Day 01 SOAP Notes & HIPAA

Day 1: AI-Assisted SOAP-Note Drafting (HIPAA-Safe)

Cut documentation time by two-thirds without ever pasting protected health information into a public model. The de-identify-first workflow, the structured-template prompt, and the line you must never cross.

~45 min Hands-on By Bo Peng

Today's Goal

Use AI to draft SOAP notes about three times faster while staying HIPAA-compliant. By the end, you will know exactly which patient details to strip before any prompt and you will have a reusable template that turns shift fragments into a clean S-O-A-P structure.

What you'll learn

Why nurses are drowning in documentation

The American Nurses Association has found, in repeated workforce surveys, that documentation now consumes roughly a third of every nursing shift. That is not because nurses write slowly. It is because every patient encounter generates a long tail of structured fields, mandatory free-text narratives, and risk-flagged assessments that must each be entered in the right place, in the right format, with the right vocabulary. The clinical thinking happens at the bedside. The translation into a chartable note happens later, often after the shift, often unpaid.

Generative AI is genuinely good at the translation step. Given a few clinical fragments — "82F admitted yesterday for CHF exacerbation, lung sounds clearer this AM, weight down 1.8 kg, foley draining clear yellow" — a model like Claude or ChatGPT can produce a fully structured S-O-A-P note in three to five seconds. The drafting time savings, when the workflow is set up correctly, is real. Studies of AI-assisted clinical documentation in 2024 and 2025 have shown 60-70 percent reductions in note-writing time without measurable accuracy loss when a clinician reviews the output.

The catch — and it is a serious one — is that consumer AI tools are not Business Associates under HIPAA. You cannot send protected health information to them. The good news is that you do not need to. The drafting workflow that follows never requires any identifier to leave your hospital's secure environment.

The HIPAA line: 45 CFR 164.514 Safe Harbor

The HIPAA Privacy Rule defines de-identified information at 45 CFR 164.514(b)(2). The "Safe Harbor" method requires removal of all 18 of the following identifiers for the patient and any relatives, employers, or household members:

  1. Names
  2. Geographic subdivisions smaller than a state (street, city, county, ZIP except first three digits in some cases)
  3. All elements of dates (except year) directly related to an individual — birth date, admission date, discharge date, date of death
  4. Telephone numbers
  5. Fax numbers
  6. Email addresses
  7. Social Security numbers
  8. Medical record numbers
  9. Health plan beneficiary numbers
  10. Account numbers
  11. Certificate or license numbers
  12. Vehicle identifiers and serial numbers, including license plate numbers
  13. Device identifiers and serial numbers
  14. Web URLs
  15. IP addresses
  16. Biometric identifiers, including finger and voice prints
  17. Full-face photographs and any comparable images
  18. Any other unique identifying number, characteristic, or code

Plus the residual rule: you must have no actual knowledge that the remaining information could identify the individual.

In practice, for a nurse drafting a SOAP note, this collapses to a short pre-flight checklist: strip the name, the room number, the MRN, any specific date (use "today" / "yesterday" / "POD-2"), any ZIP code, any device serial, and any rare condition that could identify a small-town patient. What remains — vitals, lung sounds, medications, response to therapy — is clinical content, not identifying content, and is the only thing the AI actually needs.

The de-identify-first workflow

Here is the sequence. Do not skip steps.

  1. Capture clinical content only in a scratch pad (paper, your facility-approved app, or a notes app on a hospital device that does not sync to consumer cloud).
  2. Replace identifiers with placeholders: Patient becomes "the patient." Mr. Garcia becomes "the patient." 03/14/2026 becomes "today." Rm 412B becomes nothing — delete it. Allergic to "her cat Buster" becomes "allergic to cats."
  3. Run the de-ID prompt (below) to scrub anything you missed.
  4. Run the SOAP-structure prompt on the scrubbed content.
  5. Review every line of the output. The AI is your draft assistant; you are the licensed clinician. Fix anything wrong.
  6. Paste into the EHR. Your hospital's EHR is the only place that should ever see the chart with identifiers attached.

The de-identification prompt

de-identify.txt
PROMPT
You are a HIPAA de-identification assistant. The text below is from
a nurse's scratch pad. Scan for the 18 HIPAA Safe Harbor identifiers
(45 CFR 164.514(b)(2)) and remove or generalize any that remain.

Replace names with "the patient."
Replace specific dates with relative terms (today, POD-1, etc.).
Replace specific places (room numbers, addresses, ZIPs) with deletion.
Flag anything you are uncertain about with [REVIEW].

Return ONLY the cleaned text. Do not add commentary.

---
[paste your scratch-pad content here]

The SOAP-structure prompt

soap-structure.txt
PROMPT
You are an experienced RN documenting an end-of-shift assessment.
Take the clinical fragments below and produce a SOAP note in this
exact structure:

S (Subjective): patient-reported symptoms, in patient's own language
   where quoted. If patient is non-verbal, write "Non-verbal."
O (Objective): vitals, physical exam findings, lab/diagnostic data,
   medications administered, intake/output. Use measurable values.
A (Assessment): clinical impression, change from baseline, response
   to interventions. Tie each finding to a working problem.
P (Plan): nursing interventions for the next shift, monitoring
   parameters, parameters that would trigger a provider call,
   patient/family education provided.

Hard rules:
- Do not invent any data. If a field is missing, write "Not assessed."
- Do not draw conclusions beyond the data. Use "consistent with"
  rather than diagnostic language.
- Flag any finding that warrants provider notification with
  [PROVIDER NOTIFICATION].
- Output should be 150-250 words.

Fragments:
[paste de-identified scratch-pad content here]

What nurses must NEVER paste into a public AI

Hard prohibitions. Even in a "private" chat, even on a paid plan, public consumer AI tools are not HIPAA Business Associates unless your facility has a signed Business Associate Agreement (BAA) with the vendor. Until you have written confirmation from your compliance officer that a BAA is in place, do not paste:

Penalties under HIPAA range from $137 per violation (unknowing) to $2,067,813 per category per year (willful neglect, uncorrected) under the 2024 inflation-adjusted civil monetary penalties. State boards can also discipline a license independently.

HIPAA / nursing-board-compliance pitfalls

Homework before Day 2

  1. Find your facility's policy on generative AI use (intranet search "artificial intelligence" or "ChatGPT"). If there is none, email your supervisor and ask — this is itself a useful nudge.
  2. Take three SOAP notes you wrote this week. De-identify them with the prompt above. Time how long it takes.
  3. Run the SOAP-structure prompt on the de-identified content. Compare to what you originally wrote.
  4. Note one thing the AI got wrong and one thing it got right. You will need this list for Day 5.

Day 1 Checkpoint

Disclaimer. This lesson is educational and does not constitute legal or compliance advice. Always check with your facility's compliance officer, privacy officer, and state board of nursing before adopting any AI tool in patient documentation. Regulations change; verify against the current Code of Federal Regulations and your state's nurse-practice act.

← First lesson Course index Day 2: Patient Education →