, auditing your own documentation patterns for missed billing codes, and rehearsing difficult family conversations. Where it does not belong: anywhere near a treatment decision, a med dose, an imaging read, a triage call, or a chart that contains identifiable patient data on a public cloud model. The rule that organizes this entire track is simple: PHI never leaves the room. Patient identifiers — name, MRN, DOB, room number, anything from the HIPAA 18 — do not go into ChatGPT, Claude.ai consumer, Gemini consumer, or any model your hospital has not signed a Business Associate Agreement with. Local Ollama on your laptop or a BAA-covered enterprise LLM is the entire universe of legal options. Everything else is a reportable breach.

::the safety rules · non-negotiable

What you never do.

• 01Never paste PHI into a consumer AI tool. The HIPAA 18 identifiers — name, address, dates more specific than year, phone, email, SSN, MRN, account numbers, license numbers, vehicle IDs, device IDs, URLs, IPs, biometric IDs, photos, any other unique identifier — must be stripped or replaced with placeholders before any prompt leaves the building.
• 02Local Ollama on your own laptop OR a hospital-procured LLM with a signed Business Associate Agreement are the only legal places PHI can go. Everything else is a reportable breach under 45 CFR 164.
• 03AI does not diagnose. AI does not prescribe. AI does not decide. If an AI suggests a clinical action, you re-derive that action from your own training, your protocols, and the chart — or you don't take it.
• 04Document what the AI did, not the AI doing it. Your note says 'patient educated on diabetic foot care,' not 'AI generated education sheet given to patient.' You own the content the moment you hand it to a patient.
• 05Never use AI to make a triage decision, a med reconciliation decision, a discharge-vs-admit decision, or an end-of-life decision. These are non-delegable clinical judgments.
• 06AI hallucinates drug doses, drug interactions, and dosing intervals. Always verify against Lexicomp, Micromedex, UpToDate, or your formulary. A wrong dose is not a typo — it's a med error with your license attached.
• 07AI-generated patient education must be reviewed by a licensed clinician before it leaves the building. The reading level should be 6th grade or below. Plain language is a regulated skill, not a stylistic preference.
• 08If your state has a scope-of-practice statute that says only an MD/DO can do X (e.g., diagnose), no amount of AI output makes it legal for non-physicians to do X. Scope is statutory; AI does not amend it.
• 09Photos of patients, photos of monitors, photos of the inside of a room, photos of paperwork — none of these go into cloud AI. The HIPAA 18 includes 'full-face photos and comparable images.'
• 10If a vendor cannot show you the signed BAA and the data-handling addendum, the tool does not touch a patient encounter. 'It says HIPAA-compliant on the website' is not a BAA.

::the workflows · 8 named plays

The plays. Each with the exact prompt.

You are drafting an SBAR handoff for a registered nurse. Use this raw shift information and produce a tight SBAR. Do NOT invent any clinical detail not provided. Flag anything missing in a 'GAPS' section at the end.

SITUATION (one line): [age range, sex, primary admit reason, hospital day #]
BACKGROUND: [relevant PMH, allergies, code status, baseline mental status, isolation precautions]
ASSESSMENT: [current vitals range, mental status, pain level, lines/drains/airway, last labs of note, last imaging of note, current concerns]
RECOMMENDATION: [what needs done this shift, pending consults, pending results, family situation]

Draft the SBAR. Use clinical shorthand a nurse would use. No fluff. End with 'GAPS:' listing what's missing from the above that would normally be in a handoff (e.g., 'no code status documented', 'no allergy info').

Rewrite the following discharge instructions for a patient at a 6th-grade reading level. Constraints:
- Use short sentences, one idea per sentence.
- Replace medical terms with plain words but include the medical term in parentheses the first time (e.g., 'blood thinner (anticoagulant)').
- Group into sections: 'What you have', 'Medicines to take', 'Things to watch for', 'When to call us', 'When to go to the ER', 'Your follow-up'.
- 'When to go to the ER' must use the exact red-flag symptoms I provide and nothing else.
- Do NOT add medical advice I didn't give you. Do NOT invent symptoms, doses, or follow-up timing.
- At the end, list any medical claim you weren't sure about under 'CLINICIAN REVIEW NEEDED'.

ORIGINAL INSTRUCTIONS: [paste clinical instructions here — no patient identifiers]
RED FLAGS THAT MEAN GO TO ER: [list the specific symptoms]
FOLLOW-UP: [appointment type and timing only — no scheduling specifics]

I'm a [role: RN, NP, MD, etc.] preparing for a family meeting. Help me rehearse. Do NOT script me — give me a framework, anticipated questions, and language options.

SCENARIO (sanitized, no identifiers): [e.g., '78-year-old admitted 6 days ago with sepsis from pneumonia, now intubated and on three pressors, family hasn't accepted prognosis']
MEETING GOAL: [e.g., 'introduce hospice as next step', 'clarify code status', 'deliver new cancer diagnosis']
FAMILY DYNAMIC (what I know): [e.g., 'spouse is primary, two adult kids disagree, one wants 'everything done'']
MY ROLE: [bedside RN, primary nurse, attending, charge, palliative consult, etc.]

Give me:
1) A 4-step framework for opening the meeting.
2) Three things I should NOT say (common landmines for this scenario).
3) Five anticipated questions and language options for responding.
4) Two scripts I can use to redirect if it goes sideways.
5) A closing move that ends the meeting cleanly without forcing a decision the family isn't ready to make.

Use VitalTalk and Ariadne Labs Serious Illness Conversation patterns if you know them. Cite the framework you're drawing from.

Synthesize the recent peer-reviewed literature on: [specific clinical question, e.g., 'optimal duration of empiric antibiotics for ventilator-associated pneumonia in adults without immunocompromise'].

Constraints:
- Only use peer-reviewed sources from the last 5 years unless landmark older work is essential.
- Cite every claim with author, year, journal, and DOI or PMID.
- Structure: (1) Current standard of care per major guideline (IDSA, ATS, NICE — name which), (2) Areas of recent change or controversy, (3) Strength of evidence (RCT, meta-analysis, observational, expert opinion), (4) Open questions.
- If a citation doesn't exist or you can't verify it, say 'UNVERIFIED' next to it. Do NOT fabricate DOIs.
- End with 'PRACTICAL TAKEAWAY' — three bullets I can use at the bedside today.

Do NOT include patient-specific advice. This is a literature scan, not a treatment plan.

You are auditing my clinical documentation for completeness and billing optimization. I'm going to paste 5-10 of my recent notes (PHI scrubbed and replaced with placeholders).

For each note, identify:
1) Missing required elements for the documented level of service (e.g., for a 99214, did I document 2/3 of: detailed history, detailed exam, moderate MDM?).
2) Vague language that an auditor or payer would flag (e.g., 'doing well', 'unchanged', 'as previously').
3) Documented work I didn't get credit for (procedures, time-based services, complexity not reflected in the code).
4) Risk patterns (incomplete med reconciliation, missing allergy verification, absent code status, etc.).

Do NOT add clinical detail I didn't document. Only flag what's missing. End with a 'PATTERN' section showing what shows up across multiple notes — that's the habit to fix.

NOTES (all PHI replaced with [PATIENT], [DOB], etc.): [paste]

You are helping me prep for on-call coverage. I'm going to paste the outgoing handoff sheet (PHI scrubbed). For each patient, produce:
1) A one-sentence 'why they're here' summary.
2) The three things most likely to need a call overnight (based on the documented issues).
3) The one thing I should verify before the outgoing team leaves (a gap, an ambiguity, a pending result).
4) A risk tier (LOW / MED / HIGH) for overnight instability.

Group the HIGH-tier patients at the top. Be terse. This is for memorization, not narrative.

Do NOT invent clinical detail. If the handoff sheet is incomplete, flag it. End with a 'CALL OUT' section listing the 3-5 patients I should physically lay eyes on within the first hour of the shift.

HANDOFF SHEET (PHI scrubbed): [paste]

Draft a response to a patient portal message. Constraints:
- Stay strictly within the scope of [your role: RN, MA, NP, MD].
- If the question requires a clinical decision outside my scope, the response should be 'I need to forward this to [provider role] — they'll respond within [timeframe].'
- Use 6th-grade reading level.
- Acknowledge the patient's concern in the first sentence.
- Give one clear action step.
- Include red-flag symptoms that mean call back or go to ER (only if clinically relevant).
- Sign-off appropriate to my role.
- Length: under 150 words.

PATIENT MESSAGE (identifiers removed): [paste]
KNOWN CONTEXT (from chart, identifiers removed): [paste relevant clinical context — diagnoses, current meds class only, recent visits, allergies]
MY ROLE: [RN, MA, NP, etc.]
MY SCOPE LIMITS HERE: [e.g., 'cannot adjust meds', 'cannot order labs', 'can advise on home care for existing conditions']

Draft an insurance appeal letter for a denied [service: prior auth, claim, level of care, etc.]. Format per standard payer expectations.

DENIAL REASON (from EOB or denial letter, identifiers removed): [paste]
SERVICE BEING APPEALED: [e.g., 'inpatient admission days 4-7', 'MRI lumbar spine', 'GLP-1 agonist for T2DM']
CLINICAL JUSTIFICATION (deidentified): [the actual clinical story]
GUIDELINE OR EVIDENCE SUPPORTING: [name the guideline — e.g., MCG, InterQual, IDSA, ACC/AHA — and any relevant studies]
PATIENT-SPECIFIC FACTORS: [comorbidities, failed prior treatments, contraindications — all deidentified]

Produce:
1) A formal appeal letter with: (a) policy citation if you have one, (b) clinical narrative, (c) evidence/guideline citation, (d) requested resolution.
2) Flag any claim where the citation needs verification.
3) Suggest the two strongest arguments to lead with based on the denial reason.

Tone: firm, factual, no hostility. Length: under one page.

::case study · composite anonymized

One human. One year. The pattern.