Turn a Clinical Policy Document into a Plain-Language Patient FAQ in One Prompt Session

The task

Patient experience leads, compliance officers, and clinic communications staff have to translate dense clinical or regulatory policy documents into patient-facing FAQs whenever a hospital adopts (or updates) an AI tool, changes a consent workflow, or responds to new guidance. With the UK commission's fresh recommendations on regulating AI in medicine landing this week, expect a wave of internal policy memos that need patient-ready versions fast.

Before AI

A comms lead reads the policy, flags anything a patient would ask about, drafts 8-15 Q&As at roughly a 7th-grade reading level, then routes to Clinical for accuracy, Legal for scope, and Accessibility for readability. Two to three hours of writing plus a day or two of back-and-forth — and that's before translation.

The workflow

1. Extract patient-relevant claims from the policy

Paste the policy text. The first pass is not writing — it's identifying which parts of the document actually affect a patient's experience, choices, or rights. Everything else (governance, procurement, internal audit cadence) gets set aside.

You are a healthcare patient-communications analyst. I will paste a clinical or regulatory policy document below.

Your job in this step ONLY: extract the items that materially affect a patient. Do NOT write an FAQ yet.

For each patient-relevant item, output a row with:
- Claim (one sentence, quoting or closely paraphrasing the policy)
- Why a patient would care (one sentence)
- Section reference (heading or paragraph number if present)
- Sensitivity flag: [Consent] [Data/Privacy] [Clinical decision] [Access/Eligibility] [Cost] [Rights/Recourse] [None]

Ignore purely internal governance, procurement, staffing, or IT-ops language unless it changes what a patient sees, signs, pays, or can appeal.

Return a clean markdown table. If the document contains no patient-relevant items, say so and stop.

POLICY DOCUMENT:

MEMO — Meadowbrook Regional Health System
Policy #CP-2026-047: Deployment of AI-Assisted Radiology Triage (AIRT) Tool "LumenRead v3.2"
Effective: 01 October 2026
Owner: Dr. A. Okafor, Chief Medical Informatics Officer

1. Purpose
This policy governs the deployment of LumenRead v3.2, a CE-marked AI tool that pre-reads chest X-rays and flags suspected pneumothorax, large pleural effusion, and suspected malignancy for radiologist prioritization.

2. Scope
Applies to all adult outpatient and ED chest radiographs performed at Meadowbrook Main and the Northside Ambulatory Center. Pediatric imaging is excluded. Inpatient imaging is out of scope for Phase 1.

3. Clinical Workflow
3.1 Every in-scope chest X-ray is analyzed by LumenRead within ~90 seconds of acquisition.
3.2 A board-certified radiologist reads and signs every study. The AI output is advisory only and does not replace radiologist interpretation.
3.3 Studies flagged "high suspicion" are moved to the top of the radiologist worklist. Median time-to-read for flagged studies fell from 42 min to 11 min in the pilot (n=3,180).

4. Patient Notification & Consent
4.1 Patients will be informed via a one-page handout at check-in that an AI tool assists in reading their X-ray. A poster will be displayed in waiting areas.
4.2 No separate written consent is required, consistent with Legal opinion 2026-L-118. Patients may request that AI assistance NOT be used; such requests will be honored and documented in the EHR (SmartPhrase .noAIRT).
4.3 Interpreter services are available; the handout is translated into Spanish, Vietnamese, Somali, and Arabic.

5. Data & Privacy
5.1 Images and DICOM metadata are processed on-premises. No PHI leaves the Meadowbrook network.
5.2 Vendor receives only de-identified aggregate performance metrics on a quarterly basis, per BAA signed 14 Aug 2026.
5.3 Patients may request an accounting of disclosures under HIPAA §164.528; AI-assisted reads are logged as an internal use, not a disclosure.

6. Clinical Governance
6.1 A monthly QA committee reviews discordant cases (AI flag vs. radiologist read).
6.2 Model performance is monitored for drift; a >2% drop in sensitivity or specificity triggers suspension pending revalidation.
6.3 Adverse events attributable to AI output are reported to Risk Management within 24 hours and to the MHRA per post-market surveillance obligations.

7. Patient Rights
7.1 Patients who believe an AI-influenced read contributed to a delayed or missed diagnosis may file a complaint via Patient Relations (ext. 4-2200) and are entitled to a second independent read at no charge.
7.2 Patients may request their imaging record and the associated AI output metadata via Health Information Management.

8. Training & Staffing
8.1 All radiologists complete a 2-hour LumenRead onboarding module before go-live.
8.2 Techs receive a 30-minute workflow briefing.

9. Review
This policy is reviewed annually or upon major model version change (defined as any change to the training data, thresholds, or intended use).

2. Draft the FAQ from the extracted claims

Now convert the shortlist into patient-facing Q&As. Constrain the reading level explicitly — LLMs will drift to 10th-grade prose if you don't.

Using ONLY the patient-relevant claims table you produced in the previous step, draft a patient FAQ.

Requirements:
- 8 to 12 Q&A pairs. Questions phrased the way a real patient would ask them ("Will a computer be reading my X-ray?" not "What is the AI-assisted triage protocol?").
- Plain English, target reading level Grade 6-7. Short sentences. No acronyms without a one-line definition on first use.
- Every answer must be grounded in a specific claim from the table. Do NOT invent capabilities, statistics, or rights not present in the source.
- If a patient would reasonably ask something the policy does NOT answer, include it under a final section "Questions the policy doesn't answer yet" with a one-line note on who to contact instead. Do not fabricate an answer.
- For any Q&A touching [Consent], [Data/Privacy], or [Rights/Recourse], end the answer with a short line: "Ask us if you want this in writing."
- Do NOT include marketing language ("state-of-the-art", "cutting-edge", "world-class"). Neutral tone only.

Output format: markdown, each Q as `### Q: ...` followed by the answer as a normal paragraph.

3. Run a clinical-accuracy and safety pass

This is the step people skip and regret. Have the model audit its own draft against the source, flagging drift, overclaims, and anything that could read as medical advice.

Audit the FAQ you just drafted against the original policy document.

For each Q&A, produce a review row:
- Q number
- Verdict: OK | Drift | Overclaim | Missing citation | Reads as medical advice | Consent/privacy issue
- Evidence: quote the specific policy sentence(s) that support the answer, or write "NOT IN SOURCE"
- Suggested fix (if not OK): one line

Then, at the bottom, output a REVISED FAQ that incorporates every suggested fix. Keep the same Q&A structure and reading level from the previous step.

Flag especially:
- Any answer that tells the patient what a clinical outcome will be for them personally (that's medical advice, not policy communication).
- Any statistic or number in the FAQ that is not present verbatim or as a direct paraphrase in the source.
- Any implied guarantee ("your scan will be read faster") vs. the policy's actual language (median improvement in a pilot).

4. Package for review routing

Final step gets it out of the model and into the hands of the humans who sign off.

Produce a single handoff package containing:

1. **Final patient FAQ** (the revised version from the previous step, clean, no audit annotations).
2. **Reviewer cover note** — 5 bullet points max, addressed to Clinical, Legal, and Patient Experience, calling out:
   - Which Q&As touch consent or privacy and need Legal eyes
   - Which Q&As paraphrase clinical performance claims and need Clinical eyes
   - Which questions the source policy does not answer (from the "doesn't answer yet" section) and who owns getting an answer
3. **Translation-ready flag list** — list every Q number that uses an idiom, a number, or a named product/tool, since those are the items translators most often need clarification on.
4. **Change log stub** — a one-line entry in the format: `YYYY-MM-DD | Policy #___ v___ | FAQ generated from source | Pending Clinical/Legal review`.

Output as one markdown document with those four sections in that order.

Gotchas

• The model will confidently invent patient rights that aren't in the policy. Step 3 catches most of this, but read every "you have the right to…" line against the source before it goes out.
• Statistics drift. A policy that says "median time-to-read fell from 42 to 11 minutes in a 3,180-patient pilot" will get compressed to "reads your scan 4x faster" unless you explicitly forbid it. The audit step flags this — don't disable it.
• Consent language is jurisdiction-specific. A US HIPAA-flavored answer won't fit a UK NHS trust operating under MHRA and the emerging AI regulatory framework the UK commission just recommended. If your policy cites a specific regulator, tell the model in Step 2 to preserve that citation verbatim.
• Reading level is a promise, not a measurement. The model targets Grade 6-7 but doesn't verify. Run the final FAQ through a real readability checker (Flesch-Kincaid in Word works) before publishing.
• PHI safety. The workflow assumes you're pasting a policy document, not a patient chart. Do not paste EHR excerpts, case notes, or anything with identifiers into a general-purpose model. Use your institution's approved, BAA-covered environment.
• Not a substitute for legal review. Compliance sign-off is still required. This workflow gets you to a reviewable draft faster; it does not replace the reviewers.

Time saved

Roughly 60-90 minutes per policy update on the drafting side — the first-pass FAQ, the audit table, and the routing packet are the parts that used to eat an afternoon. Downstream review time (Clinical, Legal, Accessibility) is unchanged; you still need those humans. The gain is that reviewers receive a cleaner artifact with citations already tied back to the source, which tends to shorten the review loop by one round.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

| # | Claim | Why a Patient Would Care | Section Reference | Sensitivity Flag | |---|-------|--------------------------|-------------------|------------------| | 1 | Every in-scope adult outpatient and ED chest X-ray at Meadowbrook Main and Northside Ambulatory Center is analyzed by the LumenRead AI tool; pediatric and inpatient imaging are excluded. | A patient needs to know whether this AI process applies to their specific visit and setting. | §2 Scope; §3.1 | [Clinical decision] | | 2 | Patients are informed via a one-page check-in handout and waiting-area poster that an AI tool assists in reading their X-ray; no separate written consent is required. | A patient may not realise they are consenting passively simply by proceeding, with no signature required to acknowledge AI involvement. | §4.1; §4.2 | [Consent] | | 3 | Patients may request that AI assistance NOT be used on their study, and that opt-out will be honored and documented in their medical record. | A patient who objects to AI involvement has a concrete, exercisable right to exclude it — but must proactively ask. | §4.2 | [Consent] [Rights/Recourse] | | 4 | The patient-facing handout is available in Spanish, Vietnamese, Somali, and Arabic, and interpreter services are available. | Non-English-speaking patients can receive AI-use information in their language, affecting informed decision-making. | §4.3 | [Consent] [Access/Eligibility] | | 5 | All images and personal health information (DICOM metadata) are processed on-premises and no PHI leaves the Meadowbrook network. | Patients can be assured their imaging

…output truncated at 250 words.

Step 2 output

# Meadowbrook Regional Health System ## Your Questions About AI-Assisted Chest X-Ray Reading --- ### Q: Will a computer be reading my X-ray? A computer tool called LumenRead will look at your chest X-ray first — usually within about 90 seconds of the image being taken. But a real, qualified radiologist (a doctor who specialises in reading medical images) reads and signs off on every single X-ray. The computer's job is to assist, not to replace, that doctor. The final call always comes from a human. --- ### Q: Do I have to agree to the computer being used on my X-ray? You will be given a one-page handout at check-in that explains how the computer tool works. You do not need to sign a separate consent form. However, if you do not want the computer tool used on your X-ray, you can say so and your request will be honoured and noted in your medical record. Ask us if you want this in writing. --- ### Q: What if I don't want the computer involved in reading my X-ray? You have the right to opt out. Just let a staff member know before your X-ray is taken. Your request will be recorded in your medical record. You do not need to give a reason. Ask us if you want this in writing. --- ### Q: Will the computer help make sure serious problems are spotted faster? Yes. When the computer flags an X-ray as possibly serious — for example,

…output truncated at 250 words.

Step 3 output

Part 1 — Audit Table | Q# | Verdict | Evidence | Suggested Fix | |----|---------|----------|---------------| | 1 | **Drift** | §3.1: "Every in-scope chest X-ray is analyzed by LumenRead within ~90 seconds of acquisition." §3.2: "A board-certified radiologist reads and signs every study. The AI output is advisory only and does not replace radiologist interpretation." | Remove "usually within about 90 seconds" — the policy says "~90 seconds" as a system performance figure, not a patient-facing promise. Rephrase as "shortly after" to avoid implying a guarantee. | | 2 | **OK** | §4.1: "Patients will be informed via a one-page handout at check-in…" §4.2: "No separate written consent is required… Patients may request that AI assistance NOT be used; such requests will be honored and documented in the EHR." | — | | 3 | **OK** | §4.2: "Patients may request that AI assistance NOT be used; such requests will be honored and documented in the EHR." | — | | 4 | **Overclaim / Reads as medical advice** | §3.3: "Studies flagged 'high suspicion' are moved to the top of the radiologist worklist. Median time-to-read for flagged studies fell from 42 min to 11 min in the pilot (n=3,180)." | The phrase "potentially serious findings may be reviewed more quickly" edges toward a personal clinical outcome promise. Anchor firmly to the pilot statistic and remove implied personal benefit. Also clarify the 42→11 min figure is a pilot median, not a current guarantee. | | 5 | **OK**

…output truncated at 250 words.

Step 4 output

# Patient FAQ Handoff Package ### Meadowbrook Regional Health System | Policy #CP-2026-047 | LumenRead v3.2 ### Prepared for: Clinical, Legal, and Patient Experience Review --- ## Section 1 — Final Patient FAQ ### Your Questions About AI-Assisted Chest X-Ray Reading --- ### Q1: Will a computer be reading my X-ray? A computer tool called LumenRead will look at your chest X-ray shortly after the image is taken. But a real, qualified radiologist — a doctor who specialises in reading medical images — reads and signs off on every single X-ray. The computer's job is to assist, not to replace, that doctor. The final call always comes from a human. --- ### Q2: Do I have to agree to the computer being used on my X-ray? You will be given a one-page handout at check-in that explains how the computer tool works. You do not need to sign a separate consent form. However, if you do not want the computer tool used on your X-ray, you can say so and your request will be honoured and noted in your medical record. Ask us if you want this in writing. --- ### Q3: What if I don't want the computer involved in reading my X-ray? You have the right to opt out. Just let a staff member know before your X-ray is taken. Your request will be recorded in your medical record. You do not need to give a reason. Ask us if you want this in writing. --- ### Q4:

…output truncated at 250 words.

---

This content is for informational purposes only and is not medical advice. AI tools used with patient data must meet your organization's HIPAA and privacy requirements.