Best Claude prompts for therapists in 2026

Twelve Claude prompts working therapists actually use — DAP-note structuring from de-identified notes, treatment-plan goal/objective drafting aligned to CPT codes, psychoeducation handouts in patient language, case conceptualization scaffolds for CBT/IFS/EMDR, supervision-meeting prep, telehealth intake scripts, insurance-appeal narratives for medical necessity, sliding-scale conversation scripts, niche marketing bios, weekly caseload reflection, between-session worksheet drafts, and ethical decision-making for novel situations. Each prompt assumes de-identified input. Claude is a drafting partner; the clinical decision and signature remain yours under the APA Ethics Code and NASW Code of Ethics.

→ **Generate a custom clinical prompt with our builder** (free, no signup)

---

Marketing prompts reward variety and persuasion. Clinical prompts reward the opposite — explicit framework, conservative language, refusal to diagnose, and a clear paper trail. The APA's 2024 guidance on the use of generative AI in psychological practice emphasizes that clinicians retain full professional responsibility for any AI-assisted product, including notes, treatment plans, and client communications.

Three ground rules sit underneath every prompt below:

- **De-identify before pasting.** Strip the 18 HIPAA identifiers under the Safe Harbor method described in 45 CFR 164.514(b)(2). Substance-use disorder records carry the additional protections of 42 CFR Part 2 and should never be routed through a non-BAA-covered LLM session, even de-identified, without a clinical-leadership policy. - **Use Claude for structure, not clinical judgment.** Ask it to scaffold a DAP note from your bullet points; do not ask it to diagnose, prescribe, or determine risk level. The NASW Code of Ethics §1.04 (Competence) makes clinical judgment a non-delegable obligation. - **Sign your work, edit every line.** Recent peer-reviewed work in JAMA Psychiatry on AI in mental health workflows consistently finds clinician-edited AI drafts outperform either AI-only or clinician-only baselines on accuracy and tone — but only when the clinician actually edits, not just signs.

The prompts below assume Claude Sonnet 4 or newer with a covered Team/Enterprise plan, or a HIPAA-compliant deployment via a partner that signs a BAA. Consumer Claude.ai does not carry a BAA by default.

---

**When to use:** You finished a session, jotted bullet observations, and want a clean DAP (Data / Assessment / Plan) draft you will then edit and sign.

``` You are helping a licensed clinician structure a session note in DAP format. Do not invent clinical content. Use only the bullet points provided. Do not assign diagnoses, risk levels, or symptom severity ratings. Do not write the signature line — the clinician signs. Format: D (Data): Observable session content — client presentation, themes discussed, interventions used. Behavior described in concrete terms, no interpretation. A (Assessment): Clinical impression in cautious language ("client appears to be...", "consistent with prior reported pattern of..."). No new diagnosis. Note progress toward existing treatment goals. P (Plan): Next session focus, between-session task if any, referral/coordination needs, risk monitoring if applicable. Constraints: - Use only the bullets below. - Flag any bullet that lacks enough specificity to write without inference. - Keep under 250 words. - Plain professional register, no jargon padding. De-identified bullets: [paste — names replaced with CLIENT, ages bucketed, dates relative ("session 7"), no MRN] ```

**Sample output (excerpt):** _"D: CLIENT arrived on time, oriented x3. Reported a difficult week following job-search rejection; described sleep disruption (3–4 hrs/night, 4 nights). Therapist used Socratic questioning to examine catastrophic prediction about employability. A: CLIENT's presentation is consistent with prior reported pattern of rumination under perceived failure. Progress on Goal 2 (cognitive restructuring) appears modest; Goal 1 (sleep hygiene) regressed this week. P: Reintroduce sleep-hygiene checklist; assign thought-record on rejection week…"_

DAP note structure is well-documented across clinical training literature; SAMHSA's TIP 35 on documentation in behavioral health offers a useful structural reference if your supervisor wants a citation in your policy memo.

---

**When to use:** Intake is done, you need a treatment plan with measurable goals and objectives that map cleanly to the CPT code you bill (90832 / 90834 / 90837 / 90847 / 90791).

``` Draft a treatment plan section with one long-term goal and three short-term objectives, structured so the work is consistent with [CPT CODE] (define the time and modality the code represents) and the working diagnosis is [ICD-10 CODE]. For each objective: - Behaviorally measurable (use "will demonstrate", "will identify", "will report", not "will understand"). - Includes a target frequency and a review date. - Tied to a specific intervention modality. - Does NOT include any client identifiers. After the goal/objective block, output: - The clinical rationale linking objectives to the working diagnosis. - Three risk factors that would prompt a plan revision. - A note on what coordination-of-care documentation would support medical-necessity language for payer review. Client context (de-identified): [paste presenting concerns, prior treatment, functional impacts — no PHI] ```

**Sample output (excerpt):** _"Long-term goal (90834, 45-min individual psychotherapy): Client will reduce frequency of panic episodes from baseline of 4/week to 1/week over 12 sessions. Objective 1: Client will identify three personal panic-onset cues and report them weekly for four consecutive sessions. Objective 2: Client will demonstrate paced-breathing technique to therapist by session 4…"_

Treatment-plan language that supports medical necessity is shaped by payer policy and by CMS Local Coverage Determinations; the CMS Medicare Benefit Policy Manual, Chapter 15 on covered mental health services is a useful anchor.

---

**When to use:** A client would benefit from a one-page handout on a concept you cover often (window of tolerance, cognitive triangle, polyvagal ladder, parts work basics).

``` Draft a single-page client psychoeducation handout on [CONCEPT]. Reader: an adult client with no clinical background, reading after a hard session. Aim for an 8th-grade reading level. Structure: 1. One-sentence definition in plain language. 2. "What it looks like in real life" — three short examples a client might recognize. 3. "Why it happens" — brief, normalizing, not pathologizing. 4. "What you can try this week" — two small, concrete experiments (not assignments). 5. "When to bring it up in session" — two prompts the client can use. Constraints: - No diagnosis language. - No claims about brain regions unless I provide them. - No "always" or "never." - Warm, second-person voice, but not childish. - Maximum 350 words. ```

**Sample output (excerpt):** _"Your window of tolerance is the zone where you can feel things, think clearly, and make choices that match your values. When stress, sleep loss, or old triggers push you outside that zone, your body shifts into either a 'too much' state (overwhelm, racing, anger) or a 'too little' state (shut down, numb, disconnected). Neither state means you're broken — it means your nervous system is trying to protect you…"_

The Anthropic guidance on adjusting register and reading level is genuinely useful here; specifying the 8th-grade target and the "after a hard session" reader keeps the output from sliding into textbook tone.

---

**When to use:** Intake is complete and you want a written case conceptualization that fits a specific modality (CBT, IFS, EMDR) for your own clinical thinking and your supervision file.

``` Scaffold a case conceptualization in the [MODALITY: CBT / IFS / EMDR] framework. Use only the de-identified information below. Do not invent background details. Where information is missing, list it as an "information gap" rather than guessing. For CBT, structure: - Predisposing, precipitating, perpetuating, protective factors - Core beliefs, intermediate beliefs, automatic thoughts - Behavioral patterns reinforcing the cycle - Treatment targets and the cognitive/behavioral interventions matched to each For IFS, structure: - Initial parts mapped (managers, firefighters, exiles named generically) - Burdens and roles each part may carry - Constraints on Self-energy access - Sequencing hypothesis for unburdening work For EMDR, structure: - Target memory clusters (touchstones, present triggers, future templates) - Negative cognitions, desired positive cognitions - Resourcing needs before reprocessing - Stability and dual-attention readiness factors Close with three falsifiable hypotheses you'd want to test in the first four sessions, and what observation would disconfirm each. De-identified intake summary: [paste] ```

Case conceptualization is the area where Claude is most useful and most dangerous — useful because it organizes your thinking faster than any blank-page template, dangerous because a well-written conceptualization sounds confident even when it is wrong. Make the model surface information gaps explicitly. Persons reading Persons et al.'s work on CBT case formulation will recognize the structure.

---

**When to use:** You have 15 minutes of supervision and three cases you want feedback on; you need a tight, structured ask for each.

``` Write a one-paragraph supervision-meeting summary for the case below. Structure each paragraph: 1. Two-sentence orientation: who the client is functionally (no name, no DOB), modality used, session count. 2. The clinical question I want supervision on — phrased as a question, not a complaint. 3. What I have already tried. 4. Two specific options I am considering, with the case for and against each. 5. The decision I need to make before the next session. Constraints: - Maximum 180 words. - No diagnostic certainty language ("presents with", not "is"). - Flag countertransference if it shows up in my notes. De-identified case notes: [paste] ```

Supervision time is scarce and the APA Ethics Code §2.05 on Delegation of Work and Supervision (and the parallel NASW provisions) treats supervision quality as part of competence. A well-structured ask is a way of honoring both your supervisor's time and your client's case.

---

**When to use:** You are about to onboard a new telehealth client and want a script for the first 10 minutes that handles consent, technology, emergency planning, and rapport.

``` Draft a 10-minute opening script for a telehealth intake session. The client is an adult presenting for individual psychotherapy. Cover: 1. Confirm identity and current physical location (state, address) for licensure and emergency planning. 2. Confirm the client is in a private space and has uninterrupted time. 3. Review limits of confidentiality including mandatory reporting and duty-to-warn applicable in [STATE]. 4. Telehealth-specific informed consent: technology limits, what happens if connection drops, agreed reconnect protocol, alternate phone number. 5. Emergency plan: nearest ER, support person on call, local crisis line number, 988 reminder. 6. Brief orientation to how sessions run and what to expect today. Constraints: - Warm, plain language, second person. - Pause prompts where I should let the client speak. - Flag any element my state may require to be in writing rather than verbal. ```

Telehealth-specific informed consent is required as a standalone element under most state boards and is referenced in the APA Guidelines for the Practice of Telepsychology. State requirements vary; verify against your licensing board's current rule set.

---

**When to use:** A payer has denied continued authorization and you need to draft a clinically grounded appeal letter that argues medical necessity using their own criteria.

``` Draft an insurance appeal letter for a denied continued-authorization request. Use the medical-necessity criteria from the payer's published clinical policy where I provide them; if I do not, default to MCG or InterQual-style framing (functional impairment, treatment response, risk factors, plan-revision logic). Structure: 1. Header: client identifiers in placeholder form ([CLIENT_ID], [DOB_REDACTED], [DOS_RANGE]), policy number placeholder. 2. The clinical reason continued treatment is medically necessary, tied to functional impairment domains (work, relationships, ADLs, safety). 3. Treatment response to date — what changed, what has not. 4. Risk factors that would emerge with premature termination. 5. The revised treatment plan with measurable objectives and a review date. 6. Citation to applicable parity protections under the [Mental Health Parity and Addiction Equity Act] where the denial appears to apply a non-quantitative treatment limitation more stringent than for medical/surgical care. 7. Closing request and required documentation list. Constraints: - Professional, factual, no emotional appeal. - No diagnostic certainty beyond what my notes support. - Maximum 600 words. De-identified clinical context: [paste] ```

Parity-based appeal language has real teeth: the Department of Labor's MHPAEA enforcement materials and CMS parity rules require non-quantitative treatment limitations to be applied no more stringently to mental health than to medical/surgical benefits. Cite parity where the denial pattern justifies it.

---

**When to use:** A prospective client has asked about reduced fees and you want a script that handles fee transparency, your own sustainability, and the relational frame.

``` Write a script for a 5-minute sliding-scale fee conversation with a prospective client who has emailed asking about reduced fees. Structure: 1. Acknowledge the question warmly, without apology. 2. Briefly state your practice's sliding-scale model in concrete terms (e.g., "I hold X reduced-fee slots; current availability is Y"). 3. Offer the specific options available to this client — full fee, sliding fee with a stated number, or referral to a lower-cost option you respect. 4. Name the boundary clearly: rates are reviewed annually; reduced-fee slots have a duration cap. 5. Invite the client to decide without pressure and offer a clear next step either way. Constraints: - No guilt language on either side ("if you can afford", "I really need"). - No diagnostic prejudgment. - 200-250 words. - Match the tone of the warm-but-clear practitioners my client will have encountered if they've shopped therapists before. ```

Fee transparency is both ethical practice (per APA Ethics Code §6.04 on Fees and Financial Arrangements) and clinically therapeutic — ambiguity around money introduces rupture risk. A scripted version helps you stay on your stated policy.

---

**When to use:** You are updating your Psychology Today profile, your group practice page, or your private-practice site, and you want a bio that draws the clients you actually want to work with.

``` Write a 200-word therapist bio for a clinical webpage. Reader is an adult researching therapists for [NICHE — e.g., perinatal anxiety, first-gen professionals, complex PTSD, ADHD adults]. The reader is skeptical, has been burned by a prior bad fit, and is scanning multiple bios in one sitting. Structure: - Opening: what the client is likely struggling with, in their own language, not clinical labels. - Middle: who I help and how the work tends to unfold, named modalities in plain terms ("I draw from… which means…"). - Close: one sentence on what working with me feels like, a clear next step (free consult / how to schedule), and an honest "not a fit if…" line that helps mis-fit clients self-select out. Constraints: - No "safe space," "holistic," "journey," "empower," "unique." - No diagnostic claims I am not licensed to make. - Active voice. - 8th-9th grade reading level. - Compliant with state advertising rules — flag any phrase that could read as guarantee of outcome. My actual credentials, modalities, and target client: [paste] ```

State licensing boards regulate therapist advertising; the broad rules around no outcome guarantees and accurate credential representation are summarized in most state board advertising statutes. Cross-check phrases the model flags before publishing.

---

**When to use:** Friday afternoon, week is closing, you want a structured 15 minutes to look at your caseload before the next week starts.

``` Guide me through a structured weekly caseload reflection. For each de-identified case I paste below, ask me four questions in order. After my answer to each, summarize what you heard in one sentence and ask the next question. Do not skip questions even if I ramble. Questions: 1. What changed for this client this week — even slightly? 2. Where am I stuck in the work? 3. What is one hypothesis I am holding too tightly? 4. What is the smallest experiment I could try next session? After all cases, give me: - The two cases I sounded most clinically stuck on. - The two cases where I sounded most energized. - A one-sentence reflection on what those patterns might mean for my own self-care or training direction this month. De-identified cases: [paste — use generic descriptors only] ```

Reflective practice is associated with better therapist outcomes in longitudinal therapist-development research summarized by the APA Division 29 (Psychotherapy). Claude is genuinely useful as a structured-questions partner here — not as a clinical voice.

---

**When to use:** A session ended on a workable theme and a between-session worksheet would let the client keep practicing the work.

``` Draft a one-page between-session worksheet for a client working on [THEME — e.g., thought-record practice, urge-surfing, parts-mapping, values-action linkage]. Structure: 1. One-sentence reminder of the in-session insight, in the client's register — not clinical. 2. "Notice it this week" — three small prompts for noticing the theme as it shows up. 3. "Try it on" — one structured exercise with 3-5 short prompts and space for the client to write. 4. "Bring back to session" — two questions to bring into next session based on what they noticed. Constraints: - 8th-grade reading level. - No homework framing — use "try it on" or "experiment." - No assumed outcome. - Designed to be useful even if the client only does part of it. - Print-friendly with clear white space. ```

Between-session work is associated with better outcomes across modalities; meta-analytic findings reviewed in the APA Society for the Advancement of Psychotherapy literature back this up. The trick is reducing friction — a worksheet a client does halfway beats a workbook page they avoid.

---

**When to use:** Something has come up that does not fit a clean rule — a dual-relationship question, a confidentiality edge case, a request for records under unusual conditions — and you want a structured think-through before consultation.

``` Help me think through an ethical decision using a structured framework. Do not tell me what to do. Surface considerations and tensions. Framework (six steps): 1. State the dilemma in one sentence as a tension between two values. 2. Identify the stakeholders — client, family, third parties, profession, self — and what each has at stake. 3. Name the relevant standards: applicable APA/NASW/AAMFT code sections, state board rules, HIPAA/42 CFR Part 2 if data involved, mandatory reporting laws. 4. List possible courses of action — at least four — including the uncomfortable ones. 5. For each, name the predictable consequences across stakeholders and the worst-case if the assumption is wrong. 6. Identify what consultation or supervision the situation requires before acting, and what I should document. Constraints: - No final recommendation. Decision stays with me and my consultation partner. - Cite code section numbers only where you are confident; otherwise say "verify the applicable standard." - Flag any element that requires immediate consultation (risk to client, board exposure, legal obligation timing). De-identified situation: [paste] ```

Structured ethical reasoning is a competency under the APA Ethics Code Preamble and Decision-Making Process and the NASW Code Ethical Decision-Making framework. The frameworks are public; Claude can scaffold the thinking, but the consultation step and the decision both remain human.

---

---

Three overlapping legal frameworks shape every paste into Claude: the HIPAA Privacy Rule for protected health information, the heightened protections of 42 CFR Part 2 for substance use disorder records, and your professional code of ethics. None of them pause for AI tooling.

Practical rules I follow and recommend to the practices I consult with:

1. **De-identify under HIPAA Safe Harbor before pasting.** Remove the 18 identifiers listed in 45 CFR 164.514(b)(2): names, geographic subdivisions smaller than a state, dates more specific than year (except year of birth for adults under 89), phone, fax, email, SSN, MRN, plan numbers, account numbers, certificates, vehicle IDs, device IDs, URLs, IPs, biometrics, full-face photos, any other unique identifier. 2. **SUD records require more, not less.** Under 42 CFR Part 2, records from Part 2 programs require specific consent for redisclosure and carry penalties beyond HIPAA. Default position: do not route Part 2 content through any LLM session without organizational policy approval. 3. **Consumer Claude.ai is not HIPAA-covered.** Anthropic offers a Business Associate Agreement for eligible enterprise customers; without a signed BAA, treat consumer Claude as you would any non-covered third-party tool. De-identified data is acceptable; PHI is not. 4. **Set up training opt-out for business plans.** Anthropic's data usage policy for commercial customers confirms enterprise/team inputs are not used for training by default. Verify settings before clinical use. 5. **Document AI use in your records-handling policy.** A short policy memo — "clinicians may use Claude for X, Y, Z with de-identified content; clinical judgment remains the clinician's; signature on any AI-assisted note attests to clinician review" — protects you in a board complaint and meets the supervision/consent expectations under the APA and NASW codes.

A simple rule: if you would not paste it into a public forum, prep it before pasting it into Claude.

→ **See our full prompt library for clinical and helping-profession workflows**

---

Each prompt above went through the iteration protocol described in our 7-point prompt grading rubric guide. The two dimensions that lifted clinical prompts most were constraint-compliance (especially the no-diagnosis and no-invention rules) and explicit refusal pathways (telling Claude what to do when information is missing rather than letting it fill gaps).

Structure draws on Anthropic's published prompt engineering guidance on role assignment, structured outputs, and explicit reasoning steps. The clinical authority sources behind the reasoning — APA Ethics Code, NASW Code of Ethics, HIPAA, 42 CFR Part 2, MHPAEA, CMS Chapter 15, APA Telepsychology Guidelines — were chosen because they are publicly verifiable and because the model will occasionally cite outdated language; naming the source forces it to cite accurately or admit uncertainty.

→ **Try our prompt builder to adapt these for your practice**

---

### Is it HIPAA-compliant to use Claude for clinical documentation?

Only under specific conditions: a signed Business Associate Agreement with Anthropic via an eligible enterprise plan or a HIPAA-aligned partner deployment, plus de-identification of inputs before paste, plus your own organizational policy authorizing the use. Consumer Claude.ai without a BAA is not appropriate for PHI. De-identified content paired with a covered plan is the standard pattern.

### Will Claude replace a therapist?

No. Claude drafts notes, scaffolds case formulations, and helps you stay organized. Clinical judgment, diagnosis, risk assessment, intervention selection, and the signature remain the clinician's under the APA Ethics Code, NASW Code of Ethics, and state licensing rules. Recent JAMA Psychiatry coverage of AI in mental health workflows is consistent on this point.

### Which Claude model works best for clinical work?

Claude Sonnet 4 or newer. Older models miss the structured reasoning these prompts require, especially the case-conceptualization scaffold (Prompt 4) and the ethical-reasoning framework (Prompt 12). For high-stakes outputs — medical-necessity appeals, treatment-plan revisions — run the prompt twice and compare; divergence is signal that the question is genuinely ambiguous and warrants slower thinking.

### Can I use these prompts for client-facing chatbots or psychoeducation tools I host?

Not without significant additional safeguards. The prompts above are designed for clinician-facing drafting workflows where you review and sign. Client-facing deployment introduces duty-of-care, FDA software-as-medical-device questions, and state telehealth regulation that the prompts do not address. The APA guidance on AI in psychological practice and emerging FDA guidance on digital therapeutics are the starting points.

### What if my client asks whether I use AI for their notes?

Tell them. A short paragraph in your intake paperwork — "I sometimes use AI tools to help structure notes and draft documents. I de-identify any clinical content before using these tools, and I review and sign every document myself." — is consistent with the informed-consent obligations under the APA and NASW codes and is becoming a community standard.

### How do I know if Claude's clinical output is wrong?

Read every line. Check diagnosis-adjacent language for certainty it should not carry. Check any cited code section, ICD/CPT code, or regulation. Watch for plausible-sounding but generic case formulation that does not match the actual client. Hallucinated citations and overconfident clinical claims are the two most common failure modes and the easiest to catch when you are looking for them.

### Does this apply to 42 CFR Part 2 (SUD) records?

Part 2 deserves its own policy. Default position: do not route Part 2 content through any LLM, even de-identified, without specific organizational-leadership policy approval. SAMHSA's 42 CFR Part 2 resource page is the authoritative reference.

---

- APA Ethics Code - APA Guidance on AI in Psychological Practice - APA Guidelines for the Practice of Telepsychology - NASW Code of Ethics - HHS — HIPAA De-Identification of Protected Health Information - SAMHSA — 42 CFR Part 2 Confidentiality Regulations - DOL — MHPAEA Parity Enforcement - CMS — Medicare Benefit Policy Manual, Chapter 15 - JAMA Psychiatry — ongoing AI in mental health coverage - Anthropic — Prompt engineering overview - Anthropic — Commercial terms and data handling

→ **Start with our prompt builder — free, no signup**

<script type="application/ld+json" dangerouslySetInnerHTML={{ __html: JSON.stringify({ "@context": "https://schema.org", "@type": "Article", "headline": "Best Claude prompts for therapists in 2026", "description": "Twelve clinician-tested Claude prompts for DAP notes, treatment plans, psychoeducation handouts, case conceptualization, supervision prep, telehealth intake, insurance appeals, sliding-scale conversations, marketing bios, caseload reflection, between-session worksheets, and ethical decisions — with HIPAA and 42 CFR Part 2 de-identification rules.", "datePublished": "2026-06-10", "dateModified": "2026-06-10", "author": { "@type": "Person", "name": "Dr. Sarah Chen", "jobTitle": "Licensed clinical psychologist and clinical-AI consultant" }, "publisher": { "@type": "Organization", "name": "AIPromptsHub", "url": "https://aipromptshub.co" }, "mainEntityOfPage": "https://aipromptshub.co/blog/best-claude-prompts-for-therapists-2026" }) }} />

<script type="application/ld+json" dangerouslySetInnerHTML={{ __html: JSON.stringify({ "@context": "https://schema.org", "@type": "FAQPage", "mainEntity": [ { "@type": "Question", "name": "Is it HIPAA-compliant to use Claude for clinical documentation?", "acceptedAnswer": { "@type": "Answer", "text": "Only with a signed Business Associate Agreement via an eligible enterprise plan or HIPAA-aligned partner, de-identification of inputs before paste, and an organizational policy authorizing the use. Consumer Claude.ai without a BAA is not appropriate for PHI." } }, { "@type": "Question", "name": "Will Claude replace a therapist?", "acceptedAnswer": { "@type": "Answer", "text": "No. Claude drafts notes, scaffolds case formulations, and supports organization. Clinical judgment, diagnosis, risk assessment, and signature remain the clinician's under APA/NASW codes and state licensing rules." } }, { "@type": "Question", "name": "Which Claude model works best for clinical work?", "acceptedAnswer": { "@type": "Answer", "text": "Claude Sonnet 4 or newer. Older models miss the structured reasoning required by the case-conceptualization scaffold and the ethical-reasoning framework." } }, { "@type": "Question", "name": "What if my client asks whether I use AI for their notes?", "acceptedAnswer": { "@type": "Answer", "text": "Tell them. A short paragraph in your intake paperwork describing that you sometimes use AI tools to structure notes, that you de-identify before use, and that you review and sign every document, is consistent with APA and NASW informed-consent obligations." } }, { "@type": "Question", "name": "How do I know if Claude's clinical output is wrong?", "acceptedAnswer": { "@type": "Answer", "text": "Read every line. Check diagnosis-adjacent language for unwarranted certainty. Verify any cited code section, ICD/CPT code, or regulation. Hallucinated citations and overconfident clinical claims are the two most common failure modes." } }, { "@type": "Question", "name": "Does this apply to 42 CFR Part 2 (SUD) records?", "acceptedAnswer": { "@type": "Answer", "text": "Part 2 deserves its own policy. Default position: do not route Part 2 content through any LLM, even de-identified, without specific organizational-leadership policy approval." } } ] }) }} />